36 أنظمة التداول على أساس ريست

غارمين الجديد المتقدم 35 ساعة لديها المعصم القائم على أجهزة الاستشعار معدل ضربات القلب.
واحد من عدة.
هناك عدد قليل من اليقين المعروفة في العالم، ولكنني بدأت أعتقد أن غارمين الإفراج عن جديد يمكن ارتداؤها كل شهر هو واحد منهم. في الأشهر الأربعة الماضية وحدها شهدنا مقدمة من 735XT السبق، و فيفوموف، و فيفوسمارت هر +، و X40 نهج تعقب الغولف، و غارمين فينيكس كرونوس.
اليوم ليست استثناء: غارمين قد أعلنت للتو 35 المتقدم، تحديث لشعبية 25 المتقدم. أكبر الاختلافات بين الإصدار السابق وهذا الجديد هو بناء أكثر أناقة، وعرض عالية الدقة، مطالبات عمر أطول للبطارية، والأهم من ذلك، إضافة أجهزة الاستشعار معدل ضربات القلب القائم على المعصم إلى السباق 35 (هل يمكن أن الزوج FR25 مع حزام الصدر، ولكن لم يكن لديك أجهزة استشعار معدل ضربات القلب بنيت في).
غارمين في مقدم 35 عمر البطارية تسعة أيام.
يقول غارمين أن الساعة تأتي مع عدة أنماط رياضية مختلفة، مثل المشي والجري وركوب الدراجات والمشي على التوالي، ومجموعة متنوعة من أنشطة القلب الأخرى. انها لا تتبع السباحة، على الرغم من أنها للماء تصل إلى 50 مترا. وهو يتتبع المسافات مع المدمج في نظام تحديد المواقع، وهي جزء أساسي من ساعات اليد غارمين السابقة. مطالبات عمر البطارية الجديدة هي حوالي تسعة أيام في وضع "العادية" ووتش، مع تتبع النشاط الأساسي. و 13 ساعة في وضع غس.
كما هو الحال مع معظم الأجهزة القابلة للارتداء المعصم في هذه الأيام أيضا يؤدي 35 المتقدم أيضا بعض الوظائف الأساسية للاللياقة البدنية تعقب المائل سمارتواتش: عد الخطوات والسعرات الحرارية المحروقة، والتي تبين تنبيهات الإخطار من الهاتف الذكي، وتقديم ضوابط الموسيقى من المعصم. فإنه أزواج مع غارمين الاتصال التطبيق المحمول، والذي يتوفر على دائرة الرقابة الداخلية، الروبوت، ويندوز فون، وأجهزة الكمبيوتر المكتبية.
سوف السفينة 35 سوف السفينة في وقت ما في خريف هذا العام 200 $، وهو ما يعني انها أقل تكلفة بكثير من السباق 735XT (450 $) وتكاليف أقل من حتى فيفواكتيف هر (250 $)، وكلاهما تتبع السباحة. أحيانا أبسط هو أفضل.
التالي حتى في قواطع دوائر.
الآن تتجه.
سطر الأوامر.
يوفر سطر الأوامر التحديثات اليومية من المستقبل القريب.

36 أنظمة التداول على أساس ريست
الانتماء: قسم الطب التشخيصي / علم الأمراض المرضية، كلية الطب البيطري، جامعة ولاية كانساس، مانهاتن، كس، الولايات المتحدة الأمريكية.
بياتريز مارتينيز-لوبيز.
الانتماء: مركز لنمذجة الأمراض الحيوانية والترصد (كادمس)، جامعة كاليفورنيا ديفيس، ديفيس، كاليفورنيا، الولايات المتحدة الأمريكية.
دارلا كويجادا.
الانتماء: المركز الوطني للأمن البيولوجي الزراعي، جامعة ولاية كانساس، مانهاتن، كس، الولايات المتحدة الأمريكية.
كينيث برتون.
الانتماء: المركز الوطني للأمن البيولوجي الزراعي، جامعة ولاية كانساس، مانهاتن، كس، الولايات المتحدة الأمريكية.
الانتماء: قسم الطب التشخيصي / علم الأمراض المرضية، كلية الطب البيطري، جامعة ولاية كانساس، مانهاتن، كس، الولايات المتحدة الأمريكية.
وتعتمد سالمة الثروة الحيوانية في الواليات المتحدة بشدة على قدرتها على منع إدخال األمراض الحيوانية العابرة للحدود) تادس (. ولذلك، فإن المعلومات الدقيقة والمحدثة عن موقع وأصل مخاطر تلك المخاطر المحتملة أمر ضروري، لذلك يمكن وضع تدابير وقائية مثل القيود المفروضة على السوق. الهدف من هذه الدراسة هو تقييم الخطر الحالي للحمى الخنازير الأفريقية (أسف) وحمى الخنازير الكلاسيكية (كسف) مقدمة في الولايات المتحدة من خلال الاستيراد القانوني من الخنازير الحية ومنتجات الخنازير باستخدام النهج الكمي التي يمكن تطبيقها لاحقا على مخاطر أخرى. تم تطوير أربعة نماذج لتقييم المخاطر العشوائية الكمية لتقدير الاحتمالات الشهرية لإطلاق أسف و كسف في الولايات المتحدة، وتعرض السكان المعرضين للإصابة (الخنازير المنزلية والحشائرية) لهذه المداخل على مستوى الدولة. وتشير النتائج إلى احتمال سنوي منخفض لإدخال أسف أو كسف في الولايات المتحدة، من خلال أي من المسارات تحليلها (5.5 * 10 -3). كونه احتمال إدخال من خلال الواردات القانونية من الخنازير الحية (1.8 * 10 -3 ل أسف، و 2.5 * 10 -3 ل كسف) أعلى من خطر المنتجات الخنازير المستوردة قانونيا (8.90 * 10 -4 ل أسف، و 1.56 * 10 -3 ل كسف). ويمكن أن يحدث ذلك بسبب احتمال التعرض المنخفض المرتبط بهذا النوع من السلع (المنتجات). وكان خطر دخول الخنازير الوحشية إلى منتجات الخنازير التي تم التخلص منها في مدافن القمامة أعلى قليلا من التعرض المحتمل للخنازير المحلية من خلال تغذية الخنازير. إن تحديد الأشهر المعرضة لأعلى المخاطر، وأصل الواردات المرتفعة المخاطر، وموقع الولايات المتحدة الأكثر ضعفا على تلك المداخلات (أيوا ومينيسوتا وويسكونسن للخنازير الحية وكاليفورنيا وفلوريدا وتكساس لمنتجات الخنازير) هو معلومات قيمة من شأنها أن تساعد في وضع استراتيجيات الوقاية والتخفيف من المخاطر والكشف المبكر التي من شأنها أن تساعد على التقليل إلى أدنى حد من العواقب الكارثية للمقدمات أسف / كسف المحتملة في الولايات المتحدة.
اقتباس: هيريرا-إباتا دم، مارتينيز-لوبيز B، كويجادا D، بيرتون K، مور L (2017) النهج الكمي لتقييم مخاطر حمى الخنازير الأفريقية وحمى الخنازير الكلاسيكية مقدمة في الولايات المتحدة من خلال الواردات القانونية من الخنازير ومنتجات الخنازير . بلوس وان 12 (8): e0182850. دوى: 10.1371 / journal. pone.0182850.
المحرر: شو سو، جامعة نانجينغ الزراعية، الصين.
تاريخ الاستلام: 11 أبريل 2017؛ مقبول: 25 يوليو 2017؛ تاريخ النشر: 10 أغسطس 2017.
حقوق الطبع والنشر: © 2017 هيريرا-إباتا إت آل. هذه المقالة هي مقالة الوصول المفتوح الموزعة بموجب شروط ترخيص كريتيف كومونس أتريبوتيون، والتي تسمح بالاستخدام غير المقيد والتوزيع والاستنساخ في أي وسيط، بشرط أن يقيد المؤلف الأصلي والمصدر الأصلي.
توافر البيانات: جميع البيانات ذات الصلة هي داخل ورقة.
التمويل: تستند هذه المادة إلى العمل الذي تدعمه هيئة كانساس للعلوم البيولوجية بموجب اتفاقية منحة برنامج المطابقة الاتحادية من خلال مركز التميز للأمراض الحيوانية الناشئة والحيوانية المنشأ (سيزاد)، ودولة كانساس، والمرفق الوطني للحيوية والدفاع الزراعي (نباف ) صندوق الانتقال. ولم يكن للممولين دور في تصميم الدراسة، وجمع البيانات وتحليلها، أو قرار نشرها، أو إعداد المخطوطة.
تضارب المصالح: أعلن المؤلفون أنه لا توجد مصالح متنافسة.
المقدمة.
حمى الخنازير الأفريقية (أسف) وحمى الخنازير الكلاسيكية (كسف) هي أمراض حيوانية يمكن إخطارها بالمنظمة العالمية لصحة الحيوان (أوي) [1] والتي لها تأثير مدمر في البلدان المتضررة. كل من الأمراض تسبب متلازمات الحمى النزفية في الخنازير، مع عرض سريرية مماثلة ومعدلات وفيات عالية. ومع ذلك، فإنها تسببها الفيروسات المصنفة في عائلات مختلفة: فيروس أسف (أسف) هو فيروس دنا معقدة عضو فريد من عائلة أسفارفيريداي [2]، في حين أن سبب كسف عن طريق فيروس الحمض النووي الريبي من جنس فيروس، عائلة فلافيفيريداي [3] . أسف و كسف تؤثر حصرا على الخنازير، المحلية والبرية، ويمكن أن تنتقل عن طريق إما الاتصال المباشر بين الخنازير أو نقل غير مباشر، في المقام الأول من خلال ابتلاع المنتجات المصابة والاتصال مع فوميتس الملوثة الأخرى [2، 3]. كما تنتقل أسف عن طريق القراد لينة (أورنيثودوروس سب.) [4]، ويجري O. موباتا، ال التعريف، اصلي، سهم التوجيه، إلى داخل، إفريقيا، أيضا، O. إراتيكوس في البحر المتوسط ​​[5].
وقد وصفت منظمة أطباء بلا حدود لأول مرة في أفريقيا في عام 1921، والتي لا تزال متوطنة في معظم بلدان جنوب الصحراء الكبرى [6]. ومع ذلك، لم يكن هذا المرض مقتصرا بشكل فريد على القارة الأفريقية. بين 1960 و 1990 كان أسف حاضرا في شبه الجزيرة الايبيرية مما تسبب في تفشي متفرقة في بلدان أخرى من أوروبا ومنطقة البحر الكاريبي وأمريكا الجنوبية (البرازيل). في عام 1995، تم استئصال المرض بنجاح من جميع تلك الأراضي باستثناء جزيرة سردينيا الإيطالية حيث لا يزال متوطنا. تغير الوضع الوبائي العالمي لل أسف تغيرا جذريا بعد إدخال أسف للمرة الأولى في جورجيا في عام 2007 [7]. من هناك، أسف تنتشر بكفاءة إلى مناطق واسعة من منطقة القوقاز، التي تؤثر على جنوب وغرب روسيا [8] وبيلاروس وأوكرانيا، حتى وصلت أخيرا إلى الاتحاد الأوروبي في عام 2018. بحلول عام 2017، ومن المعروف أن أسف موجودة في روسيا وبيلاروسيا وأوكرانيا ودول البلطيق (إستونيا ولاتفيا وليتوانيا) وبولندا ومولدوفا [9]. استمرار انتشار المرض نحو المناطق الغربية مع الحالات المستمرة من أسف في الخنازير البرية والخنازير المحلية، تعكس عدم نجاح برامج السيطرة في المنطقة والتهديد المحتمل لصناعة الخنازير في جميع أنحاء العالم بسبب عدم وجود لقاح [10] .
في حين أن أسف كان يرتبط تقليديا مع القارة الأفريقية (وسردينيا) حتى إعادة إدخاله في أوروبا، وقد كسف أوسع بكثير وزعت تسبب مشاكل هامة في جميع أنحاء العالم لعقود [11]. ومن الجدير بالذكر أن الخسائر نشأت في العديد من البلدان الأوروبية خلال التسعينات بعد سياسة حظر التطعيم [12] وانتشار التراكيب الوراثية 2 و 3 في الصين [13]. وعلى الرغم من توافر اللقاحات، إلا أنه بحلول عام 2017، لا يزال مرض السرطان الوبائي مستوطنا في أجزاء كثيرة من آسيا وأمريكا الجنوبية وبعض الجزر الكاريبية القريبة من الولايات المتحدة (كوبا) وهايتي والجمهورية الدومينيكية [14].
الولايات المتحدة الأمريكية هي ثالث أكبر منتج ومستهلك لحم الخنزير في العالم، وأكبر مصدر لحم الخنزير في العالم، بمتوسط ​​20٪ من لحم الخنزير سنويا المنتجة في الولايات المتحدة يتم تصديرها [15]. ولذلك، من الضروري وضع استراتيجيات للوقاية ونظم للإنذار المبكر تقوم على تحليل المخاطر للحد من احتمال إدخال الأمراض الحيوانية العابرة للحدود. لم يحدث هذا المرض أبدا في الولايات المتحدة ولم يكن هناك علاج مضاد للفيروسات ولا لقاح متاح، ولا يتوقع أن يكون متاحا على المدى القصير، لذا فإن الوقاية من الأمراض أمر ضروري لأن تدابير المكافحة تستند حصرا إلى ختم السياسات. في المقابل، بالنسبة لفيروس كسف هناك عدة لقاحات فعالة متاحة في السوق [16]، حتى في شكل الطعم الشفوي، والتي استخدمت بنجاح في الحياة البرية من خلال حملات التطعيم الشامل [17]. ومع ذلك، وكما واجهت الولايات المتحدة بالفعل في الماضي، ومراقبة واستئصال كسف يطرح تحديات الرقابة الهامة والآثار الاقتصادية المدمرة لصناعة الخنازير (أي التكلفة التقريبية للسيطرة كسف في الولايات المتحدة في عام 1978 كان 140 مليون $).
و أسف و كسف يمثلان خطرا على صناعة الخنازير العالمية، ومن المتوقع حدوث عواقب اقتصادية هامة إذا تم إدخالها في بلدان إنتاج الخنازير مثل الولايات المتحدة [18، 19]. ولذلك، فإن الهدف من هذه الدراسة هو تقييم خطر كل من أسف و كسف مقدمة في الولايات المتحدة على مستوى الولايات من قبل الواردات القانونية من الخنازير الحية ومنتجات الخنازير على أساس شهري. وستساعد نتائج هذه التقييمات على تحديد الطرق والمواقع والأوقات المحتملة التي يكون فيها عدد المواشي في البلد معرضا لمخاطر أعلى، وعندما ينبغي تنفيذ تدابير وقائية لتجنب إدخال أسف و كسف إلى الولايات المتحدة. وسيتم إدراج النتائج التي تم الحصول عليها في منصة على الانترنت (المرض بيوبورتال ™ يمكن الوصول إليها في bioportal. ucdavis. edu/) لتسهيل التصور والتحليل، والنهج الكمي وضعت هنا سوف تساعد على تطوير قالب لتقييم إمكانية إدخال تادس أخرى.
المواد والأساليب.
تم تطوير أربعة نماذج لتقييم المخاطر العشوائية الكمية لتقدير الاحتمالات الشهرية لإدخال أسف و كسف في الولايات المتحدة من خلال الواردات القانونية من الخنازير الحية والمنتجات خلال فترة عالية المخاطر (أي فترة من الزمن من العدوى في بلد المنشأ إلى الكشف والإخطار). وعلى وجه التحديد، تناولت النماذج الأربعة مسارات الأحداث التالية: (1) خطر إدخال أسف عن طريق الواردات القانونية من الخنازير الحية، و 2) خطر إدخال أسف عن طريق الواردات القانونية من منتجات الخنازير، و 3) خطر إدخال كسف عن طريق الاستيراد القانوني للخنازير الحية و (4) خطر إدخال كسف عن طريق الواردات القانونية من منتجات الخنازير.
وأخيرا، تم الجمع بين هذه النماذج الأربعة لتقدير 1) احتمالات مقدمة أسف و 2) احتمال مقدمة كسف من قبل أي من المسارات المحلل. تم الجمع بين هذين الاحتمالين أخيرا لتقدير احتمال إما، أسف أو كسف، مقدمة في الولايات المتحدة عن طريق الاستيراد القانوني للخنازير أو المنتجات. تم تطوير جميع النماذج فيRISK 7.5 (باليسيد كوربوراتيون، نيوفيلد، ني، أوسا) على ميكروسوفت إكسيل 2007 ® وتشغيل 1000 التكرار باستخدام طريقة أخذ العينات مونت كارلو.
* ملاحظة: من أجل تبسيط النص وتجنب التكرار، أحيلت الجمل والمعلمات والاحتمالات التي تنطبق على كل من نماذج الأمراض (النموذج الذي تم تطويره ل أسف ونموذج كسف) باسم "أسف / كسف". على سبيل المثال "احتمال إدخال أسف / كسف في الولايات المتحدة من قبل الواردات القانونية من الخنازير الحية" يقرأ باسم "احتمال إدخال أسف في الولايات المتحدة واحتمال مقدمة كسف في الولايات المتحدة". ومن المهم التفريق بين هذه الملاحظة واحتمال وقوع الحدثين (أسف و كسف) في نفس الوقت (P أسف ∩P كسف).
مواصفات النماذج.
استنادا إلى المبادئ التوجيهية لمنظمة أوي [20]، ينقسم تحليل المخاطر لإدخال مسببات الأمراض من خلال الواردات إلى ثلاث خطوات: الدخول (سابقا وفي الدراسة الحالية تسمى الإفراج [21])، وتقييم التعرض والنتائج. قيمت نماذج تقييم المخاطر التي تم تطويرها في هذه الدراسة احتمال وجود أسفف واحتمال إطلاق سففف في الولايات المتحدة والتعرض اللاحق للسكان المعرضين للإصابة (الخنازير المحلية والخنازير الوحشية). تم حساب الاحتمال النهائي لكل مسار على النحو التالي:
حيث P F هو الاحتمال النهائي، P R هو احتمال إطلاق أسف / كسف في الولايات المتحدة، و P E هو احتمال التعرض. وتلخص أشجار الحدث المبينة في الشكلين 1 و 2 هيكل وسلسلة أحداث مسارات الخطر للواردات القانونية من الخنازير الحية ومنتجات الخنازير على التوالي. ويتضمن الجدول 1 معلومات مفصلة عن بلدان المنشأ والواردات إلى الولايات المتحدة، في حين تتضمن الجداول 2 و 3 و 4 معلومات تفصيلية عن المعلمات المستخدمة لإطعام النماذج الأربعة. وعلى وجه التحديد، يتضمن الجدول 2 المعلمات المستخدمة في إصدار وتقييم التعرض للمخاطر المرتبطة باستيراد الخنازير الحية، في حين تم تقسيم المعلومات المتعلقة بمخاطر واردات منتجات الخنازير إلى إصدار (الجدول 3) والتعرض (الجدول 4).
توسيع الشكل 1. شجرة الحدث من مقدمة أسف / كسف في الولايات المتحدة من قبل الواردات القانونية من الخنازير الحية.
ملاحظة: المعلومات حول المدخلات التي تحمل علامة * مدرجة في الجدول 2.
تقييمات الإصدار.
وقد تم وضع نموذج لاحتمال إطلاق سراح خنزير محلي واحد على الأقل من أسف / كسف أو كيلوغرام من منتج الخنازير عن طريق الاستيراد القانوني من بلد المنشأ (ج) إلى حالة من الولايات المتحدة (الولايات المتحدة) عملية ثنائية الشكل من النموذج:
حيث يشير الرقم إلى عدد الخنازير أو الكيلوغرامات من منتجات الخنازير (حسب المسار الذي تم تقييمه) المستورد قانونا من بلد المنشأ (ج) إلى دول المقصد؛ و هو احتمال أن تصل الحيوانات أو المنتجات المصابة (بما في ذلك اللحوم المبردة والمجمدة والمدخنة والدهون والمخلفات من الخنازير) من بلد المنشأ (ج) إلى دول المقصد.
تم الحصول على المعلومات عن الخنازير الحية المستوردة ومنتجات الخنازير من نظام التجارة الزراعية العالمي التابع لوزارة الزراعة الأمريكية [22] وتجارة الولايات المتحدة الأمريكية مكتب التعداد [23] من 2008 إلى 2018. وتضمنت قاعدة بيانات الجاتس الكميات والقيمة ( $) من الحيوانات الحية المستوردة (عدد الخنازير) والمنتجات (كلغ) حسب الشهر وبلد المنشأ، ولكن مجمعة لكل الولايات المتحدة. من ناحية أخرى، تضمنت قاعدة بيانات مكتب الإحصاء الأمريكي قيمة ($) من الواردات شهريا، وبلد المنشأ والولايات الأمريكية المقصد. تم الجمع بين كل من قواعد البيانات لتقدير كمية (عدد الخنازير و كغ من المنتجات) المستوردة شهريا من كل بلد المنشأ إلى كل ولاية من الولايات المتحدة. ويمكن الاطلاع على مزيد من المعلومات المفصلة عن بيانات الاستيراد وبلدان المنشأ في الجدول 1.
تم حساب الاحتمال p كس كمنتج من الاحتمالات الشرطية المختلفة على النحو التالي:
على وجه التحديد، تم حساب احتمال الإفراج عن الاستيراد القانوني للخنازير الحية (p سل) على أنها نتاج أربعة احتمالات مشروطة (الشكل 1). احتمالية إصابة أسف / كسف في بلد المنشأ (P 1)؛ احتمالية اختيار خنزير مصاب بمرض أسف / كسفف في البلد ج (P 2L)؛ احتمال البقاء على قيد الحياة من عدوى أسف / كسف (P 3). واحتمال أن الخنازير المصابة تبقى على قيد الحياة النقل من البلد ج إلى المقصد (1-P 4). ونظرا لنقص المعلومات المتعلقة بالحيوانات الحاملة (انتشارها، وقدرات انتقالها، وما إلى ذلك)، لم ينظر في النماذج السريرية الواضحة للمرض إلا في النماذج.
وقد تم تقدير احتمال الإفراج عن الاستيراد القانوني للمنتجات (p سسب) (الشكل 2) على أنه ناتج عن احتمالين مشروطين: P 1 (احتمال عدوى أسف / كسف في بلد المنشأ، بعد نفس هيكل الخنازير الحية النماذج) و P 2P كاحتمال اختيار كيلوغرام من منتجات الخنازير المصابة مع أسف / كسفف في البلد ج.
احتمال العدوى في بلدان المنشأ (P1).
وقد اختيرت احتمالية الإصابة في بلد المنشأ ج بشكل مختلف تبعا لحالة الأمراض في البلد المعني. أما بالنسبة للبلدان التي أصيبت حاليا بالعدوى بفيروس نقص المناعة البشرية / متلازمة نقص المناعة المكتسب (أسف / كسف)، فقد تم استخدام دالة أسيية لتقدير احتمال حدوث تفشي واحد على الأقل في الفاصل الزمني المحدد بالشكل التالي:
حيث t هو الفاصل الزمني (شهر واحد) و λ هو متوسط ​​عدد فاشيات المرض في الشهر التي تم الحصول عليها من المعلومات الفاشية التاريخية [14] المقدرة ب λ = أ / ب. كونه العدد الإجمالي للفاشيات في الخنازير المحلية و ب العدد الإجمالي للأشهر من فترة الدراسة.
أما بالنسبة للبلدان التي ليس لديها سجلات لوجود أسف / كسف على مدى السنوات العشر الأخيرة، فقد تم تقدير احتمال العدوى من خلال نموذج مقسوم (S) (إنفكتد) (S) قابل للتأثر (S) يعتبر شبكة تجارة الخنازير وحالة المرض في البلدان التجارية. على وجه التحديد، استخدم النموذج بيانات عن تجارة الخنازير العالمية من آخر عام متاح (2018) [24]، وحالة المرض في البلدان التجارية على أساس المعلومات القطرية لمنظمة أوي-واهيس (وجود أو عدم وجود أسف / كسف) [14]. تم بناء نموذجين، واحد حيث تم توزيع البذور من العدوى بشكل عشوائي في البلدان الإيجابية أسف، والآخر حيث تبدأ عمليات المحاكاة في البلدان إيجابي كسف. في كلتا الحالتين، أنشأنا مصفوفات المجاورات الموجهة بعد اتجاه التجارة (من المصدر إلى المستورد). تم افتراض فترة عالية المخاطر لمدة 30 يوما ومستويين من درجات الاتصال المحتملة بين الدول التجارية لجميع النماذج. تم تقييم عدة سيناريوهات مع معدلات انتقال مختلفة (بت) لكل مرض بما في ذلك 1٪، 2.5٪، 5٪ و 10٪ ل أسف [25]؛ و 0.4٪، 6٪ و 12٪ ل كسف [26]. وقد بنيت نماذج سي في R البرمجيات [27] باستخدام حزم إيغراف [28] و سنا [29]. تم تشغيل كل سيناريو مع 1000 محاكاة. واعتبرت نسبة المرات التي أصيب فيها كل بلد مستلم بالعدوى خلال عمليات المحاكاة الكلية بمثابة خطر الإصابة بالعدوى في بلد المنشأ. واستخدمت القيم الدنيا والمتوسطية والقصوى للمخاطر التي تم الحصول عليها في سيناريوهات النموذج المختلفة لتصنيف توزيعات بيرت (الحد الأدنى والأرجح والأقصى) لمخاطر العدوى في بلدان المنشأ (انظر الأمثلة في الجدولين 2 و 3) .
احتمال اختيار الخنازير الحية المصابة (P 2L) أو المنتج (P 2P) في البلد ج قبل الكشف عن الأمراض.
في النماذج التي تقيِّم المخاطر المرتبطة باستيراد الخنازير الحية، تم وضع نموذج لاحتمال اختيار الخنازير الحية المصابة في بلد المنشأ ج باستخدام توزيع بيتا يحدده البارامترات التالية: α 1L على أنه عدد الخنازير المقدر ة المسبقة قبل أسف / كسف في البلد ج، و α 2L كما العدد الإجمالي للخنازير في البلد ج. وقدر عدد الخنازير المصابة في البلد ج على أنه نتاج ثلاثة معايير مستقلة: (1) الفاشيات غير المكتشفة خلال فترة الخطورة العالية في البلد (ج) [14]، 2) متوسط ​​حجم قطيع الخنازير في بلد المنشأ (1) إلى [24]، و 3) انتشار القطيع (هب) ل أسف / كسف [14] (الجدول 2). ولتقدير البارامترات المتعلقة بالمرض (أوو و هب)، لم يتم النظر إلا في التقارير المتعلقة بالتفشي الأوروبي الأخير، في حالة مرض الالتهاب الرئوي الحاد، لأن التقارير الواردة من أفريقيا كثيرا ما تفتقر إلى البيانات الأساسية المطلوبة. وبالنسبة لحسابات أوو، قمنا بحساب عدد الفاشيات التي تحدث في البلدان المتضررة حديثا بين الحدث الأول وتاريخ التقرير الأول (فترة عالية الخطورة)، في حين تم تقدير معدل انتشار القطيع في كل حدث بالنظر إلى عدد الحالات مقابل عدد الحيوانات الحساسة الموجودة في الفاشية.
في النماذج التي تم تطويرها لتقدير المخاطر المرتبطة باستيراد منتجات الخنازير، كان احتمال اختيار المنتجات المصابة من بلد المنشأ (ج) محددا أيضا بتوزيع بيتا يتميز بمعلمين: ألف 1P ككمية من الخنازير التي يحتمل أن تكون مصابة المنتجات في البلد ج، و α 2P كما كمية (كغ) من منتجات الخنازير المنتجة (نب) في البلاد ج. وقدرت كمية المنتجات المصابة في بلد المنشأ (كيم) بنتيجة ثلاث معلمات: ني) العدد المقدر للخنازير المصابة في بلد المنشأ c، P m) احتمال تحول الخنازير المصابة ب أسف / كسف إلى اللحوم، والنائب) متوسط ​​وزن المنتجات (كجم) التي تم الحصول عليها في خنزير ذبح (الجدول 3).
تم حساب عدد الخنازير المصابة (ني) بشكل مماثل لنماذج الخنازير الحية، ومنتج أو * هب * ل. وقد تم تقدير احتمال وجود خنزير مصاب بالعدوى بالصفائح (أسف / كسف) يتحول إلى لحم (P m) على النحو التالي: P m = P 3 * (1-P 4) * P سم * P.
حيث P 3 هو احتمال وجود الخنازير على قيد الحياة العدوى أسف / كسف، P 4 هو احتمال الخنازير لا البقاء على قيد الحياة النقل، P سم هو احتمال أن يتم ذبح الخنزير خلال شهر معين، و P كما احتمال أسف / كسف الخنازير المصابة التي لم يتم الكشف عنها خلال الفحوص السريرية التي أجريت في المسلخ. وقد تم تقدير نسبة الشلل الرخو الحاد باستخدام البيانات التاريخية لتعداد الخنازير والنسبة الشهرية للخنازير المرسلة إلى الذبح في عدد من البلدان. واستخدمت بيانات من بلدان الاتحاد الأوروبي لهذا الغرض ([36])، بافتراض أن جميع أنواع نظم إنتاج الخنازير ممثلة فيها.
تقييم التعرض.
وتختلف احتمالات السكان المعرضين للإصابة (الخنازير المحلية والخشبية) في الولايات المتحدة التي تتلامس مع الفيروسات بشدة تبعا لنوع المصفوفة المصابة (الخنازير الحية أو منتجات الخنازير). لذلك، تم تحليل تقييمات التعرض لكلا المسارين بشكل مختلف وشرح بشكل منفصل أدناه.
احتمال تعرض الخنازير المحلية الأمريكية للخنازير المحلية المستوردة (بي L).
تم تقدير احتمال وجود الخنازير المحلية في الولايات المتحدة للحصول على اتصال فعال مع الخنازير المصابة المستوردة قانونا (بي L) على النحو التالي:
حيث P د هو احتمال وجود خنازير مستوردة ذات وجهة زراعية، ف ف احتمالية الخنازير المستوردة التي يتم عزلها، و P احتمال عدم الكشف عن الخنازير المصابة أثناء الحجر الصحي. P q تم توزيعه كتوزيع بيتا باستخدام بيتابوستر [37] بالنظر إلى القيمة الأكثر احتمالا من 0.95 و 90٪ من الثقة من احتمال أعلى من 0.9. وقد افترض المؤلفون هذه القيم استنادا إلى التشريع الأمريكي بشأن إجراءات الحجر الصحي [34] ودرجة معينة من عدم اليقين بسبب الامتثال غير الملائم للتشريعات. وبالنظر إلى أوجه الشبه في العرض السريري لل أسف و كسف، افترضنا نفس احتمال عدم الكشف عن كل من الأمراض، وبالتالي، نفس احتمال التعرض للخنازير المستوردة التي يحتمل أن تكون مصابة. وترد جميع المعلومات المتعلقة بهذه الاحتمالات في الجدول 2.
لم يتم تضمين الاتصال المحتمل من الخنازير المحلية المستوردة مع الخنازير الوحشية في الولايات المتحدة في النموذج، نظرا لخصائص الخنازير الخنازير المستوردة (99.9٪ خنازير المغذية القادمة من كندا [38])، وجهتهم الأكثر احتمالا سيكون التشطيب مرفق مع معايير الأمن البيولوجي عالية وأي اتصال محتمل مع الخنازير الوحشية.
احتمال التعرض لمنتجات الخنازير المستوردة إلى الولايات المتحدة (بي P).
ويمكن الاطلاع على المعلومات التفصيلية عن المدخلات والحسابات المستخدمة لتقدير احتمال التعرض لمنتجات الخنازير المستوردة إلى الولايات المتحدة في الجدول 4 والشكل 2. ويمكن التخلص من منتجات الخنازير المستوردة قانونيا في الولايات المتحدة وتصبح النفايات الغذائية إما في نقطة البيع إذا لم يتم بيعها في الوقت المحدد (تسمى التجزئة هنا)، أو على مستوى المستهلك (المطاعم وحجز المنازل). مصادر أخرى من النفايات الغذائية هي المؤسسات والمستشفيات والكليات والسجون. ومع ذلك، فإن مساهمتها في إجمالي النفايات الغذائية المتولدة في الولايات المتحدة محدودة (بين 5.4 و 8٪) [39، 40]. ولذلك، ونظرا لعدم وجود معلومات تفصيلية عن إدارة النفايات المتولدة في هذا القطاع، لم تدرج في التحليل سوى ثلاث فئات. واستخدمت إحصاءات محددة لنسبة الأغذية واللحم الخنزير المهدر في كل مستوى لتقدير احتمال أن تكون منتجات الخنازير غير مستهلكة في القطاعات الثلاثة المختلفة لسلسلة الاستهلاك التي تم تحليلها (ور، W رت و و، يحمل، على التوالي). تم تقدير احتمال أن يتم التخلص من لحم الخنزير على مستوى المنزل (و) كمنتج لنسبة الطعام الذي يضيع على مستوى المستهلك (W C) بنسبة لحم الخنزير المستهلك في المنازل بدلا من المطاعم (1-C رت). وينطبق العكس على النفايات في مستوى المطاعم W رت = W C * C رت.
وبمجرد التخلص منها، يمكن أن تكون لنفايات الأغذية وجهات مختلفة، بما في ذلك عمليات الاستعادة (إنتاج الوقود الحيوي، وإعادة التدوير، والتسميد، وما إلى ذلك)، وتغذية الحيوانات (ومعظمها خنازير) أو التخلص منها في مقالب القمامة. الخيار الأول لا ينطوي على الاتصال مع الخنازير، وبالتالي لم يتم إجراء المزيد من الدراسات على ذلك. وقد قام النموذج بتقييم طريقين محتملين للاتصال مع السكان الخنازير المعرضين للخطر: P سف، واحتمال الخنازير المحلية في الحصول على اتصال مع المنتجات المصابة من خلال تغذية سويل، و P لف، واحتمال الخنازير الوحشية في الحصول على اتصال مع المنتجات المصابة الوصول إلى مقالب القمامة .
ويختلف تواتر ممارسات التغذية السائلة واللوائح المطبقة، اعتمادا شديدا على مصدر نفايات الأغذية (أي التجزئة والمطاعم مقابل الأغذية المنزلية)، ولذلك فإن النفايات المنزلية تعتبر منفصلة عن النفايات الأخرى. وبالنسبة لاحتمال التغذية السائلة الناتجة عن التجزئة (سف R) والمطاعم (سف رت)، فقد اعتبرنا نسبة النفايات الغذائية المستخدمة في وقت لاحق كغذاء سائل من قطاع التجزئة (سف | ور) [41] والمطاعم (سف | W رت ) [41، 42]، وتفويض هذه الممارسة في الدولة (غف)، وفعالية العلاج (إف) والممارسات غير القانونية المحتملة (بيل). يسمح الطعام والنفايات الغذائية المطاعم فقط لاستخدامها كعلف للحيوانات في ولايات معينة من الولايات المتحدة وفقا للوائح قانون التغذية القمامة [43]. واستنادا إلى هذا التشريع، ينبغي معالجة نفايات الأغذية عند 212 درجة فهرنهايت لمدة 30 دقيقة على الأقل، مما يضمن تعطيل مسببات الأمراض. وقد تم تضمين درجة معينة من عدم اليقين حول فعالية هذه العملية (إف) للدول التي يسمح فيها هذه الممارسة (تسمى غف) مع توزيع بيتا على افتراض أن مع 95٪ الثقة المعالجة التدفئة من النفايات الغذائية فعالة أكثر من 90 ٪ من الحالات، وعلى الأرجح 95٪ كفاءة. في الولايات التي تمنع هذه الممارسة (نغف)، تم افتراض درجة معينة من عدم اليقين بسبب الممارسات غير القانونية (P إيل)، من خلال توزيع بيتا مع 95٪ من الثقة أن الممارسات غير المشروعة تغذية سويل أقل شيوعا من 10٪ و على الأرجح أقل من 5٪. وقد استخدمت هذه الافتراضات لحساب توزيعات بيتا عن طريق استخدام بيتابوستر [37]. وقدرت احتمالية وصول الخنازير المحلية إلى مخلفات لحم الخنزير من التجزئة عن طريق تغذية السويل لكل ولاية على النحو التالي: سف R = W R * سف | W R * (غف * ((1- إف) + P إيل)). وفيما يتعلق بالتغذية من المطاعم (سف رت)، فإن المعادلة هي نفسها تماما، ولكنها تطبق القيم المتعلقة بنفايات المطاعم (أي W رت و سف | W رت).
لتقدير احتمالية ممارسات التغذية السائلة في الأسر (سف H) استخدمنا نهجا مماثلا للنهج المستخدم من قبل [44]. أولا، نظرنا في نسبة نفايات الطعام في الأسر (و)، واحتمال استخدام هذه النفايات كغذاء سويل (سف | W H) [45]. بعد ذلك، تم تقدير احتمال أن تكون منتجات الخنازير المستهلكة في أسرة مع وجود الخنازير (P هب) معتبرا عدد الأسر المعيشية مع الخنازير مقابل العدد الإجمالي للمنازل المحتلة في الولايات المتحدة [23]. وكشفت الدراسات الاستقصائية السابقة التي أجريت في الولايات المتحدة أن احتمال ممارسات التغذية سويل تختلف تبعا لحجم مزرعة الخنازير [46]. ولذلك، وبالنظر إلى نسبة مختلفة من حجم "مزارع الخنازير لكل ولاية [47] واحتمال ممارسة التغذية سويل لكل حجم المزرعة [48]، تم تقدير احتمال التغذية سويل بالتناسب لكل ولاية (سف | P هب). وأخیرا، تم تقدیر المخاطر المرتبطة بممارسات التغذیة السائلة مع نفایات الأغذیة من الأسر علی النحو التالي: سف H = W C * (1-C رت) * سف | و * P هب * (سف | P هب).
أما الوجهة الثانية المحتملة لنفايات الأغذية التي يمكن أن تؤدي إلى تفشي مرض أسف / كسف فهي التخلص منها في مدافن القمامة وإمكانية وصول الخنازير البرية إلى مدافن النفايات (P لف). وبما أن نسبة الخنازير المحلية الحرة (بدون حواجز) في الولايات المتحدة منخفضة جدا، فإن إمكانية حصول خنازير محلية على مدافن النفايات لم تؤخذ في الاعتبار في التقييم. ولتقييم هذا الخطر، حصلنا أولا على نسبة النفايات الغذائية التي تم التخلص منها في مدافن النفايات من قطاع التجزئة (لف R)، والمطاعم (لف رت) [41، 42] والأسر (95.2٪) [45]. بعد ذلك، تم إجراء تحليل مكاني باستخدام أرتجيس 10.3 (إسري ®) لتقدير نسبة مدافن النفايات في كل ولاية مع وجود محتمل للخنازير الوحشية (لف فب). ولتحقيق ذلك، تم تداخل خرائط توزيع الخنازير الوحشية ([49]) والمعلومات المتعلقة بتوزيع مدافن القمامة ([50])، وتم الحصول على إحصاءات لكل ولاية بما في ذلك الحد الأدنى والمتوسط ​​والأقصى لنسب عدد مدافن القمامة السطحية وقدرة المدافن الموجودة في مناطق مع وجود الخنازير الوحشية مقابل إجمالي مدافن القمامة في كل ولاية. واستخدمت هذه القيم لتصنيف احتمال وجود الخنازير الوحشية في مدافن النفايات في كل ولاية (لف فب). لم يتم العثور على بيانات ذات صلة باحتمال وجود الخنازير الوحشية التي يمكنها الوصول إلى محتوى مدافن القمامة. ولذلك، افترضنا نهجا متحفظا، مع احتمال الحد الأدنى من الخنازير الوحشية الوصول إلى مدافن 0.05، احتمال الأرجح من 0.1 والحد الأقصى 0.2 (فب A). وقدر احتمال أن تكون الخنازير الوحشية التي تتلامس مع نفايات الطعام التي يتم التخلص منها في مدافن القمامة القادمة من أي من القطاعات الثلاثة على النحو التالي:
وعلى افتراض أن كلا الطريقين للتعرض يستبعد أحدهما الآخر، بمعنى أنهما لا يمكنهما أن يحدثا في وقت واحد (أي إذا لم يستخدم المنتج المستهلك كغذاء سويل لا يمكن أن ينتهي في مكب النفايات، والعكس بالعكس)، فقد تم حساب التعرض النهائي على أنه المجموع of all three probabilities of domestic pigs being exposed by swill feeding (∑ SFi ) plus the probability of feral pigs getting in contact with infected imported swine products disposed in landfills (PLF), all of them calculated at state level.
The results of the models were presented as annual means (95% CI). The annual mean (sum of monthly probabilities) probabilities of ASF /CSF introduction were mapped in ArcGis 10.0 (ESRI®) using Natural Breaks (5 classes) as classification methods calculated by Jenks algorithm [59] and the Cartographic Boundary Shapefiles from the US Census Bureau.
Combined probabilities.
In order to provide a global picture of the risk for the introduction of these diseases in the US, we combined the four analyzed pathways. Firstly, we calculated the annual probabilities of introduction per country for each disease by any of the two pathways analyzed. Considering both pathways (legal import of pigs and legal importation of pig products) as independent events not mutually exclusive, the combined probability per disease was estimated as follows:
where P ASF is the combined probability of ASF being introduced into the US by legal imports, P ASFL is the probability of ASF being introduced into the US by legal imports of live pigs and P ASFP the probability of ASF being introduced into the US by legal imports of pig products. The same calculations applied for the combined probability of CSF (P CSF ).
Finally, assuming the introduction of each disease being independent and not mutually exclusive from each other, the final combined probability of introduction of ASF or CSF into the US by legal imports (P C ) was estimated as follows:
Sensitivity analysis.
Sensitivity analyses were performed for all the models previously described in two steps. Firstly, the most influential parameters in each of the models were identified by calculating the regression coefficients (β i ) between each input and the annual probability of ASF/CSF introduction in the US. Afterwards, the inputs that were most likely to influence the final results (β i ≥ 0.1) were analyzed in detail using the advanced sensitivity analysis tool of @RISK 5.5 running 500 iterations for each scenario. A total of 10 scenarios were assessed for each selected parameter, by changing the base values in ten consecutive steps, from a minimum of 50% reduction to a maximum of 50% increase.
Probability of ASF/CSF introduction into the US by legal imports of live pigs.
Considering the current situation of both diseases (as of November 2018), the probability of ASF introduction into the US by legal imports of live pigs was estimated as 3.6*10 −3 (2.0*10 −4 , 1.5*10 −2 ), while the probability of CSF by the same pathway was 2.5*10 −3 (4.4*10 −5 , 1.1*10 −2 ). These values approximately correspond with one outbreak of ASF in 276 years, and one introduction of CSF in 201 years, if conditions remain constant. The overall mean annual probabilities of ASFV being released by imported pigs was 8.4*10 −2 (6.8*10 −3 , 3.2*10 −1 ), whereas for CSFV this probability reached 1.2*10 −1 (2.1*10 −3 , 5.1*10 −1 ). The mean exposure of US domestic pig population to imported pigs (potentially infected either with ASFV or CSFV) pigs was 4.4*10 −2 (1.6*10 −2 , 8.7*10 −2 ).
The distribution of the risk of introductions through imports of live pigs into the US was very similar for both diseases analyzed. Although small differences were found between states for ASF and CSF, when the values were categorized for producing the risk maps, the result was the same, so only one map was presented (Fig 3). The highest probabilities were located in Iowa, Minnesota and Wisconsin, which concentrate 57% of the total probability of both viruses introduction. With 99% of the domestic pigs imported to the US coming from Canada, this was the country of origin that poses the highest risk for both, the introduction of ASF and the introduction of CSF through legal imports of pigs, with annual probabilities of 3.6*10 −3 for ASF, and 4.9*10 −3 for CSF. The monthly disaggregation of the risk revealed that January and March were the months at higher risk in both models, including also February for CSF. However, no important seasonal differences were observed for none of the diseases (data not shown).
Expand Fig 3. Final risk of ASF/ CSF introduction by legal imports of live swine into the US.
The graduated color map represents the final risk (release*exposure) from the highest (darker) to the lowest (lighter).
Probability of ASF/CSF introduction into the US by legal imports of swine products.
The probabilities of both viruses ASFV/CSFV being released into the US by infected swine products were estimated as 7.8*10 −2 (8.3*10 −3 , 2.9*10 −1 ), and 6.9*10 −2 (3.9*10 −3 , 2.8*10 −1 ), respectively. The location of the risks of ASFV /CSFV being released into the US through import of swine products varies between both viruses and from the risk associated with live pigs imports. Specifically, the risk of ASFV release was highest in the states of New Jersey (2.6*10 −2 ), Virginia (1.2*10 −2 ), California (9.5*10 −3 ) and Florida (9.4*10 −3 ). However, the risk of release of CSF potentially infected products was concentrated in Florida (2.6*10 −2 ), Illinois (8.1*10 −3 ) and California (6.9*10 −3 ) (Fig 4). As it can be observed in the map, Florida and California presented relatively high risk for both diseases, either ASF/CSF release.
Expand Fig 4. Risk maps of ASFv and CSFv release by legal imports of swine products.
(A) Risk of ASFv release into the US. (B) Risk of CSFv release in to the US. The graduated color maps represent the risk from the highest (darker) to the lowest (lighter).
A total of 91% of the risk of ASFV being released into the US was originated from Denmark (5.8*10 −2 ) and Poland (1.4*10 −2 ). However, for the introduction of CSF the origins of the risk were wider distributed, being Finland (4.8*10 −2 ), Canada (7.1*10 −3 ), Cayman Islands (7.1*10 −3 ) and Denmark (1.7*10 −3 ) the origins of the highest risk.
The average probability of US swine populations of being exposed to imported products per state was 4.4*10 −3 (4.6*10 −5 , 1.9*10 −2 ), existing important differences between states. This risk of exposure was concentrated in the southern states, including California, mainly due to the presence of feral pigs with potential access to landfills, as well as to the smaller size of swine premises, which present higher risk of using swill feeding (Fig 5). The risk of exposure was almost 10 times higher for the potential access of feral pigs to landfills (2.5*10 −1 ) than the risk associated with swill feeding activities to domestic pigs (2.0*10 −2 ). The waste originated from households contributed with 64% of the total risk of exposure, mainly due to the disposal of 94.5% of the generated waste in landfills. This risk was followed by the waste from the risk associated with retail sector (20%) and finally the restaurants waste (16%).
Expand Fig 5. Risk of exposure to legal imports of swine products.
The graduated color map represents the risk from the highest (darker) to the lowest (lighter) of US susceptible swine populations being exposed to the legally imported swine products.
Although the risk of release was higher for ASF, the final probability of CSF being introduced into the US through infected swine products was almost two times higher than the risk of ASF by this pathway. Specifically, for ASF the final probability was estimated as 4.5*10 −4 (4.2*10 −5 , 1.9*10 −3 ), and 8.3*10 −4 (3.8*10 −5 , 3.5*10 −3 ) for CSF. Those probabilities were approximately eight times (ASF) and 6 times (CSF) lower than the estimated probabilities of introduction through imports of live pigs. The final risk maps for both diseases were quite similar, being in both cases Florida and California the states at highest risk (Fig 6). However, while 90% of the total risk of CSF was concentrated in these states, the risk of ASF was broader distributed. In contrast to the live pig pathway, the ASF/CSF risk associated with the imports of swine products presents certain seasonality being April and October the months at highest risk for both diseases analyzed.
Expand Fig 6. Final risk of ASF and CSF introduction by legal imports of swine products.
The graduated color maps represent the final risk (release *exposure) from the highest (darker) to the lowest (lighter). (A) Risk of ASF introduction into the US by legal imports of swine products. (B) Risk of CSF introduction into the US by legal imports of swine products.
Combined probabilities.
The combined probability of ASF introduction into the US through legal importations (of pigs or products) was estimated as 2.2*10 −3 (1.6*10 −4 , 8.7*10 −3 ). For CSF, the combined probability resulted in 3.3*10 −3 (8.1*10 −5 , 1.5*10 −2 ). The final combined probability, which estimates the risk of ASF or CSF being introduced into the US by legal importations of animals or products was estimated as 5.5*10 −3 (2.4*10 −4 , 2.3*10 −2 ), which is approximately equivalent to one outbreak every 181 years.
Sensitivity analysis.
Based on the correlation coefficient assessment (β i ≥ 0.1), the following inputs were selected for the advanced sensitivity analysis from the pigs imports model: the probability of ASF/CSF infection in Canada (P1 CAN), the probability of selecting an infected pig from Canada (P2 CAN), the probability of survival to ASF/CSF infection (P3), the probability of pigs being undetected during quarantine (P u ) and the probability of imported pigs going through quarantine (P q ). The advanced sensitivity analysis reveals that both models were only noticeable influenced by one parameter: Pq or the probability of imported pigs going through quarantine (Fig 7). The rest of the parameters analyzed didn’t influence substantially the final result, even when they were increased or decreased up to 50%.
Expand Fig 7. Advanced sensitivity analysis for the models of ASF/CSF risk of introduction by pigs imports.
The spider graph plots the percent change of the selected input parameters against the output results.
Five parameters were selected from each legal imports of products model for the advanced sensitivity analysis based on their correlation coefficients. The probability of feral pigs accessing to landfills (FP A ) and the food losses at consumer level (W C ) were selected for both models. In addition, from the ASF model of swine products the probability of selecting infected meat from Denmark (P2 ASF DEN), Canada (P2 ASF CAN) and The Netherlands (P2 ASF NETH) were included. Whereas for the CSF model, the probability of selecting infected meat from Finland (P2 CSF FIN), Cayman Island (P2 CSF CAY) and United Kingdom (P2 CSF UK) were selected (β i ≥ 0.1). Based on the advanced sensitivity analysis, both models were highly influenced by the likelihood of feral pigs accessing to landfills and the food losses at consumer level (Fig 8). Based on the scenarios run on the ASF model, the rest of the parameters assayed have none (P2 ASF CAN and NETH) or very little influence (P2 ASF DEN). In contrast, the probability of selecting pigs from Finland (P2 CSF FIN) highly influenced the CSF model results.
Expand Fig 8. Advanced sensitivity analysis for the models of ASF/CSF risk of introduction by swine products imports.
The spider graph plots the percent change of the selected input parameters against the output results.
Discussion.
Following the OIE guidelines [20], four quantitative risk assessment models were developed to evaluate the risks of ASF and CSF introduction into the US by legal importations of live pigs and products during the high risk period (i. e., time between infection and detection/notification in the source country). The primary aim of import risk analysis is to provide importing countries an objective and defensible method for assessing the risks associated with the importation of animals and animal related products (i. e. genetic material, feedstuffs, biological products and pathological material) [60], as well as to identify hazards and how those hazards could eventually become a risk. Therefore, risk analysis should be always transparent, based on the best available information and fully referenced, as their results can be used to regulate international trade. In this study, when available, only accredited sources of data were used. For those parameters for which no detailed information exists, expert opinion and certain assumptions (assuming the highest risk scenario) were included in the model, being all of them fully described in detail in Tables 2, 3 and 4, ensuring the transparency of the assessment. The intense sensitivity analyses performed in the study allowed us to estimate the impact of these parameters in the final model outcomes, and their relative importance in the final results.
In the present analysis, significant novelties were incorporated to address some of the biggest challenges previously identified in similar studies [61]. For example, the estimation of the risk of infection in the origin countries has been always a critical parameter, difficult to address in the import risk assessments. If the disease is present in the origin countries, the prevalence data is normally used as proxy for the risk of infection in origin [20]. However, when the disease is absent in the countries of origin, estimations based on the time from the last outbreak are sometimes used [61–63], assuming that the occurrence of the diseases is seasonal (i. e. disease outbreaks will occur every X years). This assumption could be valid for some vector-borne diseases highly dependent on climatic factors (i. e. Rift Valley Fever). However, it is not valid for the two diseases analyzed here (ASF and CSF), as the outbreaks in new affected areas are commonly initiated by movement of infected animals or contaminated material.
Therefore, in this study we used a Susceptible-Infected compartmental model, assuming that the risk of a country to become infected by ASF/CSF strongly depends on their commercial relationships with currently infected countries. We are aware that this approach only addresses one of the routes of introduction of animal diseases in free countries, and other important sources of infection were not considered (i. e. movement of wild animals, illegal imports, etc.). However, the lack of detailed data and controversy associated with the estimation of those other pathways make these estimations very difficult and require to be supported by a series of risk assessments of each of the source countries, which was out of the scope and possibilities of this study. Therefore, although is not the perfect solution, it was considered that this approach provides more realistic information than traditional estimates, and could be easily applied in other risk assessment models.
Another strength of this study was the integration of the release and exposure assessments in one single stochastic model for the assessment of the risk associated with imported products. The integration of both steps in a common model is not common, as in most of the cases, the quality of information are not comparable and each step is assessed separately (i. e. release and exposure in [64]). In addition, this is the first time that the risk of exposure of US swine populations through swill feeding and landfill disposal are numerically compared. Around one third of the food produced in the world for human consumption (approximately 1.3 billion tons) gets lost or wasted every year [65]. Innovative approaches for promoting food waste reduction are necessary, as well as new options to ensure the correct disposal of food waste, maximizing the profitability of proteins. One of the options for that is swill feeding (which is mostly used to feed swine). Although the practice is controversial as it could involve serious health risks, as demonstrated with the introduction of ASF in Georgia in 2007 [7] or the spread of many other TADs, swill feeding has been historically largely used as it is a way to maximize the use of the food.
If done in controlled manner (with appropriate heat treatment, regulations and controls in place), our model suggests that it implies less risks than disposing the food waste in landfills.
In our model, the risk associated with swill feeding practices was mostly related with the food waste generated in households (64% of the risk of swill feeding was associated with waste from households vs retail and restaurants). Opposite to EU where swill feeding is totally banned, many states in the US allow swill feeding to swine under certain conditions (properly heat-treated and fed by a licensed facility) [66]Therefore, in order to reduce the risk of exposure and optimize the use of meat, swill feeding practices should be promoted in a controlled manner and communication and informative campaigns should be performed in small swine premises or swine hobby farmers.
Most of the food waste generated in our houses ends in landfills (94.5%) [42], and many of those landfills are located in areas are inhabited by feral pigs, that potentially have access and get in contact with the material discarded on them. Our results indicate that the potential risk of exposure of feral pigs to food waste present in open landfills is not negligible (2.5*10 −2 ), posing a higher risk compared to other exposure pathways analyzed here as the swill feeding activities in domestic pigs. In addition, although not included in this study, feral pig could potentially have access to the food waste in bins and containers, before it arrives to the landfills, as well as contacting domestic pigs. The contacts and potential transmission of pathogens between domestic pigs and feral pigs has been documented in certain areas of the US [67] and pose a risk for disease controlling. In these models this pathway of exposure was not included, due to the characteristics of swine pigs imported into the US (99.9% are feeder pigs coming from Canada [38] with destination to large feeding units with no outdoor access). However, it would be an interesting research area to explore for other scenarios. The sensitivity analysis revealed that the probability of feral pigs accessing landfills (which was a high risk scenario assumption) and the proportion of food waste at consumer level highly influenced our results. Therefore, we strongly recommend further research efforts to evaluate the potential access of feral pigs to landfills and other sources of food waste, and estimate the consequences derived from it, as this could be a risk not only for ASF and CSF but for many other diseases/ health issues that could be transmitted through food waste (i. e. E. coli, antimicrobial resistance, toxins, etc.).
Despite the global burden situation of ASF and CSF, and the assumptions used in the model, the risk of introduction of either ASF or CSF into the US by legal import of animal and products is considerably low (combined probability of 5.5*10 −3 ). However, it is important to remember that other pathways should be analyzed such as the importation of other biological products (semen, ova, etc.), animal feed and, importantly, illegal pathways, including the waste from planes and other international transports, in order to have a more complete picture of the risk of ASF and CSF introduction into the US. Among the four pathways analyzed, the imports of pigs posed the highest risk for the introduction of CSF into the US. This could be caused, firstly, due to the fact that CSF is more widely distributed in the world, and the probabilities of CSF infection in origin derived from the SI model were higher comparing with ASF. Secondly, the risk associated with the legal import of swine products was lower in both diseases (4.5*10 −4 for ASF and 3.8*10 −4 for CSF) compared to the live swine imports (3.6*10 −3 for ASF and 4.9*10 −3 for CSF), potentially due to the lower risk of exposure to the imported products, as they usually go directly to human consumption. However, the risk of ASF and CSF released by imports of animal products was considerable higher (0.1), which reflects the importance of controlling the exposure to these products to avoid potential outbreaks of these diseases in the US.
As it was expected, the risk of ASF and CSF introduction by legal imports of pigs was concentrated in the US pork production states, which are the main importers of pigs in the country as Iowa (35%), Minnesota (12%) and Wisconsin (10%). The models presented here estimated the existing risk associated with exporting animals before the detection in the country of origin (high risk period), assuming that US will not accept importation of swine from ASF or CSF infected countries. Canada is an ASF/CSF free country but the risk rises with increasing the volume of commodity imported. According to FAOSTAT [24], in the past Canada has had trade of live pigs with non-free countries such as Russia (ASF/CSF), Colombia (CSF), Republic of Korea (CSF), China (CSF), Poland (ASF), Peru (CSF), Italy (ASF) and Philippines (CSF). The strong connections between both markets implies that the swine disease situation in Canada is crucial to the US swine market, for ASF and CSF and any other infectious disease potentially appearing. Consequently, as soon as any change occur in swine status in Canada, the US should re-evaluate its risk levels, and consider the implementation of preventive measures.
The probability of domestic pigs in the US being exposed to a potentially infected (either with ASFV or CSFV) imported swine was estimated as 4.35*10 −2 . However, the results of the sensitivity analysis identified the probability of imported pigs going through quarantine (Pq) as the essential parameter in the model. Therefore, the correct application of quarantine procedures is an essential component for maintaining the free status and reduce the risks associated with animal importations.
For the legal importation of live pigs, no differences were found on the location of the risk between the analyzed diseases (ASF or CSF). However, for the risk associated with the legal imports of swine products, the location of the risk substantially varies between ASF and CSF, due to the trade differences between the states. Whereas for ASF the probability of release was highest in New Jersey (33%), Florida (14%), California (13%) and Virginia (11%); for CSF the risk of release was concentrated in Florida (38%), Illinois (12%), California (10%), and Texas (8%) (Fig 5). The probability of exposure to swine products was concentrated in the southern states and California (Fig 6) due to the abundance of feral pigs that could access to landfills and usually smaller pig farms that present higher probability of using swill feeding [47, 48]. As a result of the combination of release and exposure risks, the final risk of ASF/CSF being introduced into the US through legal products importations is highly concentrated in the states of Florida (38% for ASF and 58% of CSF risk), California (39% of ASF risk and 16% of CSF) and Texas (11% of ASF risk vs 16% CSF risk). In this pathway, although Canada was again the highest risk country for the origin of these products, other countries presented also a relative high importance in this pathway including Denmark and Poland for ASF. The imports from Poland only represents 3.4% of the total products imported. However, it was the second highest risk origin country due to the presence of ASF in the country since February 2018. Interestingly, the imports of swine products from Poland have not been stopped for ASF, but continue gradually increasing. In this case of an already infected country, the preventive measures should be focused on analyzing and ensuring the freedom from disease in the units approved for exports. Additional checks of biosecurity compliance in origin farms, regionalization procedures, or even periodic diagnostic testing of the swine products in origin would be also recommended. In the case of CSF introduction, surprisingly Finland was one of the highest risk, potentially due to the swine trade connections with China, Russia, Estonia and Poland [24]. Indeed, the probability of selecting infected swine products from Finland was one of the most influential parameters of the model for CSF risk of imported products (Fig 8).
Although the consequences of the introduction of ASF and CSF into the US were not considered in this work, the example of previous outbreaks in other free territories suggest a potential huge economic impact. Presumable, the impact would be more serious in the case of ASFV introduction, as no vaccines are available for its control. However, in the case of CSF, there are highly efficacious and safe vaccines available, including a bait format for oral immunization which could be used in wild boar and even in domestic pigs in backyard conditions [17].
Quantitative risk assessments present certain benefits and advantages versus qualitative models, as they incorporate the variability of data, uncertainty of the estimations and evaluate the influence of the parameters in the final outcome through sensitivity analysis. However, this type of analysis results in very time intensive process due to the substantial efforts for collecting all the data required plus the computational requirements. For example, in this study, more than 15,000 inputs were used in the models for the legal importations of pig products. On the other hand, once the models’ structures are defined, data sources identified and available, the models can be easily updated. The studies developed here not only served to estimate the probabilities of ASF and CSF introduction through legal imports, but will constitute the basis for the documentation and quantitative analysis of the risk of other FADs into the US, that would help to prevent the negative consequences associated with these types of diseases. The epidemiological information obtained in the present study could be used to develop risk-based surveillance, prevention and early detection strategies that would help to prevent ASF/CSF introduction and protect US swine livestock and consumers as well as allocate resources effectively and efficiently. In addition, the model parameters and calculations of the present study will be integrated in an online user-friendly risk assessment platform (Disease BioPortal™ accessible from bioportal. ucdavis. edu/) for the easy update and visualization of the risk estimates. Other pathways of introduction such as the illegal importation of pork and other swine products into the US by the different ports of entry (airports, maritime, mail, etc.) are currently being addressed and will be also incorporated in the platform.
استنتاج.
Four quantitative stochastic risk assessments models were developed to estimate the risk of ASF/CSF introduction into the US through legal importation of live swine and swine products during the high risk period. The models’ results suggest that the risk of both diseases being introduced into the US through the analyzed pathways was very low, being the risk of CSF by legal imports of pigs the analyzed pathway that poses the highest risk. The risk of introduction through live swine imports is higher for CSF than ASF, being for both diseases concentrated in the pork production states (Iowa, Minnesota and Wisconsin), with most of the live pigs coming from Canada. In contrast, the final risk of ASF/CSF introduction for the products model was concentrated in states of California, Florida and Texas. However, the risk of entrance of potentially infected swine products into the US clearly differs between CSF and ASF. The epidemiological information obtained in the present study could be useful to develop prevention and early detection strategies that would help to prevent ASF/CSF introduction as well as allocate resources effectively and efficiently.
Acknowledgments.
This material is based upon work supported by Kansas Bioscience Authority under the Federal Matching Program Grant Agreement through the Center of Excellence for Emerging and Zoonotic Animal Diseases (CEEZAD) and by the State of Kansas, National Bio and Agro-defense Facility (NBAF) Transition Fund.
Darla Quijada was able to develop her research work thanks to the "Institute for Infectious Animal Diseases Undergraduate Internship Program”.
Special thanks to all the contributors on the qualitative pathway assessment studies that facilitate this work: Mollie Burton, Joe Fund, Marie Keith, Diane Larson Jonathan Miller; Loren Riccioni, Lauren Sawyer and Adrian Self.
المراجع.
1. OIE. World Organisation for Animal Health 2018. Available from: oie. int/animal-health-in-the-world/oie-listed-diseases-2018/. 2. Sánchez-Vizcaíno JM, Arias M. African swine fever. In: Zimmerman J, Karriker LA, Ramirez A, Schwartz K, Stevenson G, editors. Diseases of swine. 10 ed. Ames, Iowa: Blackwell Publishing Professional; 2018. p. 396–404. 3. ICTV. The International Committee on Taxonomy of Viruses 2018. Available from: ictvonline/virusTaxonomy. asp. 4. Plowright W, Parker J, Pierce MA. African Swine fever virus in ticks (Ornithodoros moubata) collected from animal burrows in Tanzania. Nature. 1969;221:1071–3. pmid:5813153 5. Costard S, Mur L, Lubroth J, Sanchez-Vizcaino JM, Pfeiffer DU. Epidemiology of African swine fever virus. Virus Res. 2018;173(1):191–7. doi: 10.1016/j. virusres.2018.10.030 pmid:23123296. 6. Montgomery RE. On a form of swine fever occurring in British East Africa. J Comp Pathol. 1921;34:59–191. 7. Rowlands RJ, Michaud V, Heath L, Hutchings G, Oura C, Vosloo W, et al. African swine fever virus isolate, Georgia, 2007. Emerging infectious diseases. 2008;14(12):1870–4. Epub 2008/12/03. doi: 10.3201/eid1412.080591 pmid:19046509; PubMed Central PMCID: PMCPMC2634662. 8. Gogin A, Gerasimov V, Malogolovkin A, Kolbasov D. African swine fever in the North Caucasus region and the Russian Federation in years 2007–2018. Virus research. 2018;173(1):198–203. doi: 10.1016/j. virusres.2018.12.007 pmid:23266725. 9. Sanchez-Vizcaino JM, Mur L, Gomez-Villamandos JC, Carrasco L. An update on the epidemiology and pathology of African swine fever. Journal of comparative pathology. 2018;152(1):9–21. Epub 2018/12/03. doi: 10.1016/j. jcpa.2018.09.003 pmid:25443146. 10. Sánchez-Vizcaíno J, Mur L, Sánchez-Matamoros A, Martínez-López B. African Swine Fever: new challenges and measures to prevent its spread 2018. Available from: oie. int. 11. Saatkamp HW, Berentsen PB, Horst HS. Economic aspects of the control of classical swine fever outbreaks in the European Union. Veterinary microbiology. 2000;73(2–3):221–37. Epub 2000/04/28. pmid:10785330. 12. Moennig V, Floegel-Niesmann G, Greiser-Wilke I. Clinical signs and epidemiology of classical swine fever: a review of new knowledge. Veterinary journal. 2003;165(1):11–20. pmid:12618065. 13. Luo Y, Li S, Sun Y, Qiu HJ. Classical swine fever in China: a minireview. Veterinary microbiology. 2018;172(1–2):1–6. Epub 2018/05/06. doi: 10.1016/j. vetmic.2018.04.004 pmid:24793098. 14. OIE WAHIS. OIE WAHIS Interface 2018. Available from: oie. int/wahis_2/public/wahid. php/Wahidhome/Home. 15. USDA Economic Research Service. Hogs & Pork 2018 [November 3 2018]. Available from: ers. usda. gov/topics/animal-products/hogs-pork/. 16. Madera R, Gong W, Wang L, Burakova Y, Lleellish K, Galliher-Beckley A, et al. Pigs immunized with a novel E2 subunit vaccine are protected from subgenotype heterologous classical swine fever virus challenge. BMC veterinary research. 2018;12(1):197. Epub 2018/09/11. doi: 10.1186/s12917-016-0823-4 pmid:27612954; PubMed Central PMCID: PMCPMC5016919. 17. Rossi S, Staubach C, Blome S, Guberti V, Thulke HH, Vos A, et al. Controlling of CSFV in European wild boar using oral vaccination: a review. Front Microbiol. 2018;6:1141. Epub 2018/11/12. doi: 10.3389/fmicb.2018.01141 pmid:26557109; PubMed Central PMCID: PMCPMC4615961. 18. Costard S, Wieland B, de Glanville W, Jori F, Rowlands R, Vosloo W, et al. African swine fever: how can global spread be prevented? Philos Trans R Soc Lond B Biol Sci. 2009;364(1530):2683–96. Epub 2009/08/19. doi: 10.1098/rstb.2009.0098 pmid:19687038; PubMed Central PMCID: PMCPMC2865084. 19. Martinez-Lopez B, Perez AM, Sanchez-Vizcaino JM. A stochastic model to quantify the risk of introduction of classical swine fever virus through import of domestic and wild boars. Epidemiology and infection. 2009;137(10):1505–15. Epub 2009/02/27. doi: 10.1017/S0950268808001623 pmid:19243649. 20. OIE World Organisation for Animal Health. Handbook on Import Risk Analysis for Animals and Animal Products. Quantitative Risk Assessment. France: OIE; 2004. 126 p. 21. OIE World Organisation for Animal Health. Handbook on Import Risk Analysis for Animals and Animal Products. Introduction and qualitative risk analysis. France: OIE; 2018. 88 p. doi: 10.1111/j.1539-6924.2018.01386.x. 22. USDA Global Agricultural Trade System. 2018. Available from: https://apps. fas. usda. gov/gats/default. aspx. 23. US Census Bureau Trade Online. 2018. Available from: https://usatrade. census. gov/. 24. FAOSTAT. 2018. Available from: faostat. fao/. 25. de Carvalho Ferreira HC, Backer JA, Weesendorp E, Klinkenberg D, Stegeman JA, Loeffen WL. Transmission rate of African swine fever virus under experimental conditions. Veterinary microbiology. 2018;165(3–4):296–304. Epub 2018/05/15. doi: 10.1016/j. vetmic.2018.03.026 pmid:23664069. 26. Stegeman JA, Elbers AR, Boum A, de Jong MC. Rate of inter-herd transmission of classical swine fever virus by different types of contact during the 1997–8 epidemic in The Netherlands. Epidemiology and infection. 2002;128(2):285–91. Epub 2002/05/11. pmid:12002547; PubMed Central PMCID: PMCPmc2869822. 27. R Core Team. R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. 2018. 28. Csardi G, Nepusz T. The igraph software package for complex network research, InterJournal, Complex Systems 2006. Available from: igraph. 29. Butts C. SNA: Tools for Social Network Analysis. R package version 2.4. 2018. Available from: https://CRAN. R-project/package=sna. 30. Martínez-López B, Perez A, Sánchez-Vizcaíno J. A stochastic model to quantify the risk of introduction of classical swine fever virus through import of domestic and wild boars. Epidemiology and infection. 2009;137(10):1505–15. doi: 10.1017/S0950268808001623. pmid:19243649 31. Spickler AR, Roth JA. Peste porcina africana. Iowa State University, College of Veterinary Medicine 2006. Available from: cfsph. iastate. edu/Factsheets/es/peste_porcina_africana. pdf. 32. National Agricultural Biosecurity Center. Pathways Analysis of Classical Swine Fever (CSF) risk to the United States 2004. Available from: k-state. edu/nabc/docs/2004-PA-CSF. pdf. 33. Murray AC, Johnson CP. Impact of the halothane gene on muscle quality and pre-slaughter deaths in Western Canadian pigs. Can J Anim Sci. 1998;78:543–8. 34. Animal and Plant Health Inspection Service USDA § 93.503. 1990. Available from: https://gpo. gov/fdsys/pkg/CFR-1998-title9-vol1/pdf/CFR-1998-title9-vol1-sec93-503.pdf. 35. Romero L. Spanish Ministry of Agriculture, MAPA, personal communication. 36. Eurostat. Pig farming sector—statistical portrait 2018. Available from: ec. europa. eu/eurostat/statistics-explained/index. php/Pig_farming_sector_-_statistical_portrait_2018. 37. Su C-L. BetaBuster. Available from: cadms. ucdavis. edu/diagnostictests/betabuster. html. 38. USDA USDoA-. Hogs & Pork trade 2017. Available from: https://ers. usda. gov/topics/animal-products/hogs-pork/trade/#trade. 39. Agency). USEEP. Food Waste Management in the United States, 2018. 2018. 40. Office USEEPAR. Summary Analysis of Massachusetts Commercial/Institutional Food Waste Generation Data. 2018. 41. BSR. Analysis of U. S. Food Waste Among Food Manufacturers, Retailers, and Restaurants. 2018. 42. United States Environmental Protection Agency. Postconsumer Food Diverted Through Donation, Animal Feed, Anaerobic Digestion, and Composting for 2018. 2018. 43. USDA APHIS. Swine Health Protection; Feeding of Processed Product to Swine. [Docket No APHIS–2008–0120]. 2009;74(63). 44. Hernandez-Jover M, Schembri N, Holyoake PK, Toribio JL, Martin PA. A Comparative Assessment of the Risks of Introduction and Spread of Foot-and-Mouth Disease among Different Pig Sectors in Australia. Frontiers in veterinary science. 2018;3:85. Epub 2018/10/08. doi: 10.3389/fvets.2018.00085. pmid:27713881 45. United States Environmental Protection Agency. Municipal Solid Waste Generation, Recycling, and Disposal in the United States: Facts and Figures for 2018. EPA United States Environmental Protection Agency. 2018;EPA-530-F-14-001:13. 46. United States Environmental Protection Agency. Food waste scoping analysis. 2018. 47. USDA APHIS NAHMS. Swine 2018 Part I: Baseline Reference of Swine Health and Management in the United States, 2018 2018. Available from: https://aphis. usda. gov/animal_health/nahms/swine/downloads/swine2018/Swine2018_dr_PartI. pdf. 48. USDA APHIS NAHMS. Small-Enterprise Swine 2007. Reference of Management Practices on Small-Enterprise Swine Operations in the United States, 2007 2007. Available from: https://aphis. usda. gov/animal_health/nahms/swine/downloads/swine2007/Swine07_dr_SmallSwine. pdf. 49. USDA. Feral Swine populations 2018 by county. Southeastern Cooperative Wildlife Disease Study 2018. Available from: swine. vet. uga. edu/nfsms/information/map2018.htm. 50. United States Environmental Protection Agency. LMOP Database 2018. Available from: https://epa. gov/lmop. 51. Buzby J, Bentley J, Padera B, Campuzano J, Ammon C. Updated Supermarket Shrink Estimates for Fresh Foods and Their Implications for ERS Loss-Adjusted Food Availability Data. Economic Information Bulletin, 2018. 52. Buzby J, Wells F, Axtman B, Mickey J. Supermarket Loss Estimates for Fresh Fruit, Vegetables, Meat, Poultry, and Seafood and Their Use in the ERS Loss-Adjusted Food Availability Data. Economic Information Bulletin, 2009. 53. FAO. Global food losses and food waste–Extent, causes and prevention. 2018. 54. Buzby J, Wells H, Hyman J. The Estimated Amount, Value, and Calories of Postharvest Food Losses at the Retail and Consumer Levels in the United States. Economic Information Bulletin, 2018. 55. USDA. Agricultural Research Service. 1994–96 Continuing Survey of Food Intakes by Individuals and 1994–96 Diet and Health Knowledge Survey. 2000. 56. Davis C, Lin B. Factors Affecting U. S. Pork Consumption. 2005. 57. USDA APHIS. Rules and Regulations. 2009. 58. USDA Economic Research Service. ARMS Farm Financial and Crop Production Practices 2018. Available from: ers. usda. gov/data-products/arms-farm-financial-and-crop-production-practices/tailored-reports-farm-structure-and-finance. aspx#P5c4799ca4f61405682bd0c34bee3c10f_25_66iT0R8T0R0x0. 59. Environmental Systems Research Institute (ESRI). ArcGIS Release 10.0. Redlands, CA.2018. 60. Welte V. Risk Analysis and OIE 2018. Available from: fao/docrep/003/x7354e/x7354e12.htm. 61. Mur L, Martinez-Lopez B, Martinez-Aviles M, Costard S, Wieland B, Pfeiffer DU, et al. Quantitative risk assessment for the introduction of African swine fever virus into the European Union by legal import of live pigs. Transboundary and emerging diseases. 2018;59(2):134–44. Epub 2018/08/13. doi: 10.1111/j.1865-1682.2018.01253.x pmid:21831148. 62. Mur L, Martinez-Lopez B, Costard S, de la Torre A, Jones BA, Martinez M, et al. Modular framework to assess the risk of African swine fever virus entry into the European Union. BMC veterinary research. 2018;10:145. Epub 2018/07/06. doi: 10.1186/1746-6148-10-145 pmid:24992824; PubMed Central PMCID: PMCPMC4112856. 63. Sanchez-Vizcaino F, Perez A, Lainez M, Sanchez-Vizcaino JM. A quantitative assessment of the risk for highly pathogenic avian influenza introduction into Spain via legal trade of live poultry. Risk analysis: an official publication of the Society for Risk Analysis. 2018;30(5):798–807. Epub 2018/02/09. doi: 10.1111/j.1539-6924.2009.01351.x pmid:20186740. 64. Costard S, Jones BA, Martinez-Lopez B, Mur L, de la Torre A, Martinez M, et al. Introduction of African swine fever into the European Union through illegal importation of pork and pork products. PloS one. 2018;8(4):e61104. Epub 2018/04/25. doi: 10.1371/journal. pone.0061104 pmid:23613795; PubMed Central PMCID: PMCPmc3627463. 65. FAO. Save food: Global Initiative on Food Loss and Waste Reduction 2018. Available from: fao/save-food/resources/keyfindings/en/. 66. Broad Leib E, Balkus O, Rice C, Maley M, Taneja R, Cheng R, et al. Leftovers for livestock: A Legal Guide for Using Food Scraps as Animal Feed 2018. Available from: chlpi/wp-content/uploads/2018/12/Leftovers-for-Livestock_A-Legal-Guide_August-2018.pdf. 67. Wyckoff AC, Henke SE, Campbell TA, Hewitt DG, VerCauteren KC. Feral swine contact with domestic swine: a serologic survey and assessment of potential for disease transmission. Journal of Wildlife Diseases. 2009;45(2):422–9. doi: 10.7589/0090-3558-45.2.422. pmid:19395751.
PLOS is a nonprofit 501(c)(3) corporation, #C2354500, and is based in San Francisco, California, US.

PLOS ONE.
Submissions Getting Started Submission Guidelines Figures Tables Supporting Information LaTeX Revising Your Manuscript Submit Now Policies Best Practices in Research Reporting Human Subjects Research Animal Research Competing Interests Disclosure of Funding Sources Licenses and Copyright Data Availability Materials and Software Sharing Ethical Publishing Practice Authorship Downloads and Translations Manuscript Review and Publication Criteria for Publication Editorial and Peer Review Process Reviewer Guidelines Accepted Manuscripts Corrections and Retractions Comments Article-Level Metrics.
Submit Your Manuscript.
Discover a faster, simpler path to publishing in a high-quality journal. PLOS ONE promises fair, rigorous peer review, broad scope, and wide readership – a perfect fit for your research every time.
Click through the PLOS taxonomy to find articles in your field.
For more information about PLOS Subject Areas, click here.
Regression Model-Based Walking Speed Estimation Using Wrist-Worn Inertial Sensor.
Affiliation School of Mechatronic Systems Engineering, Simon Fraser University, 250–13450 102nd Avenue, Surrey, BC, V3T 0A3, Canada.
Affiliation School of Mechatronic Systems Engineering, Simon Fraser University, 250–13450 102nd Avenue, Surrey, BC, V3T 0A3, Canada.
Regression Model-Based Walking Speed Estimation Using Wrist-Worn Inertial Sensor.
Shaghayegh Zihajehzadeh, Edward J. Park.
Published: October 20, 2018 https://doi/10.1371/journal. pone.0165211.
نبذة مختصرة.
Walking speed is widely used to study human health status. Wearable inertial measurement units (IMU) are promising tools for the ambulatory measurement of walking speed. Among wearable inertial sensors, the ones worn on the wrist, such as a watch or band, have relatively higher potential to be easily incorporated into daily lifestyle. Using the arm swing motion in walking, this paper proposes a regression model-based method for longitudinal walking speed estimation using a wrist-worn IMU. A novel kinematic variable is proposed, which finds the wrist acceleration in the principal axis (i. e. the direction of the arm swing). This variable (called pca-acc ) is obtained by applying sensor fusion on IMU data to find the orientation followed by the use of principal component analysis. An experimental evaluation was performed on 15 healthy young subjects during free walking trials. The experimental results show that the use of the proposed pca-acc variable can significantly improve the walking speed estimation accuracy when compared to the use of raw acceleration information ( p <0.01). When Gaussian process regression is used, the resulting walking speed estimation accuracy and precision is about 5.9% and 4.7%, respectively.
Citation: Zihajehzadeh S, Park EJ (2018) Regression Model-Based Walking Speed Estimation Using Wrist-Worn Inertial Sensor. PLoS ONE 11(10): e0165211. https://doi/10.1371/journal. pone.0165211.
Editor: Houbing Song, West Virginia University, UNITED STATES.
Received: May 17, 2018; Accepted: October 7, 2018; Published: October 20, 2018.
Copyright: © 2018 Zihajehzadeh, Park. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The experimental protocol of this study (No. 2018s0750) was approved by the Office of Research Ethics of Simon Fraser University. This ethical restriction prohibits the authors from making the minimal data set publicly available. However, the anonymized data are available to all interested researchers upon request. Interested readers may contact the Office of Research Ethics of Simon Fraser University (dore@sfu. ca) and Dr. Edward J. Park (ed_park@sfu. ca) to request data.
Funding: This work was fully funded by the Natural Sciences and Engineering Research Council of Canada (nserc-crsng. gc. ca/): SPG/430592-2018, and Vanier Canada Graduate Scholarship. The funder provided support in the form of salaries for the first author [SZ] and research materials, but did not have any additional role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. The specific roles of these authors are articulated in the ‘author contributions’ section.
Competing interests: I have read the journal’s policy and the authors of this manuscript have the following competing interest: Provisional US Patent Application No. 62/344,566 filed June 2, 2018 for Systems and Methods for Walking Speed Estimation. We confirm that our provisional patent does not alter our adherence to PLOS ONE policies on sharing data and materials.
المقدمة.
Walking speed is widely used to study human health status. Based on previous studies, walking speed can be used as a marker of mild cognitive impairment (MCI) [1–3]. For example, the trajectories of weekly walking speed and the coefficient of variation of the walking speed are shown to be among the most important parameters for the early detection of MCI in older adults [1]. In addition to MCI, walking speed can also be used as a marker of multiple sclerosis (MS) [4], Parkinson’s disease [5, 6], risk of falls [7], kidney disease [8], and adverse outcomes in aging [9]. Hence, it can be considered as a powerful predictor of hospitalization, disability, and survival [10, 11]. In a clinical setting, different protocols including the 4-meter [12], 10-meter [8], and 6-minute walking tests [13] and the timed up and go (TUG) test [9, 14] have been used as standard tools to evaluate walking speed and gait parameters. However, the short walking tests (e. g. the 10-meter walking test) are subject to bias due to their brevity [15] and the longer tests are less accepted due to the space and time constraints in clinical exams [16]. Additionally, the walking speed results of clinical tests cannot be fully applied to the free-living environment [17]. Furthermore, building precise pathological models of some disease for development of monitoring and treatment guidelines, requires access to longitudinal measurements [18,19]. This emphasizes the need for a reliable system/method for longitudinal (i. e. over long periods of time) and continuous walking speed measurement in real-world situations.
Aiming at longitudinal walking speed measurement outside the clinical setting, some researchers have used passive infrared (PIR) motion sensors. These PIR sensors can be mounted on ceiling [20] or walls [21] of a residence and can measure the individuals’ walking speed when they are in the field of view of the sensors. However, walking speed measurement based on PIR sensors is limited to confined areas such as hallways. Additionally, such system cannot differentiate between multiple residents, limiting its application to independent-living resident homes. Camera-based systems have also been used in the literature for in-home gait speed measurement [22]. However, camera-based systems can get affected by the lighting conditions, and similar to the PIR sensors, they are limited to confined areas and hence more suitable for clinical settings.
Fortunately, with recent advances in MEMS technology and wireless sensor networks, wearable inertial measurement units (IMU) have emerged as powerful devices for portable human motion analysis [23–29]. Being self-contained, wearable inertial sensors can facilitate walking speed measurement in an ambulatory fashion. Considering that the acceleration data from tri-axial accelerometer in a wearable inertial sensor can be integrated to get the velocity, integration-based approaches have been widely used for speed tracking [30]. The main challenge in integration-based approaches is the velocity drift over time that happens as a result of time-varying bias in MEMS-based inertial sensors [31]. To mitigate the drift, some researchers have proposed the detection of periodic foot stance phases during walking to reset the velocity to zero through a process called zero velocity update (ZUPT) [30–34]. However, the need for foot-stance detection requires the wearable sensor to be normally mounted on the leg (ideally on the foot), which is inconvenient for longitudinal walking speed monitoring, particularly indoors. Using waist-worn IMU, some studies have modeled the foot swing in walking as an inverted pendulum to find a 3D walking kinematic model for speed estimation [35]. Additionally, using a waist-mounted IMU, linear and nonlinear regression models have shown promising performances for ambulatory walking [13, 36–37] and swimming [38] speed estimation. These regression-based approaches for walking speed estimation are based on mapping the inherent pattern of acceleration and rate of turn information corresponding to the hip rotation in a gait cycle to walking speed.
For longitudinal health status monitoring, among the available state-of-the-art inertial sensing-based wearables, wrist-worn devices are the most user-friendly and compliant that do not limit the freedom of movement and do not require specific dressing style (e. g. wearing a belt in the case of waist-worn sensor). Thus, wrist-worn devices have relatively higher potential to be easily incorporated into daily lifestyle and worn for longer hours. Similar to hip rotation in each gait cycle [13], arm swing motion during walking is a periodic motion pattern that is highly correlated to walking speed: the faster the walking speed, the faster the arm swing motion. However, in walking speed estimation based on regression models, free arm motion necessitates the use of more complex algorithms to manipulate the acceleration and rate of turn information and get a variable that is more representative of the arm swing motion. Although extracting this variable is of high importance (because the accuracy of the regression models depends on the set of chosen variables and the extracted features), it has not been addressed in the existing literature.
Aiming at accurate walking speed estimation using a wrist-worn IMU, this paper provides a novel processing method based on combined inertial sensor fusion and principal component analysis (PCA) for variable extraction. Experimental results show that the extracted variable can improve the accuracy of wrist-based walking speed estimation.
Theoretical Method.
Problem Definition.
Considering that walking is represented by a set of features, this section is focused on formulating a mapping from walking-related features (predictors) to walking speed (response value) using a regression model. In a regression problem, a training set ( ) consisting of N - number of D - dimensional predictors x i and noisy observations of the response value y i is given ( ). The goal of a regression model is to find the best-fit function f ( x i ) that predicts the response values. The Gaussian process regression (GPR) and regularized least squares regression using least absolute shrinkage and selection operator (LSR-Lasso) models are the two candidate regression methods used in this study.
Gaussian Process Regression.
The objective of GPR, a well-known non-parametric regression technique, is to model the dependency as follows [39]: (1) where ε i = N (0, σ n 2 ) is the independent and identically, normally distributed noise terms. GPR has two main advantages compared to conventional regression methods [40]:
It is a non-parametric regression method and the model structure is determined from data. It uses a probabilistic approach that can model the prediction uncertainty.
A Gaussian process is completely identified by its mean μ ( x i ) and covariance function Σ ( x i , x j ). The covariance function used here is a parameterized squared exponential (SE) covariance function [39]: (2) where σ f is the signal variance and W = diag ( l 1 , …, l D ) is the diagonal matrix of length-scale parameters. This covariance function implements automatic relevance determination (ARD) as the length-scale values determine the effect of each predictor on the regression.
GPR is chosen herein because of its superior performance compared to other regression models in [36] where a waist-worn IMU is used to estimate walking speed.
Regularized Least Squares Regression Using Lasso.
Lasso is the shrinkage and selection method for regularized linear regression. LSR-Lasso, a well-known parametric regression technique, minimizes the usual sum of squared errors, with a bound on the sum of the absolute values of the coefficients to deliver a sparse solution, i. e. a set of estimated regression coefficients in which only a small number is non-zero [41]. Given a linear regression, the LSR-Lasso solves the ℓ 1 - penalized regression to minimize [41]: (6) for unknown parameters β 0 and β = [ β 1 , …, β D ]. The second term in Eq (6) is the penalty function balancing the fit of the model with its complexity with the non-negative parameter λ governing this trade-off [41]. The value of λ is chosen based on 10-fold cross-validation in this study.
LSR-Lasso is chosen in this study to provide a performance baseline for GPR with the SE-ARD covariance function.
Experimental Method.
Participants.
Fifteen young (nine males, six females) self-reported healthy students from Simon Fraser University participated in this study. The participants had an average age of 27±4 years, average height of 1.69±0.08 m, average weight of 6510±10 kg, and average body mass index (BMI) of 23.07±2.3 kg/m 2 . Informed written consent was obtained from the participants and the experimental protocol (No. 2018s0750) was approved by the Office of Research Ethics of Simon Fraser University.
Hardware and Experimental Protocol.
Raw inertial and magnetic data are collected from tri-axial accelerometers, gyroscopes, and magnetometers at the rate of 100 Hz. The sensor is Xsens MTw worn by human subjects on the wrist (Fig 1). Each subject is asked to walk for a distance of 30 m in indoor environment for three different self-selected walking speed regimes: slow, normal, and fast. The subjects are asked to keep their walking speed constant during each 30 m trial and each trial is repeated four times per chosen speed regime, resulting in 12 trials per subject. To get the ground truth average walking speed, the floor is divided into three segments of 10 m long (accurately measured by a laser distance measuring tool with sub-centimeter accuracy), as shown in Fig 1, and the time it takes for the subject to pass each segment is measured using a stopwatch with an accuracy of 0.01 s. The criterion of line passage is when the subject’s right foot passes the line and a human observer always walked with the subject to ensure a perfect sagittal plane view.
Left: a subject wearing MTw units; right: MTw unit and schematic of the test field.
For the purpose of demonstrating and further evaluating the performance of the proposed walking speed estimation method in a real-world setting, five subjects (four males, one female) are asked to perform a 12-min outdoor walking trial that includes 2 min of fast walking, 4 min of normal walking, and 6 min of slow walking. In these outdoor trials, Xsens MTi-G-700 [consisting of tri-axial accelerometers, tri-axial gyroscopes, tri-axial magnetometers, and the Global Positioning System (GPS)] is worn by the subjects on their left wrist and the reference walking speed is obtained by GPS/IMU fusion using our existing Kalman filter-based fusion algorithm previously presented in [31]. Compared to the indoor trials, these outdoor trials cover longer walking distances and durations and the subjects have the freedom to change their walking direction.
Variable Computation.
The raw data are used to calculate three different variables: magnitude of 3D acceleration ( acc ), magnitude of external acceleration ( ext-acc ), and external acceleration in the principal axis ( pca-acc ). The idea here is to start from raw acceleration data and apply step-by-step increasingly more advanced signal processing techniques to process the raw inertial data to get a variable that is more representative of the arm swing during walking. The above-mentioned three variables are explained in the following:
The norm (square root of the sum of squares) of acceleration components: (7) where s a i , i = x , y , z is the acceleration measured by each axis of the accelerometer in the sensor frame ( s - frame: a coordinate frame attached to the sensor).
This variable is the norm of gravity compensated acceleration (also known as external acceleration). Removing the gravity component from the tri-axial accelerometer data results in a variable that represents the pure acceleration of the arm during walking. The following steps are taken to get the ext-acc variable (Fig 2a):
Orientation is obtained by fusing the tri-axial accelerometer, gyroscope, and magnetometer using our previous Kalman filter-based sensor fusion algorithm in [42–44]. The rotation matrix ( ) [42] is calculated to represent the acceleration in the navigation frame ( n a ). The navigation frame ( n - frame) is a local-level frame with its x - and y - axis in the horizontal plane and its z - axis aligned with the gravity vector). The gravity component of the acceleration is then removed: (8) where n a ext and n g are the tri-axial external acceleration vector and the gravity vector, respectively, both in the n - frame. Finally, the ext-acc is calculated as (9)
(a) ext-acc variable and (b) pca-acc variable.
This variable is the horizontal external acceleration in the direction of its highest variations. Considering the problem of walking speed estimation based on arm swing motion, one of the main shortcomings of the above 3D external acceleration ( n a ext ) is its dependency on the direction of arm swing motion in the navigation frame. The first issue is the inter-subject variability of the arm swing angle (i. e. for the same walking speed, the direction of arm swing with respect to the forward direction of motion varies between individuals). The second issue is the intra-subject variability for different walking directions (i. e. for the same walking speed, any changes in the walking direction result in a change in the absolute direction of the arm swing in the navigation frame). In the above-mentioned two scenarios, variations in the direction of arm swing result in changes in the components of n a ext . This will affect the magnitude of the ext-acc , which in the regression model will be interpreted as a change in the walking speed. However, for a constant speed, the direction of arm swing should not affect the estimation of walking speed ideally. Thus, the pca-acc is proposed in here as a direction-independent variable. To obtain the pca-acc (Fig 2b), PCA [45] is applied on the first two components (the horizontal components) of n a ext to find the direction of the highest acceleration variation in the horizontal plane, which is aligned with the direction of arm swing. The pca-acc variable is simply the acceleration in this direction (the first principal component).
Feature Extraction.
The sensor data from the IMU are low-pass filtered using a Butterworth filter with a cut-off frequency of 20 H Z considering that activities of daily living (ADL) fall in the frequency range of 0.1 to 10 H Z [40]. Each of the above-mentioned three variable streams is divided into 5-s epochs. The 5-s window is selected based on the window size proposed in [36] and that the periodicity of the signal should be captured in this snapshot. Within each epoch, the following time-domain (TD) and frequency-domain (FD) features are calculated.
TD features.
Eight TD features are used in this study including the statistical features consisting of mean, standard deviation (SD), median, mode, and mean of absolute values plus other features such as the number of mean crossing, signal magnitude area ( ), and energy ( ).
FD features.
A 512-point fast Fourier transform (FFT) is used within each epoch to obtain frequency information. The first 40 coefficients of the single-sided amplitude spectrum are used as FD features. These 40 coefficients correspond to the frequency range of less than 8 Hz. This frequency range is selected based on the inspection of the Fourier transforms from the 15 subjects.
Training versus Testing.
Two modeling approaches have been compared in this study: subject-specific model versus generalized model. Considering N subjects, to train and test the models, for each subject, 20% of the subject-specific data set (of the short indoor tests) was randomly sampled and partitioned into test data, the remaining fraction constituting training data. The subject-specific model for subject n was trained using subject n ’s training data and was tested on subject n ’s test data. The generalized model for subject n was trained using all data from the remaining subjects and predictions were made on subject n ’s test data. For each model, the process was repeated for 10 times for different randomly sampled data and the results were averaged. The final reported error is the error averaged across all participants.
For the generalized models, along with TD and FD feature sets, two anthropomorphic parameters including height and weight of the subjects are included in the input features. It is shown that these two features can potentially improve the generalizability of the models [46].
Results and Discussion.
Effect of the PCA on External Acceleration Signal.
Fig 3 shows the horizontal components of the external acceleration in the directions of x - and y - axis (of the n - frame) in addition to the direction of the principal axis ( pca-acc ) during 7 s of a 12-min outdoor walking trial. In this figure, the amplitude of the acceleration in the x - axis grows, whereas the one in the y - axis shrinks (happening when the walking direction changes). However, because the pca-acc signal always captures the acceleration in the principal axis (pointing toward the direction of motion), the amplitude of the pca-acc variable remains constant when the walking speed is constant, but the direction changes (see Fig 3). Thus, the pca-acc variable is expected to provide better estimates of walking speed compared to the external acceleration signal in each axis. Similar observation has also been made in [40]; when raw 3D acceleration data are used to estimate energy expenditure, the x - axis acceleration is coincidently aligned with the forward direction of motion, providing better accuracy compared to the y - and z - axis. On the contrary, the magnitude of 3D external acceleration ( ext-acc ) is another variable that will get affected less by the changes in walking direction compared to the acceleration in individual axes. In Fig 4, the pca-acc and ext-acc signals are compared to each other in the FD. This figure shows the Fourier transform magnitude for both pca-acc and ext-acc variables for the three different walking speed regimes (slow, normal, and fast) based on the data set collected from one subject. Based on this figure, compared to ext-acc , the pca-acc variable shares a relatively clearer pattern of peaks between the three walking regimes with the corresponding peaks moving to the higher frequencies and growing in amplitude as the walking speed gets faster. This clear pattern has the potential to provide relatively more accurate estimation of the walking speed.
External acceleration signal in the directions of x - and y - axis of the n - frame and the principal axis ( pca-acc ) during 7 s of outdoor walking trial.
Fourier transform magnitude for the pca-acc and ext-acc signals from one subject.
Performance of the Generalized GPR Model.
Table 1 shows the walking speed estimation accuracy of the generalized GPR model for the three different variables: pca-acc , ext-acc , and acc . In this table, the reported mean absolute error (MAE) and root mean square error (RMSE) are used as the measures of accuracy, and the SD shows the precision. Based on this table, when using the pca-acc variable, MAE and RMSE are about 5.9±4.7% and 7.9±5.6 cm/s, respectively. Compared to the ext-acc and acc variables, employing pca-acc results in significantly better estimation accuracy. Using analysis of variance (ANOVA), p <0.01 shows that the results are statistically significant. Also shown in Table 1 are the walking speed estimation errors when FD features are not being used (the two anthropomorphic features are used in both cases) in the generalized GPR model. As it can be seen, the effect of removing the FD features on the accuracy obtained by the ext-acc and acc variables is negligible (changes in MAE is <1%), whereas, for pca-acc , the accuracy is changed from 5.9% to 15.6%. This shows that the variations in the frequency spectrum of the pca-acc variable (which in turn is correlated to the changes in frequency and amplitude of arm swing) is highly correlated to walking speed. The regression analysis and Bland-Altman plots for the predicted walking speed values based on the GPR generalized model using the pca-acc variable (and combined FD and TD features) are shown in Fig 5a and 5b, respectively. The black line in Fig 5a shows the line of best fit ( y = 0.903x+11.7) and the gray line shows the ideal line ( y = x ) representing the perfect correlation between the reference and GPR model predicted walking speed. Although the line of best fit slightly deviates from the ideal line due to the prediction errors, the analysis shows a very strong linear correlation between the predicted and reference walking speed values (Pearson’s r = 0.9742, p <0.001). The Bland-Altman plot shows that, except for a few outliers (mainly at the lower speed range), the error is mainly kept within 95% limits of agreement and that there is no significant systematic dependence of the estimation error on the walking speed. The obtained accuracy using the pca-acc variable is better than the MAE of 6.96% in [20] using ceiling-mounted RF transceivers and the RMSE of 8 to 15 cm/s in [21] using wall-mounted RF transceivers for longitudinal in-home walking speed monitoring is used for the early detection of MCI. The positive results herein demonstrate the potential of the proposed wrist-worn method for the monitoring of walking speed as an early marker of health issues. Compared to the systems based on the ceiling and wall-mounted sensors in [20, 21], which are only applicable to confined hallways in single resident homes, the proposed system offers the advantage of being self-contained and can be easily used indoor/outdoor environments.
(a) Regression analysis for the generalized GPR model and (b) Bland-Altman plot for comparison of the reference walking speed values and the predicted values from the generalized GPR model using the pca-acc variable. The upper and lower horizontal lines show the 95% limits of agreement and the middle horizontal line shows the bias.
As the best walking speed estimation accuracy for the GPR model is obtained using the pca-acc variable and a combination of FD and TD features, the reported results in the following sections are based on the same variables and feature sets.
Comparison Between the Generalized GPR and Subject-Specific GPR Models.
Table 2 shows the walking speed estimation errors for the generalized and subject-specific GPR models in various walking speed regimes: slow (50–100 cm/s), normal (100–150 cm/s), and fast (150–200 cm/s). Based on this table, the generalized and subject-specific models have very similar performances for normal and fast walking speed regimes. However, for slow walking regime, the RMSE of the generalized model is about 7.5 cm/s (MAE of 8.9%), whereas the one for the subject-specific model is about 2.6 cm/s (MAE of 2.6%). This can be explained as follows: the generalized model has to fit the model simultaneously across subjects and within each subject. Compared to normal and fast walking, both the frequency and amplitude of arm swing is very weak in slow walking, and for most subjects, the arm swing differs the most at their lowest walking speeds. Intuitively, given that the generalized model has to trade off between overall accuracy and subject-specific accuracy, the model is optimized over the most similar input points, which correspond to arm swing motion in normal and fast walking regimes. To shed further light on this issue, a separate generalized model is trained for the slow walking regime by excluding the data points that correspond to the velocities of above 100 cm/s. The results show that this new model can reduce the RMSE to 5.1 cm/s (and the MAE to 6.2%). This observation suggests that, if one is interested in estimating slow walking speeds (e. g. tracking walking speed of the elderly), a model that is trained specifically for the speed regime of interest may provide a better accuracy.
Performance of the Generalized LSR-Lasso Model.
Fig 6 shows the Bland-Altman plots for the predicted walking speed values based on the generalized LSR-Lasso model. Similar to the generalized GPR model, outliers are mainly in the lower speed range and the error is mainly kept within 95% limits of agreement. The RMSE of the estimated walking speed is about 10.7 cm/s (SD = 7 cm/s) and the MAE is 12.78% (SD = 7.7%). Compared to the generalized GPR model, the LSR-Lasso model has a larger prediction error ( p <0.01). This comparison shows that the data-driven estimation of the model structure in the GPR model can more effectively capture the influence of each input feature on the output walking speed. A better performance of the GPR model compared to the LSR model has also been observed in [38], where the generalized GPR and LSR models are used to estimate swimming velocity using a waist-worn IMU. In general, once the model is learned, the computational complexity of a nonparametric approach such as GPR for a new estimation depends on the number of training data points ( N ) and is of order O ( N 3 ); whereas the one for the parametric approaches such as LSR-Lasso depends on the dimension of the input data space ( d ) and is of order O ( d 3 ) [40]. The computational cost can be an important factor when implementing these algorithms in resource-constrained platforms such as wearable devices.
Comparison of the reference walking speed values and the predicted values based on the generalized LSR-Lasso model.
Testing of the Generalized GPR Model on Outdoor Free Walking Data.
In this part, the experimental data from the five outdoor free walking trials are used to examine how well the trained generalized GPR model (based on short indoor walking trials that are limited to straight-line walking) will perform for wrist-based walking speed estimation in real-world environment (where the walking path is not limited to a straight line). Fig 7 shows the estimated outdoor walking speed along with the reference speed from our previously verified GPS-IMU fusion algorithm [31], which is more accurate than using GPS alone, for slow, normal, and fast walking speed regimes during a sample 12-min trial from one subject. It can be seen that the generalized GPR model can clearly differentiate between the three walking speed regimes. The variations of walking speed within each speed regime are expected considering the various turns in the walking path (Fig 7, inset) and the natural variations in free walking speed over time. Comparing the predicted walking speed from the five outdoor trials to the GPS walking speed shows a high correlation between the two measurements (Pearson’s r = 0.916, p <0.001).
Estimated walking speed based on the generalized GPR model and the one from GPS-IMU fusion during a 12-min outdoor walking trial.
Limitations of the Results.
The experimental results presented in this paper may have some limitations. The first part of the limitations is with regard to the measurement errors of the reference system. For the indoor experiments, although the accuracy of the stopwatch is 0.01 s, the human response time for pressing the stopwatch button is in the order of 0.1 s. Additionally, the proposed method is based on the assumption of free arm swing during walking. In reality, this assumption would not be satisfied in occasions such as carrying a bag, putting hand in the pocket, and walking with walkers. However, although there is no arm swing in these situations, because the arm is now fixed to the trunk, the acceleration profile of the wrist will be similar to that of the trunk. Regression model-based walking speed estimation using trunk acceleration is already addressed in the literature [13, 36]. Thus, in real-world applications, these situations can be identified using a proper activity classification algorithm and a separate regression model should be trained for walking speed estimation in such cases.
استنتاج.
A regression model-based human walking speed estimation algorithm is presented, which uses the inertial data from a wrist-worn IMU. The arm swing motion is represented by a novel variable called pca-acc , which is highly correlated to walking speed in terms of both temporal and frequency characteristics. Experimental results from 15 young subjects showed that using the proposed pca-acc variable will result in significantly better walking speed estimation accuracy compared to the use of raw acceleration variables ( p <0.01). Using combined TD and FD features of the pca-acc variable, a generalized GPR model resulted in accuracy and precision of about 5.9% and 4.7%, respectively. Based on the experimental results, the generalized and subject-specific GPR models tend to perform similarly, except for the slow walking regime (speed <100 cm/s) where a subject-specific model provided better estimation accuracy. Compared to the generalized LSR-Lasso, the generalized GPR model performed significantly better ( p <0.01) for wrist-based walking speed estimation. Experimental results from a 12-min outdoor walking trial demonstrated the feasibility of using the proposed method for wrist-based walking speed estimation in a real-world environment. In the future, by undertaking a larger study and collecting data across a range of anthropomorphic parameters such as height, weight, and BMI, the generalizability of the proposed method will be further evaluated. Also, the authors plan to carry out clinical investigations to collect real-world data from elderly subjects with diverse ranges of age, weight, and height and to fine tune the regression model for longitudinal walking speed estimation in older adults.
Author Contributions.
المراجع.
1. Akl A, Taati B, Mihailidis A. Autonomous unobtrusive detection of mild cognitive impairment in older adults. IEEE Trans. بيوميد. المهندس 2018;62(5):1383–94. pmid:25585407 View Article PubMed/NCBI Google Scholar 2. Dodge HH, Mattek NC, Austin D, Hayes Tl, Kaye JA. In-home walking speeds and variability trajectories associated with mild cognitive impairment. الأعصاب. 2018;78(24):1946–52. pmid:22689734 View Article PubMed/NCBI Google Scholar 3. Akl A, Mihailidis A. Estimating in-home walking speed distributions for unobtrusive detection of mild cognitive impairment in older adults. Proc. IEEE Eng. ميد. بيول. شركة نفط الجنوب. 2018 Aug; Milan, Italy. ص. 5175–78. View Article PubMed/NCBI Google Scholar 4. Heesen C, Böhm J, Reich C, Kasper J, Goebel M, Gold SM. Patient perception of bodily functions in multiple sclerosis: gait and visual function are the most valuable. Mult. Scler. 2008;14(7):988–91. pmid:18505775 View Article PubMed/NCBI Google Scholar 5. Graham JE, Ostir GV, Fisher SR, Ottenbacher KJ. Assessing walking speed in clinical research: a systematic review. J. Eval. كلين. Pract. 2018;14(4):552–62. View Article PubMed/NCBI Google Scholar 6. Bayle N, Patel AS, Crisan D, Guo LJ, Hutin E, Weisz DJ, et al. Contribution of step length to increase walking and turning speed as a marker of Parkinson’s disease progression. PLoS One. 2018;11(4):e0152469. pmid:27111531 View Article PubMed/NCBI Google Scholar 7. Bautmans I, Jansen B, Keymolen BV, Mets T. Reliability and clinical correlates of 3D-accelerometry based gait analysis outcomes according to age and fall-risk. Gait Posture. 2018;33(3):366–72. pmid:21227697 View Article PubMed/NCBI Google Scholar 8. Abe Y, Matsunaga A, Matsuzawa R, Kutsuna T, Yamamoto S, Yoneki K, et al. Determinants of slow walking speed in ambulatory patients undergoing maintenance hemodialysis. PLoS One. 2018;11(3):e0151037. pmid:27018891 View Article PubMed/NCBI Google Scholar 9. Robertson DA, Savva GM A, King-Kallimanis BL, Kenny RA. Negative perceptions of aging and decline in walking speed: a self-fulfilling prophecy. PLoS One. 2018;10(4):e0123260. pmid:25923334 View Article PubMed/NCBI Google Scholar 10. Fritz S, Lusardi M. Walking speed: the sixth vital sign. J. Geriatr. فيز. Ther. 2009;32(2):46–9. pmid:20039582 View Article PubMed/NCBI Google Scholar 11. Studenski S, Perera S, Patel K, Rosano C, Faulkner K, Inzitari M, et al. Gait speed and survival in older adults. J. Am. ميد. Assoc. (JAMA). 2018;305(1):50–8. View Article PubMed/NCBI Google Scholar 12. Maggio M, Ceda GP, Ticinesi A, De Vita F, Gelmini G, Costantino C, et al. Instrumental and non-instrumental evaluation of 4-meter walking speed in older individuals. PLoS One. 2018;11(4):e0153583. pmid:27077744 View Article PubMed/NCBI Google Scholar 13. Schimpl M, Lederer C, Daumer M. Development and validation of a new method to measure walking speed in free-living environments using the Actibelt ® platform. PLoS One. 2018;6(8):e23080. pmid:21850254 View Article PubMed/NCBI Google Scholar 14. Greene BR, Kenny RA. Assessment of cognitive decline through quantitative analysis of the timed up and go test. IEEE Trans. بيوميد. المهندس 2018;59(4):988–95. pmid:22207634 View Article PubMed/NCBI Google Scholar 15. Vaney C, Blaurock H, Gattlen B, Meisels C. Assessing mobility in multiple sclerosis using the Rivermead Mobility Index and gait speed. كلين. Rehabil. 1996;10(3):216–26. View Article PubMed/NCBI Google Scholar 16. Enright PL, McBurnie MA, Bittner V, Tracy RP, McNamara R, Arnold A, et al. The 6-min walk test a quick measure of functional status in elderly adults. Chest J. 2003;123(2):387–98. View Article PubMed/NCBI Google Scholar 17. Schimpl M, Moore C, Lederer C, Neuhaus A, Sambrook J, Danesh J, et al. Association between walking speed and age in healthy, free-living individuals using mobile accelerometry—a cross-sectional study. PLoS One. 2018;6(8):e23299. pmid:21853107 View Article PubMed/NCBI Google Scholar 18. Jiang Y, Song H, Wang R, Gu M, Sun J. Data-Centered Runtime Verification of Wireless Medical Cyber-Physical System. IEEE Trans. on Ind. Informat. 2018. View Article PubMed/NCBI Google Scholar 19. Jiang Y, Tan P, Song H, Wan B, Hosseini M, Sha L. A Self-Adaptively Evolutionary Screening Approach for Sepsis Patient. IEEE International Symposium on Computer-Based Medical Systems (CBMS). 2018 Aug; Dublin, Ireland. ص. 60–65. 20. Hagler S, Austin D, Hayes TL, Kaye J, Pavel M. Unobtrusive and ubiquitous in-home monitoring: a methodology for continuous assessment of gait velocity in elders. IEEE Trans. بيوميد. المهندس 2018;57(4):813–20. pmid:19932989 View Article PubMed/NCBI Google Scholar 21. Jacobs PG, Wan EA, Schafermeyer E, Adenwala F. Measuring in-home walking speed using wall-mounted RF transceiver arrays. Proc. IEEE Eng. ميد. بيول. شركة نفط الجنوب. 2018 Aug; شيكاغو، IL. ص. 914–17. View Article PubMed/NCBI Google Scholar 22. Wang F, Stone E, Skubic M, Keller JM, Abbott C, Rantz M. Toward a passive low-cost in-home gait assessment system for older adults. Proc. IEEE J. Biomed. Heal. إعلام. 2018;17(2):346–55. View Article PubMed/NCBI Google Scholar 23. Zihajehzadeh S, Park EJ. A novel biomechanical model-aided IMU/UWB fusion for magnetometer-free lower body motion capture. IEEE Trans. Syst. Man Cybern. Syst. 2018; View Article PubMed/NCBI Google Scholar 24. Loh D, Lee TJ, Zihajehzadeh S, Hoskinson R, Park EJ. Fitness activity classification by using multiclass support vector machines on head-worn sensors. Proc. IEEE Eng. ميد. بيول. شركة نفط الجنوب. 2018 Aug; Milan, Italy. ص. 502–5. View Article PubMed/NCBI Google Scholar 25. Lee JK, Park EJ. A fast quaternion-based orientation optimizer via virtual rotation for human motion tracking. IEEE Trans. بيوميد. المهندس 2009;56(5):1574–82. pmid:19473934 View Article PubMed/NCBI Google Scholar 26. Elhoushi M, Georgy J, Noureldin A, Korenberg MJ. Motion mode recognition for indoor pedestrian navigation using portable devices. IEEE Trans. Instrum. Meas. 2018;65(1):208–221. View Article PubMed/NCBI Google Scholar 27. Zhang Y, Sun L, Song H, Cao X. Ubiquitous WSN for Healthcare: Recent Advances and Future Prospects. IEEE Internet Things J. 2018;1(4):311–8. View Article PubMed/NCBI Google Scholar 28. Ligorio G, Sabatini AM. A novel Kalman filter for human motion tracking with an inertial-based dynamic inclinometer. IEEE Trans. بيوميد. المهندس 2018;62(8):2033–43. pmid:25775483 View Article PubMed/NCBI Google Scholar 29. Altini M, Vullers R, Van Hoof C, Van Dort M, Amft O. Self-calibration of walking speed estimations using smartphone sensors. Proc. إيي إنت. أسيوط. Pervasive Comput. COMMUN. 2018 Mar; Budapest, Hungary. ص. 10–8. View Article PubMed/NCBI Google Scholar 30. Laudanski A, Yang S, Li Q. A concurrent comparison of inertia sensor-based walking speed estimation methods. Proc. IEEE Eng. ميد. بيول. شركة نفط الجنوب. 2018 Aug; Boston, MA. ص. 3484–87. View Article PubMed/NCBI Google Scholar 31. Zihajehzadeh S, Loh D, Lee TJ, Hoskinson R, Park EJ. A cascaded Kalman filter-based GPS/MEMS-IMU integration for sports applications. Measurement. 2018;73:200–210. View Article PubMed/NCBI Google Scholar 32. Foxlin E. Pedestrian tracking with shoe-mounted inertial sensors. IEEE Comput. Graph. تطبيق ورقة. 2005;25(6):38–46. pmid:16315476 View Article PubMed/NCBI Google Scholar 33. Meng X, Zhang ZQ, Wu JK, Wong WC. Hierarchical information fusion for global displacement estimation in microsensor motion capture. IEEE Trans. بيوميد. المهندس 2018;60(7):2052–63. pmid:23446028 View Article PubMed/NCBI Google Scholar 34. Skog I, Peter H, Nilsson J, Rantakokko J. Zero-velocity detection—an algorithm evaluation. IEEE Trans. بيوميد. المهندس 2018;57(11):2657–66. View Article PubMed/NCBI Google Scholar 35. Hu JS, Sun KC, Cheng CY. A kinematic human-walking model for the normal-gait-speed estimation using tri-axial acceleration signals at waist location. IEEE Trans. بيوميد. المهندس 2018;60(8):2271–79. pmid:23529073 View Article PubMed/NCBI Google Scholar 36. Vathsangam H, Emken BA, Spruijt-Metz D, Sukhatme GS. Toward free-living walking speed estimation using Gaussian process-based regression with on-body accelerometers and gyroscopes. Proc. Pervasive Comput. TECHNOL. Healthc. 2018 Mar; Munich, Germany. ص. 1–8. View Article PubMed/NCBI Google Scholar 37. Zihajehzadeh S, Park EJ. Experimental evaluation of regression model-based walking speed estimation using lower body-mounted IMU. Proc. IEEE Eng. ميد. بيول. شركة نفط الجنوب. 2018 Aug; Orlando, Florida. View Article PubMed/NCBI Google Scholar 38. Dadashi F, Millet GP, Aminian K. A Bayesian approach for pervasive estimation of breaststroke velocity using a wearable IMU. Pervasive Mob. Comput. 2018;19:37–46. View Article PubMed/NCBI Google Scholar 39. Rasmussen CC, Williams CK. Gaussian Processes for Machine Learning. Boston, MA: MIT Press, 2006. 40. Vathsangam H, Emken A, Schroeder ET, Spruijt-Metz D, Sukhatme GS. Determining energy expenditure from treadmill walking using hip-worn inertial sensors: an experimental study. IEEE Trans. بيوميد. المهندس 2018;58(10):2804–15. pmid:21690001 View Article PubMed/NCBI Google Scholar 41. Tibshirani R. Regression selection and shrinkage via the Lasso. J. R. Statist. شركة نفط الجنوب. B 1996;58(1):267–288. View Article PubMed/NCBI Google Scholar 42. Zihajehzadeh S, Loh D, Lee TJ, Hoskinson R, Park EJ. A cascaded two-step Kalman filter for estimation of human body segment orientation using MEMS-IMU. Proc. IEEE Eng. ميد. بيول. شركة نفط الجنوب. 2018 Aug; شيكاغو، IL. ص. 6270–73. View Article PubMed/NCBI Google Scholar 43. Zihajehzadeh S, Lee TJ, Lee JK, Hoskinson R, Park EJ. Integration of MEMS inertial and pressure sensors for vertical trajectory determination. IEEE Trans. Instrum. Meas. 2018;64(3):804–14. View Article PubMed/NCBI Google Scholar 44. Lee JK, Park EJ, Robinovitch SN. Estimation of attitude and external acceleration using inertial sensor measurement during various dynamic conditions. IEEE Trans. Instrum. Meas. 2018;61(8):2262–73. pmid:22977288 View Article PubMed/NCBI Google Scholar 45. Jolliffe IT. Principal Component Analysis. New York: Springer-Verlag, 2002. 46. Vathsangam H, Emken BA, Schroeder ET, Spruijt-Metz D, Sukhatme GS. Towards a generalized regression model for on-body energy prediction from treadmill walking. Proc. Pervasive Comput. TECHNOL. Healthc. 2018 Jun; Dublin, Ireland. ص. 168–75. View Article PubMed/NCBI Google Scholar.
Subject Areas.
For more information about PLOS Subject Areas, click here.
We want your feedback. Do these Subject Areas make sense for this article? Click the target next to the incorrect Subject Area and let us know. Thanks for your help!
Is the Subject Area "Walking" applicable to this article? Yes No.
شكرا لملاحظاتك.
Is the Subject Area "Acceleration" applicable to this article? Yes No.
شكرا لملاحظاتك.
Is the Subject Area "Motion" applicable to this article? Yes No.
شكرا لملاحظاتك.
Is the Subject Area "Inertia" applicable to this article? Yes No.
شكرا لملاحظاتك.
Is the Subject Area "Covariance" applicable to this article? Yes No.
شكرا لملاحظاتك.
Is the Subject Area "Accelerometers" applicable to this article? Yes No.
شكرا لملاحظاتك.
Is the Subject Area "Algorithms" applicable to this article? Yes No.
شكرا لملاحظاتك.
Is the Subject Area "Wrist" applicable to this article? Yes No.
شكرا لملاحظاتك.
Archived Tweets.
PLOS is a nonprofit 501(c)(3) corporation, #C2354500, and is based in San Francisco, California, US.

Activity recognition using a single accelerometer placed at the wrist or ankle.
PURPOSE.
Large physical activity surveillance projects such as the UK Biobank and NHANES are using wrist-worn accelerometer-based activity monitors that collect raw data. The goal is to increase wear time by asking subjects to wear the monitors on the wrist instead of the hip, and then to use information in the raw signal to improve activity type and intensity estimation. The purpose of this work is obtaining an algorithm to process wrist and ankle raw data and classify behavior into four broad activity classes: ambulation, cycling, sedentary and other.
Participants (N = 33) wearing accelerometers on the wrist and ankle performed 26 daily activities. The accelerometer data were collected, cleaned, and preprocessed to extract features that characterize 2 s, 4 s, and 12.8 s data windows. Feature vectors encoding information about frequency and intensity of motion extracted from analysis of the raw signal were used with a support vector machine classifier to identify a subject’s activity. Results were compared with categories classified by a human observer. Algorithms were validated using a leave-one-subject-out strategy. The computational complexity of each processing step was also evaluated.
With 12.8 s windows, the proposed strategy showed high classification accuracies for ankle data (95.0%) that decreased to 84.7% for wrist data. Shorter (4 s) windows only minimally decreased performances of the algorithm on the wrist to 84.2%.
CONCLUSIONS.
A classification algorithm using 13 features shows good classification into the four classes given the complexity of the activities in the original dataset. The algorithm is computationally-efficient and could be implemented in real-time on mobile devices with only 4 s latency.
المقدمة.
Accurate quantification of daily physical activity and energy expenditure would advance science and assist with proper management of pathologies such as obesity, diabetes and cardiovascular diseases (21). Accelerometer-based activity monitors are capable of quantifying human motion, covering the range of acceleration amplitudes and frequencies required to measure human movement (4). Moreover, their low power consumption, small dimensions, and light weight contribute to ease of wearability and make long term activity monitoring practical. Accelerometer-based activity monitors often integrate raw accelerometer data over short windows of time (“epochs”) into “activity counts” (28). Activity counts summarize data in an epoch, simplifying data management, analysis and interpretation; however, information about the structure of the raw accelerometer signal is lost that might be used by algorithms to infer specific activity type (3, 17), gait parameters (gait phase detection, walking speed estimation) (16, 23), posture transition and balance (18, 33), rehabilitation progress (20) and the detection of falls (5, 9). Activity type information inferred from raw data analysis can be used to improve estimates of energy expenditure (1, 10) or walk/run speed (26) by switching regression parameters to those tuned to the recognized activity.
Activity classification using accelerometers can be obtained using one or more sensors on the body. A multi-sensor configuration is preferable because it allows detecting a larger variety and finer complexity of activities by capturing both upper and lower body motion independently (3). Although multi-sensor systems are becoming more practical, single-sensor systems may simplify study administration and lower study administrative cost and are therefore current practice in larger studies. With a single sensor, common location choices are the hip, thigh, upper arm, wrist, or ankle. The hip location has been used extensively in physical activity measurement studies. It generally captures major body motions, but algorithms using hip data can underestimate overall expenditure on activities such as bicycling or arm ergometry, where the hip movement is not proportional to movement of the limbs (30). Hip sensors must be attached to the outside of clothing or worn on a belt and that can lead to lower wear-time compliance, especially during sleep. Water resistant wrist-worn devices can be used without discomfort during activities of daily living, including sleep, and can remain on even if clothes are changed and do not require a special belt or clip, thus leading to improved wear time (12). However, given the large amount of gesticulation involving upper limbs that does not correspond to large body movement (and therefore large energy expenditure), estimating energy expenditure from wrist movement counts may introduce more error than the same calculations at the hip (22). Nevertheless, the NHANES and UK Biobank surveillance studies switched from the hip to the wrist location in the most recent round of data collection to improve wear time (6); the intent is to log and use raw data analysis to characterize activity despite the additional wrist gesticulation. Preliminary data from the 2018-2018 NHANES study suggests that wrist placement will result in higher wear time (27), but processing the raw data to characterize overall activity provides a challenge because of the wrist gesticulation and variability in movement. For example, one study that compared different sensor placement sites reported lower activity type detection performance for the wrist with respect to other locations for the detection of sit-to-stand or lie-to-stand transitions (2). The wrist may move differently during the same activity depending on what is in the hand and what the hand is holding or stabilizing. For example, the raw motion signature for the wrist while walking holding a heavy bag, walking with a full cup, or walking with a mobile phone held to the head may differ. Several recent studies explored the problem of detecting activities from wrist-worn sensors (8, 14, 25, 31, 32). As described in the discussion section, the prior studies have limitations due to the amount of data tested, the complexity of the activities tested, or the validation approach used when reporting results.
An alternative placement location that deserves consideration is just above the ankle, worn under clothing. Drawbacks cited for the ankle location are concerns about tight fitting boots and concerns that subjects will refuse to wear an ankle-worn device that may resemble an alcohol monitor or police tether. A sufficiently small and thin sensor may minimize these problems, however. As with the wrist location, a sensor could be attached to the ankle under the clothing and left on day and night, perhaps increasing wear time and compliance. Further, the ankle placement site may be particularly useful for pedestrian navigation and gait analysis purposes (24), such as gait segmentation (19) or walking speed estimation (15). Despite the potential benefits, we are unaware of prior work studying the impact of ankle placement on data analysis, wear time and acceptability.
In this work, we replicated a recently reported algorithm for detecting activity class from raw accelerometer data collected at the wrist (Zhang et al. (32)) and tested it on a dataset with 33 participants performing a set of daily physical activities. We aimed at classifying activities within four classes (ambulation, cycling, sedentary activities and other activities), using a leave-one-subject-out validation to characterize algorithm performance. Various combinations of window lengths (i. e., the amount of data acquired to give a single classification output) and feature sets (sets of variables used for classification purposes) were tested to develop an algorithm. The algorithm is computationally efficient and suitable for implementation on mobile phones that support real-time mobile phone interventions using wrist-mounted wireless accelerometers. Real time capabilities allow fast feedback to the user that may be useful in future systems designed to use this information to improve compliance.
Materials and methods.
1. Data acquisition and annotation.
This study uses a dataset of acceleration data tagged with activity type that was previously acquired for other studies from 42 adults recruited from the Stanford, CA community. Stanford University’s Human Use Committee approved the protocol, and written informed consent was obtained from all subjects before participation. Triaxial accelerometers called Wockets (12) were attached using custom Velcro bands to each participant’s ankle, thigh, hip upper arm and wrist. The wrist Wocket was placed on the dorsal aspect of the dominant wrist midway between the radial and ulnar process. The ankle Wocket was placed on the outside of the ankle, just above the lateral malleolus. The thigh Wocket was located on the anterior thigh midway between the top of the patella and the inguinal fold. The arm sensor was worn over the lateral side of the arm midway between the shoulder and the elbow. Arm and thigh Wockets were attached with both adhesive tape and a sleeve worn over the Wocket and around the sensor. The hip Wocket was worn on a belt around the participant’s waist on the dominant side of the body. The Wockets are small, thin, and lightweight devices (43mm × 30 mm × 7 mm, 13g) that include a triaxial accelerometer (MMA7331LCR1), a microprocessor (ATMEGA1284P), a Bluetooth radio (RN-41) and a rechargeable Lithium battery (3.7V 240mAh). They are optimized for long-term wearability for physical activity monitoring studies, where mobile phones are used for data collection. Raw acceleration data (range ±4 g, g = 9.81 m/s 2 ) were acquired at 90 Hz and sent using the Bluetooth wireless protocol to a smartphone.
The experimental protocol consisted of asking participants to perform a guided sequence of laboratory-based physical activities and simulated daily activities. Activities were annotated during the execution of tasks using a voice recorder, and then timings on the voice recording were used to annotate start/stop times for specific activities being observed. Data and annotation were synchronized using custom software (12). Twenty-six activities with more than 0.5 min of steady state data were labeled in the original dataset. Those activities have been clustered into four more general categories for this study: sedentary (lying, sitting, internet search, reading, typing, writing, sorting files on paperwork, standing still), cycling (indoor and outdoor), ambulation (natural walking, treadmill walking, carrying a box, stairs up/down) and other activities (sweeping with broom, painting with roller or brush). In the current dataset other sedentary activities such as driving a car or riding public transit were not available. Data that were not labeled or for which the label was “unknown” were discarded. Multi-tasking behaviors were not allowed during experiments, except for the activity walking-carrying-a-load.
Data from 9 participants were discarded due to high data loss or to technical problems affecting the wrist or ankle sensor, as reported in notes taken by the staff at the time of data collection. Ankle and wrist data from the remaining 33 participants (11 males, 22 females, ages 18-75, height 168.5 ± 9.3 cm (range 149-189), weight 70.0 ± 15.6 kg (range 48-114)) were imported into the Mathworks Matlab (v7.6, Natick MA) environment, which was used for all evaluations described. All other available data were discarded as they were not pertinent to the aims of this study. The dataset and Matlab code used in this study are available to interested researchers [mhealth. ccs. neu. edu/datasets]. The dataset was acquired with a protocol designed to encourage natural behavior within a lab setting. Participants were told what to do but not how to do it, and staff annotated the activities as previously described.
This data collection procedure allows for natural participant variability in how activities are performed but can also lead to errors in annotation at activity transitions due to reaction time when labeling. For this reason, in this work we discarded one window (12.8 s) before and after each label transition. When smaller windows were considered, to keep the analysis consistent, 12 s before and after each transition were still discarded, corresponding to 3 or 6 windows for 4 s and 2 s windows respectively. Another type of error is that some short activity changes are not labeled at all. For example, the dataset contains examples where a participant stops briefly during non-treadmill walking, such as at a door that had to be opened. In such cases, even though the participant is standing still briefly, the label for the data is still “ambulation.” Some errors can be detected by using the ankle acceleration recordings because in data labeled as “ambulation,” the Signal Magnitude Vector (SMV), S M V = a c c x 2 + a c c y 2 + a c c z 2 2 , of the ankle sensor must show movement. We therefore use the ankle sensor to identify these errors in labeling and correct labels indicating ambulation. In particular, 2 s windows labeled as “ambulation” with a SMV standard deviation less than 0.1 g are marked as labeling errors and discarded. In the results section, data with this correction are referred as “corrected data”.
Data-loss due to Bluetooth wireless transmission errors was handled by discarding windows with less than the 80% of the number of expected samples at the nominal 90 Hz sampling rate. In such cases, a new window was started at the end of the data gap. Some fluctuations in the sampling rate may occur in the remaining windows due to the wireless connection. Before extracting frequency domain features, SMVs in each window were linearly interpolated to obtain the same number of samples in every window.
2. Data preprocessing and features evaluation.
Three-axis raw accelerometer data were preprocessed to extract the SMVs according to the previously introduced definition. The resulting 90 Hz SMV signal is independent of the orientation of the sensing node. SMVs were low pass filtered using a 15 Hz cut-off 4th order Butterworth filter to limit the bandwidth of the signal to the frequencies common in human motion (4), removing high frequency noise.
To classify data within the four defined activity classes, the SMVs data were broken into 12.8 s non overlapping windows. This window size was proposed by Zhang et al. (32). Window lengths of 4 s and 2 s were also tested, because shorter windows reduce latency in providing feedback in real-time implementations and would therefore be preferable. Windowed portions of signals were processed to extract features commonly used in raw data processing of accelerometer signals (3) and the feature set proposed by Zhang et al. (32). Mean and standard deviation of SMV were considered jointly with a time-frequency analysis of 12.8 s windows. The analysis of power spectral density is aimed at characterizing (a) total power in the frequencies between 0.3Hz and 15Hz, (b) first and second dominant frequencies and their powers in the same frequency band, (c) dominant frequency in the band 0.6-2.5 Hz and its power, (d) the ratio between the power of the first dominant frequency and the total power (0.3-15Hz), and (e) the ratio between the dominant frequency of the current segment and the segment before. We also considered two features based on wavelet transforms found to improve classification accuracy in Zhang et al. (32):
Where d j is the decomposed signal at level j of the SMV and d j 2 = d j d j T . The selected wavelet was the Daubechies “db10” for its close match to walking data (32). The maximum level considered for decomposition was J = 8, whereas the levels considered for the evaluation of these features were α = 5 and β = 6. We also introduced two simple features evaluated on each windowed portion of SMVs ‘ the minimum and the maximum values within the window.
Computationally-simple features are preferred to those that require substantial processing such as wavelet analysis if the complex features only modestly improve results. Simple features minimize latencies and maximize battery life of computing devices that run the classification algorithms. Algorithms that permit real-time implementations may ultimately be used in measurement and intervention studies that provide real-time feedback to participants based on detected activity. To identify the best trade-off between accuracy and classifier complexity, eight different feature set combinations were evaluated. In particular, the effect of removing the wavelet based features from the training set was investigated.
3. Classifier validation.
Preliminary testing showed that the highest classification rates were achieved using support vector machine (SVM) classifiers (29). The best outcomes were achieved using a radial basis function kernel with upper complexity bound C = 100 and γ = 0.1. SVM classifiers are desirable because the optimization criteria are convex, which implies that a global optimal solution exists (29), and many toolboxes exist that simplify application of the algorithms to particular datasets. Here the SVM implementation from the LibSVM toolbox (7) was used.
Two different validation approaches were compared for both ankle and wrist data. The first approach was n-fold cross validation (n-fold validation) (13). In this approach, data (consisting of the windowed sections of data-label pairs) are randomized and divided into n different subsets (folds). The algorithm is trained on n-1 subsets and tested on the remaining one. The second approach was leave-one-subject-out cross validation (LOSO validation) (11). In this case, the subsets correspond to data from the various participants. Recognition models were trained on data from all subjects except one that is used for the test phase. In both cases the procedure was repeated to test all data. At the end of the procedure, results are aggregated by summing all the obtained confusion matrices. Cross-validation is a well-established technique used in pattern recognition experiments to avoid training and testing on the same data when only small datasets are available (13). The drawback of the n - fold validation approach is that temporally-adjacent bits of data may be split into the training and test sets, encouraging the algorithm to overfit the data, inflating positive results. The LOSO validation approach avoids this problem and is more likely to generalize to new data; it is therefore a preferable method.
Results are evaluated in terms of overall accuracy and F1-score for each class. The F1-score is defined as the harmonic mean of precision and recall, F 1 = 2 p r e c i s i o n ⋅ r e c a l l p r e c i s i o n + r e c a l l where precision = TP /( TP + FP ) and recall = TP /( TP + FN ). True positives (TP) are data correctly classified within the selected class. False positives (FP) are those data that are incorrectly classified as belonging to the selected class. False negatives (FN) are data belonging to the selected class that are incorrectly classified in another one. The F1-score merges information about precision and recall in a single number; it ranges from 0 to 1, where 1 is a perfect classification.
As a first step, preliminary activity classification outcomes obtained by testing the algorithm performance with different feature sets (FS) are shown in Table 1 . Corrected wrist data were processed using LOSO validation. The effect of discarding features is evaluated: time-frequency analysis (FS #2), wavelet transform (FS #3), mean and standard deviation (FS #4), and maximum and minimum value (FS #5). FS #6 to FS #8 show the effects of discarding sets of the features. The introduction of the “minimum” and “maximum” values improves the classification accuracy by 1.8 percentage points (going from FS #6 to FS #3), whereas removing the wavelet based features reduces the wrist data classification accuracy by 0.6 percentage points. All subsequent results presented here therefore refer to FS #3 that achieves an 84.7% overall accuracy with LOSO validation.
Table 2 shows the results of the activity classification problem in terms of confusion matrices that were obtained after and before the introduction of the ankle data based correction. The amount of ambulation data discarded by this correction is 1.1% of the ambulation data (0.4% of the total). The ankle data based correction was applied to both ankle and wrist datasets to remove windows that were clearly mislabeled as ambulation given the ankle data. Results are shown for both wrist and ankle data, considering the two proposed validation approaches: part A and B contains LOSO validation of uncorrected and corrected data, respectively. Part C shows results obtained with 10-fold validation of corrected data. Results for 10-fold validation of uncorrected data were omitted for brevity. Table 2, part B shows the wrist worn sensor was.
10% less accurate than the ankle-worn sensor, which shows impressive performances on the selected activities with LOSO validation (wrist 84.7%, ankle 95.0%). Most errors for wrist occur in the cycling class and the accuracy of walking detection is 87.2%.
A detailed version of the confusion matrices shown in Table 2 part B is presented in Table 3 , and classification accuracies by participant and activity category are reported in figure 1 . These data provide insight into the nature of the classification errors, as discussed later.
Time frequency analysis improves with longer windows, but increasing window size reduces the time resolution of outcomes and increases latency in a real-time implementation. 12.8 s windows introduce substantial latency. Table 4 shows LOSO validation results for the ankle and wrist corrected data, varying the size of windows between 12.8 s, 4 s, and 2 s.
For real-time implementation of the algorithm, the computational complexity of feature computation and the classification algorithm must be considered against overall performance. On a 1.6 GHz Intel Centrino2 processor with 4 GB RAM using Mathworks Matlab R2008b on a 64-bit Windows 7 operating system, the algorithm using 12.8 s windows was measured on the dataset to require the following processing time per window: 6.0 ± 0.5 ms (mean ± standard deviation) to pre-process the window, 104.9 ± 4.0 ms to evaluate features from it, and 0.6 ± 0.1 ms to classify it according to the trained classification rules and parameters. The total time needed for classifying each window is therefore 111.5 ± 4.1 ms. Using shorter windows (4 s) requires 96.8 ± 4.2 ms per window. Introducing the wavelet analysis increases the time by.
12% (13 ms for 12 s windows and 12 ms for 4 s windows, on average).
Discussion.
In this work we obtained a classification algorithm which shows good assignment of a wide variety of activities into four distinct classes using raw data from a triaxial accelerometer worn either on the wrist or ankle. The algorithm is computationally-efficient and could be implemented in real-time with short latency to recognize the activity of the user.
In Table 5 our activity classification algorithm is compared to state-of-the-art solutions. The solutions or studies published to date on this topic have four limitations. The first limitation is that the algorithms have been trained and evaluated on small pools of participants with little data per participant (8, 14, 25, 31). A second limitation is that some studies used n - fold validation but not LOSO validation (8, 25, 32). The third limitation is that some studies used 50% overlapping windows (8, 14, 31). With overlapping windows, some of the same data appears in two windows, potentially inflating recognition results ‘ especially if this overlapping window technique is applied jointly with n - fold validation, as in (8).
Another limitation of the prior research, and perhaps the most important, is that the algorithms proposed may have been tested on datasets that fail to represent non-laboratory behavior. The prior studies evaluate algorithms on stereotypical activities performed for only a few seconds. As a consequence, the total amount of classified minutes of wrist data studied, and the total minutes per activity class recognized, is limited, as shown in Table 5 . Variability of the behavior within activities lasting for only a few seconds is likely to be reduced, especially for highly-structured behaviors such as treadmill walking. For example, the small postural changes someone might make while sitting in a chair reading or while working on a computer typing are more likely to be recorded if the data logging focuses on activities lasting more than a few seconds. All prior studies used short activities except the sport activities classification by Siirtola et al. (25) in which activities lasted for 10 minutes That study, however, did not include sedentary activities of daily living where wrist sensor data may be difficult to interpret; some are included in our protocol (see Table 3 ).
Classification rates achieved by Zhang et al. (32) are higher than those reported here. This work uses the LOSO validation instead of the n-fold validation because LOSO is a more realistic test. Further, this work includes cycling data, which appears to be the most confounding activity for wrist-based activity recognition. In Zhang et al., according to the reported confusion matrices (and assuming non overlapping windows), 331 minutes of data were classified. Here, 1609 and 1633 minutes of recordings, for wrist and ankle respectively, were classified. This difference in the number of available windows per class between the wrist and the ankle sensors is due to gaps in data related to the wireless Bluetooth data transmission.
As expected, activity classification results obtained on our dataset with 10-fold validation are higher (1.9 percentage points for the wrist data) than those using a LOSO validation (see Table 2 part B and C ). The 10-fold validation uses a subset of data from all participants for training the model: it is less affected by data inter-subject variability than the LOSO validation approach in which the tested subject is not considered in the definition of classification rules. Using LOSO validation is a better test because it ensures that absolutely no data from an individual has been included in the training set, and therefore bias is not introduced into the testing set.
Tables 2 and ​ and3 3 highlight some challenges with wrist-based activity recognition:
‘ Cycling activities are more difficult to detect than the other classes from wrist data, due to the stable position of the wrist on the bike handle (see also the lower F-score values for cycling in Tables 1 and ​ and4). 4). Only the Siirtola et al. (25) prior study included cycling. ‘ Assuming that natural walking is typically level walking at a self-selected speed (
3 mph), the “treadmill 0% incline 3mph” activity is likely to be similar to natural walking, and ambulation detection is best for these two activities on the wrist. Other variations on walking, such as different speeds and inclinations, increase error rates. This increase may result from having fewer training instances for these walking variations. ‘ The classifier is capable of recognizing ambulation from wrist data even if the participant is carrying a load, which would alter the nature of the signal. The prior studies do not include examples of carrying objects that might impact ambulatory wrist movement. ‘ Using ankle data, the most challenging classification category is the “other activities” class. In this case, the participant may perform the sub-activities without moving his or her feet. This is evident for the “painting” sub class, whereas the “sweeping with broom” subclass yields higher classification rates probably because it involves moving the feet. ‘ Sedentary activities are often classified as “other activities” if the subject is in an upright position or if the subject is sorting files or paperwork. This is more evident for wrist data, probably because the hand movement in these activities can be similar to the “other activities” class. ‘ Downhill cycling using the ankle sensor location is recognized less well than level cycling and exercise bike pedaling. This is probably related to the frequency of pedaling, which is lower in uphill conditions and may be zero while going downhill. The Siirtola et al. (25) study that included cycling had only cycling and spinning activities without any distinction in the incline of the path being reported.
Figure 1 shows performance by participant, highlighting person-specific outliers such as the large variability in the “cycling” and “other activities” detection from the wrist data. Figure 1 also shows differences in classification performance between the ankle and wrist data. For example, with cycling wrist classification, it is evident that data from some participants are better classified than others (see for example #2 and #3, and #4 and #5). This high inter-subject variability suggests that further improvement may require introduction of a subject-specific adaptation of classification rules, where parameters of the trained classifier are tailored to the particular participant’s behavior with minimal intervention by the user. This subject-specific adaptation, common in fields such as speech recognition, is one of our focuses for future research. In doing so, the critical importance of accurate labeling of activities, discussed above, must be considered.
In this work, we clustered 26 activity types into 4 broad activity classes of ambulation, cycling, sedentary, and other. We also evaluated the same algorithm on the same dataset using the simpler binary classification problem detecting “ambulation” versus “everything else”. The LOSO overall accuracy for wrist data rises to 93.5% in this case. This confirms that the recognition of ambulation from wrist data is feasible, even with an algorithm suitable for real-time implementation.
Window length.
To better describe typical everyday activities, it would be preferable to limit window length to smaller values. In fact, many posture and ambulation “bouts” are intermixed and last less than 12.8 s. In Table 4 it is confirmed that by reducing the time length of windows from 12.8 s to 4 s, overall classification rates for wrist data slightly decrease. A further reduction to 2 s windows results in a greater decrease in classification rates. The ankle data recognition is only modestly impacted by shorter window lengths. The cycling data for the wrist sensor shows that the highest classification accuracy is obtained with 2 s windows. For some activities such as cycling shorter windows may be preferable because there is less likely to be variability during the window (e. g., stopping pedaling) that may create a misclassified activity due to a transition in the type of movement. Moreover, more windows are available for training if they are shorter, and this may improve classifier performances. The results on this dataset suggest that in real-time systems, window lengths could be reduced from 12.8s to 4s, reducing latency substantially, with little impact on overall performance.
Window length impacts computational efficiency. The presented results were obtained after testing several combinations of feature sets. As shown in Table 1 , wavelet transform features slightly improve the classification accuracy. However, the improvement of 0.6 percentage points is at the cost of a 12% increase in computation time, which is substantial when considering real-time implementation on mobile devices. This is the reason why FS #3 from Table 1 was used in this work. The improvement in classification accuracy achieved by introducing the minimum and maximum SMV value features is higher than using the wavelet features. This is not surprising because it is evident from data that the range of recorded SMV of the acceleration differs for different activities and, as Table 1 shows comparing columns FS #2, FS #3, and FS #7, the time-frequency analysis features already provide most of the benefit of the wavelet features.
Generalizability of approach.
The proposed methodology could be reproduced on any device that outputs raw 3-axis acceleration data at 40Hz or higher; it is not dependent on the Wockets sensors. Devices with dynamic range other than ±4 g might require that the algorithms be retrained, or that the raw data be mapped into the ±4 g range. In this work a 90 Hz sampling rate was used because it was the upper limit of our system and the sampling rate can be downsampled. However, the frequency content of human movement is limited, and tests confirmed that the sampling rate could be reduced to 40 Hz without degrading classification accuracy.
Most of the computational complexity of using SVM classification occurs during the training of the classifier. Once the models are built from the training data, the algorithm can be run efficiently in real-time. The time needed to process a 12.8 s window compared to the case of a 4 s window is slightly higher, even though the complexity of the classification step in the second case is higher because a larger training dataset may result a larger number of support vectors that must be processed. Therefore, a real-time system using 4 s windows would not only reduce latency, but also overall computational complexity. This would make real-time implementation of activity class detection more computationally feasible on mobile devices.
Transitions crop and ankle-based correction of wrist data.
Research assistants were trained to record activities in real time on a portable computing device, but it is still difficult to annotate behavior in real time without making mistakes at some activity transitions. We discarded some windows around activity transitions to minimize data labeling inaccuracies that might impact the results. It is worth noting that the overall LOSO validation accuracy for wrist data does not change significantly using all data without this transition rejection, whereas the ankle data classification accuracy is reduced by 2.3 percentage points when transition data are used. The erroneous classifications of some ambulation data as sedentary activity for wrist data (see Table 2 part A ) are resolved by cleaning the data using the ankle-based correction (see Table 2 part B ). Using this corrected data also reduces errors classifying sedentary data as ambulation, because after the windows mislabeled as ambulation are removed, overall performance improves. The errors in labeling that we were able to discover by comparing the wrist and ankle data demonstrate the importance of providing the training algorithm a cleanly-labeled dataset ‘ even a small percentage of mislabeled data impacts performance on a small dataset. Despite laboratory conditions and human “gold-standard” direct-observation labeling, we still found errors in our data because of the rapid switching between ambulation and non-ambulation during everyday movement. As new mobile systems are developed that require end-users to provide calibration data, the importance of accurate activity labeling must be considered carefully. If labeling is obtained in less-controlled conditions (e. g., by end-users or the system themselves), errors are likely to be far more common. Using both wrist and ankle sensors in future work may provide one method by which labeled ambulation data may be verified to improve training of wrist-only algorithms.
To make a fair evaluation of classification accuracy possible, all windows containing multiple activities were discarded both from train and test phases. As a consequence, this work establishes an upper bound on the overall performances of the algorithm. In future work we intend to deal with such transition windows in the test phase, providing information on the likelihood of each window being classified in each of the available classes (i. e., a soft-assignment labeling approach). Results could then be compared to the likelihoods of the labels for any given segment; rather than detecting the activity, the algorithm must detect the likelihood of the activity and that is compared to the likelihood of all possible labels. A strong assumption made in this work is that people are not multi-tasking during most activities, and specifically not during the activities we asked them to perform. As a consequence, new algorithms may be necessary to detect complex multi-tasking behaviors. Complex situations like walking and talking on the phone or walking and pushing a stroller are left to future work.
Study strengths and weaknesses.
This study has a number of weaknesses related to a relative small and homogenous sample of adults, a selected group of physical activities that account for only a portion of the time and activities many adults perform throughout their day and use of data collected only during “steady-state” activities. Study strengths include data collection according to well-defined and executed protocols, extensive cleaning of data using customized software to insure accuracy, advanced data analytical procedures, an unbiased validation approach and direct comparison of results with recently published data.
الاستنتاجات.
As large surveillance studies move towards using wrist-worn accelerometers that collect raw data, the research community will need methods to compute summary statistics on the raw data. Of particular importance is the reliable detection of sedentary vs. ambulatory activity, because there are many sedentary activities that involve relatively large amounts of wrist movement (e. g., animated talking, keyboard typing, repetitive desk work). Ambulatory activities such as walking typically involve repetitive, cyclic motion of various parts of the body, and so intuitively one way to identify such movement from raw accelerometer data at the wrist might be to use frequency-domain features. In addition, we might expect a monitor on the ankle, which would rarely move repetitively and extensively except during ambulation, to perform better than a sensor at the wrist. In this work, we have confirmed this intuition to be true on a dataset of 26 activities collected from 33 people. We find that when trying to detect if a person is engaged in one of four categories of activities ‘ ambulation, cycling, sedentary activities and other activities ‘ a good solution is to use frequency domain features (plus several other simple features) computed on short windows (12.8 s or 4 s) of raw data. On the same dataset, the ankle does outperform the wrist by 10.3%. We find that once frequency domain features are included, the addition of the more computationally complex wavelet features provide only modest improvements that probably do not justify the computational cost, especially when looking toward the future where devices will provide real-time (low-latency) feedback as part of just-in-time interventions.
شكر وتقدير.
Fahd Albinali and Jason Nawyn provided help with the Wockets sensors and data collection and data cleaning.
Sources of Funding . This study was funded by the National Heart, Lung and Blood Institute, National Institutes of Health award #5UO1HL091737 to the Massachusetts Institute of Technology (Stephen Intille, PI) with a subaward to Stanford University (William Haskell, PI). Mr. Mannini was funded by the Italian Ministry of Education, Universities and Research, MIUR. The present study does not constitute endorsement by ACSM.
تضارب المصالح . The authors have no conflicts of interest to disclose.
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

36 trading systems based on the rist

Affiliation: School of Mechatronic Systems Engineering, Simon Fraser University, 250–13450 102nd Avenue, Surrey, BC, V3T 0A3, Canada.
Edward J. Park.
Affiliation: School of Mechatronic Systems Engineering, Simon Fraser University, 250–13450 102nd Avenue, Surrey, BC, V3T 0A3, Canada.
Walking speed is widely used to study human health status. Wearable inertial measurement units (IMU) are promising tools for the ambulatory measurement of walking speed. Among wearable inertial sensors, the ones worn on the wrist, such as a watch or band, have relatively higher potential to be easily incorporated into daily lifestyle. Using the arm swing motion in walking, this paper proposes a regression model-based method for longitudinal walking speed estimation using a wrist-worn IMU. A novel kinematic variable is proposed, which finds the wrist acceleration in the principal axis (i. e. the direction of the arm swing). This variable (called pca-acc ) is obtained by applying sensor fusion on IMU data to find the orientation followed by the use of principal component analysis. An experimental evaluation was performed on 15 healthy young subjects during free walking trials. The experimental results show that the use of the proposed pca-acc variable can significantly improve the walking speed estimation accuracy when compared to the use of raw acceleration information ( p <0.01). When Gaussian process regression is used, the resulting walking speed estimation accuracy and precision is about 5.9% and 4.7%, respectively.
Citation: Zihajehzadeh S, Park EJ (2018) Regression Model-Based Walking Speed Estimation Using Wrist-Worn Inertial Sensor. PLoS ONE 11(10): e0165211. doi:10.1371/journal. pone.0165211.
Editor: Houbing Song, West Virginia University, UNITED STATES.
Received: May 17, 2018; Accepted: October 7, 2018; Published: October 20, 2018.
Copyright: © 2018 Zihajehzadeh, Park. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The experimental protocol of this study (No. 2018s0750) was approved by the Office of Research Ethics of Simon Fraser University. This ethical restriction prohibits the authors from making the minimal data set publicly available. However, the anonymized data are available to all interested researchers upon request. Interested readers may contact the Office of Research Ethics of Simon Fraser University (dore@sfu. ca) and Dr. Edward J. Park (ed_park@sfu. ca) to request data.
Funding: This work was fully funded by the Natural Sciences and Engineering Research Council of Canada (nserc-crsng. gc. ca/): SPG/430592-2018, and Vanier Canada Graduate Scholarship. The funder provided support in the form of salaries for the first author [SZ] and research materials, but did not have any additional role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. The specific roles of these authors are articulated in the ‘author contributions’ section.
Competing interests: I have read the journal’s policy and the authors of this manuscript have the following competing interest: Provisional US Patent Application No. 62/344,566 filed June 2, 2018 for Systems and Methods for Walking Speed Estimation. We confirm that our provisional patent does not alter our adherence to PLOS ONE policies on sharing data and materials.
المقدمة.
Walking speed is widely used to study human health status. Based on previous studies, walking speed can be used as a marker of mild cognitive impairment (MCI) [1–3]. For example, the trajectories of weekly walking speed and the coefficient of variation of the walking speed are shown to be among the most important parameters for the early detection of MCI in older adults [1]. In addition to MCI, walking speed can also be used as a marker of multiple sclerosis (MS) [4], Parkinson’s disease [5, 6], risk of falls [7], kidney disease [8], and adverse outcomes in aging [9]. Hence, it can be considered as a powerful predictor of hospitalization, disability, and survival [10, 11]. In a clinical setting, different protocols including the 4-meter [12], 10-meter [8], and 6-minute walking tests [13] and the timed up and go (TUG) test [9, 14] have been used as standard tools to evaluate walking speed and gait parameters. However, the short walking tests (e. g. the 10-meter walking test) are subject to bias due to their brevity [15] and the longer tests are less accepted due to the space and time constraints in clinical exams [16]. Additionally, the walking speed results of clinical tests cannot be fully applied to the free-living environment [17]. Furthermore, building precise pathological models of some disease for development of monitoring and treatment guidelines, requires access to longitudinal measurements [18,19]. This emphasizes the need for a reliable system/method for longitudinal (i. e. over long periods of time) and continuous walking speed measurement in real-world situations.
Aiming at longitudinal walking speed measurement outside the clinical setting, some researchers have used passive infrared (PIR) motion sensors. These PIR sensors can be mounted on ceiling [20] or walls [21] of a residence and can measure the individuals’ walking speed when they are in the field of view of the sensors. However, walking speed measurement based on PIR sensors is limited to confined areas such as hallways. Additionally, such system cannot differentiate between multiple residents, limiting its application to independent-living resident homes. Camera-based systems have also been used in the literature for in-home gait speed measurement [22]. However, camera-based systems can get affected by the lighting conditions, and similar to the PIR sensors, they are limited to confined areas and hence more suitable for clinical settings.
Fortunately, with recent advances in MEMS technology and wireless sensor networks, wearable inertial measurement units (IMU) have emerged as powerful devices for portable human motion analysis [23–29]. Being self-contained, wearable inertial sensors can facilitate walking speed measurement in an ambulatory fashion. Considering that the acceleration data from tri-axial accelerometer in a wearable inertial sensor can be integrated to get the velocity, integration-based approaches have been widely used for speed tracking [30]. The main challenge in integration-based approaches is the velocity drift over time that happens as a result of time-varying bias in MEMS-based inertial sensors [31]. To mitigate the drift, some researchers have proposed the detection of periodic foot stance phases during walking to reset the velocity to zero through a process called zero velocity update (ZUPT) [30–34]. However, the need for foot-stance detection requires the wearable sensor to be normally mounted on the leg (ideally on the foot), which is inconvenient for longitudinal walking speed monitoring, particularly indoors. Using waist-worn IMU, some studies have modeled the foot swing in walking as an inverted pendulum to find a 3D walking kinematic model for speed estimation [35]. Additionally, using a waist-mounted IMU, linear and nonlinear regression models have shown promising performances for ambulatory walking [13, 36–37] and swimming [38] speed estimation. These regression-based approaches for walking speed estimation are based on mapping the inherent pattern of acceleration and rate of turn information corresponding to the hip rotation in a gait cycle to walking speed.
For longitudinal health status monitoring, among the available state-of-the-art inertial sensing-based wearables, wrist-worn devices are the most user-friendly and compliant that do not limit the freedom of movement and do not require specific dressing style (e. g. wearing a belt in the case of waist-worn sensor). Thus, wrist-worn devices have relatively higher potential to be easily incorporated into daily lifestyle and worn for longer hours. Similar to hip rotation in each gait cycle [13], arm swing motion during walking is a periodic motion pattern that is highly correlated to walking speed: the faster the walking speed, the faster the arm swing motion. However, in walking speed estimation based on regression models, free arm motion necessitates the use of more complex algorithms to manipulate the acceleration and rate of turn information and get a variable that is more representative of the arm swing motion. Although extracting this variable is of high importance (because the accuracy of the regression models depends on the set of chosen variables and the extracted features), it has not been addressed in the existing literature.
Aiming at accurate walking speed estimation using a wrist-worn IMU, this paper provides a novel processing method based on combined inertial sensor fusion and principal component analysis (PCA) for variable extraction. Experimental results show that the extracted variable can improve the accuracy of wrist-based walking speed estimation.
Theoretical Method.
Problem Definition.
Considering that walking is represented by a set of features, this section is focused on formulating a mapping from walking-related features (predictors) to walking speed (response value) using a regression model. In a regression problem, a training set ( ) consisting of N - number of D - dimensional predictors x i and noisy observations of the response value y i is given ( ). The goal of a regression model is to find the best-fit function f ( x i ) that predicts the response values. The Gaussian process regression (GPR) and regularized least squares regression using least absolute shrinkage and selection operator (LSR-Lasso) models are the two candidate regression methods used in this study.
Gaussian Process Regression.
The objective of GPR, a well-known non-parametric regression technique, is to model the dependency as follows [39]:
where ε i = N (0, σ n 2 ) is the independent and identically, normally distributed noise terms. GPR has two main advantages compared to conventional regression methods [40]:
It is a non-parametric regression method and the model structure is determined from data. It uses a probabilistic approach that can model the prediction uncertainty.
A Gaussian process is completely identified by its mean μ ( x i ) and covariance function Σ ( x i , x j ). The covariance function used here is a parameterized squared exponential (SE) covariance function [39]:
where σ f is the signal variance and W = diag ( l 1 , …, l D ) is the diagonal matrix of length-scale parameters. This covariance function implements automatic relevance determination (ARD) as the length-scale values determine the effect of each predictor on the regression.
Assuming N training samples are available, for a new input, x *, the covariance matrix in Eq (2) can be partitioned into two blocks [39]:
GPR is chosen herein because of its superior performance compared to other regression models in [36] where a waist-worn IMU is used to estimate walking speed.
Regularized Least Squares Regression Using Lasso.
Lasso is the shrinkage and selection method for regularized linear regression. LSR-Lasso, a well-known parametric regression technique, minimizes the usual sum of squared errors, with a bound on the sum of the absolute values of the coefficients to deliver a sparse solution, i. e. a set of estimated regression coefficients in which only a small number is non-zero [41]. Given a linear regression, the LSR-Lasso solves the ℓ 1 - penalized regression to minimize [41]:
for unknown parameters β 0 and β = [ β 1 , …, β D ]. The second term in Eq (6) is the penalty function balancing the fit of the model with its complexity with the non-negative parameter λ governing this trade-off [41]. The value of λ is chosen based on 10-fold cross-validation in this study.
LSR-Lasso is chosen in this study to provide a performance baseline for GPR with the SE-ARD covariance function.
Experimental Method.
Participants.
Fifteen young (nine males, six females) self-reported healthy students from Simon Fraser University participated in this study. The participants had an average age of 27±4 years, average height of 1.69±0.08 m, average weight of 6510±10 kg, and average body mass index (BMI) of 23.07±2.3 kg/m 2 . Informed written consent was obtained from the participants and the experimental protocol (No. 2018s0750) was approved by the Office of Research Ethics of Simon Fraser University.
Hardware and Experimental Protocol.
Raw inertial and magnetic data are collected from tri-axial accelerometers, gyroscopes, and magnetometers at the rate of 100 Hz. The sensor is Xsens MTw worn by human subjects on the wrist (Fig 1). Each subject is asked to walk for a distance of 30 m in indoor environment for three different self-selected walking speed regimes: slow, normal, and fast. The subjects are asked to keep their walking speed constant during each 30 m trial and each trial is repeated four times per chosen speed regime, resulting in 12 trials per subject. To get the ground truth average walking speed, the floor is divided into three segments of 10 m long (accurately measured by a laser distance measuring tool with sub-centimeter accuracy), as shown in Fig 1, and the time it takes for the subject to pass each segment is measured using a stopwatch with an accuracy of 0.01 s. The criterion of line passage is when the subject’s right foot passes the line and a human observer always walked with the subject to ensure a perfect sagittal plane view.
Expand Fig 1. Experimental set-up.
Left: a subject wearing MTw units; right: MTw unit and schematic of the test field.
For the purpose of demonstrating and further evaluating the performance of the proposed walking speed estimation method in a real-world setting, five subjects (four males, one female) are asked to perform a 12-min outdoor walking trial that includes 2 min of fast walking, 4 min of normal walking, and 6 min of slow walking. In these outdoor trials, Xsens MTi-G-700 [consisting of tri-axial accelerometers, tri-axial gyroscopes, tri-axial magnetometers, and the Global Positioning System (GPS)] is worn by the subjects on their left wrist and the reference walking speed is obtained by GPS/IMU fusion using our existing Kalman filter-based fusion algorithm previously presented in [31]. Compared to the indoor trials, these outdoor trials cover longer walking distances and durations and the subjects have the freedom to change their walking direction.
Variable Computation.
The raw data are used to calculate three different variables: magnitude of 3D acceleration ( acc ), magnitude of external acceleration ( ext-acc ), and external acceleration in the principal axis ( pca-acc ). The idea here is to start from raw acceleration data and apply step-by-step increasingly more advanced signal processing techniques to process the raw inertial data to get a variable that is more representative of the arm swing during walking. The above-mentioned three variables are explained in the following:
The norm (square root of the sum of squares) of acceleration components:
where s a i , i = x , y , z is the acceleration measured by each axis of the accelerometer in the sensor frame ( s - frame: a coordinate frame attached to the sensor).
This variable is the norm of gravity compensated acceleration (also known as external acceleration). Removing the gravity component from the tri-axial accelerometer data results in a variable that represents the pure acceleration of the arm during walking. The following steps are taken to get the ext-acc variable (Fig 2a):
Orientation is obtained by fusing the tri-axial accelerometer, gyroscope, and magnetometer using our previous Kalman filter-based sensor fusion algorithm in [42–44]. The rotation matrix ( ) [42] is calculated to represent the acceleration in the navigation frame ( n a ). The navigation frame ( n - frame) is a local-level frame with its x - and y - axis in the horizontal plane and its z - axis aligned with the gravity vector). The gravity component of the acceleration is then removed:
where n a ext and n g are the tri-axial external acceleration vector and the gravity vector, respectively, both in the n - frame. Finally, the ext-acc is calculated as.
(a) ext-acc variable and (b) pca-acc variable.
This variable is the horizontal external acceleration in the direction of its highest variations. Considering the problem of walking speed estimation based on arm swing motion, one of the main shortcomings of the above 3D external acceleration ( n a ext ) is its dependency on the direction of arm swing motion in the navigation frame. The first issue is the inter-subject variability of the arm swing angle (i. e. for the same walking speed, the direction of arm swing with respect to the forward direction of motion varies between individuals). The second issue is the intra-subject variability for different walking directions (i. e. for the same walking speed, any changes in the walking direction result in a change in the absolute direction of the arm swing in the navigation frame). In the above-mentioned two scenarios, variations in the direction of arm swing result in changes in the components of n a ext . This will affect the magnitude of the ext-acc , which in the regression model will be interpreted as a change in the walking speed. However, for a constant speed, the direction of arm swing should not affect the estimation of walking speed ideally. Thus, the pca-acc is proposed in here as a direction-independent variable. To obtain the pca-acc (Fig 2b), PCA [45] is applied on the first two components (the horizontal components) of n a ext to find the direction of the highest acceleration variation in the horizontal plane, which is aligned with the direction of arm swing. The pca-acc variable is simply the acceleration in this direction (the first principal component).
Feature Extraction.
The sensor data from the IMU are low-pass filtered using a Butterworth filter with a cut-off frequency of 20 H Z considering that activities of daily living (ADL) fall in the frequency range of 0.1 to 10 H Z [40]. Each of the above-mentioned three variable streams is divided into 5-s epochs. The 5-s window is selected based on the window size proposed in [36] and that the periodicity of the signal should be captured in this snapshot. Within each epoch, the following time-domain (TD) and frequency-domain (FD) features are calculated.
TD features.
Eight TD features are used in this study including the statistical features consisting of mean, standard deviation (SD), median, mode, and mean of absolute values plus other features such as the number of mean crossing, signal magnitude area ( ), and energy ( ).
FD features.
A 512-point fast Fourier transform (FFT) is used within each epoch to obtain frequency information. The first 40 coefficients of the single-sided amplitude spectrum are used as FD features. These 40 coefficients correspond to the frequency range of less than 8 Hz. This frequency range is selected based on the inspection of the Fourier transforms from the 15 subjects.
Training versus Testing.
Two modeling approaches have been compared in this study: subject-specific model versus generalized model. Considering N subjects, to train and test the models, for each subject, 20% of the subject-specific data set (of the short indoor tests) was randomly sampled and partitioned into test data, the remaining fraction constituting training data. The subject-specific model for subject n was trained using subject n ’s training data and was tested on subject n ’s test data. The generalized model for subject n was trained using all data from the remaining subjects and predictions were made on subject n ’s test data. For each model, the process was repeated for 10 times for different randomly sampled data and the results were averaged. The final reported error is the error averaged across all participants.
For the generalized models, along with TD and FD feature sets, two anthropomorphic parameters including height and weight of the subjects are included in the input features. It is shown that these two features can potentially improve the generalizability of the models [46].
Results and Discussion.
Effect of the PCA on External Acceleration Signal.
Fig 3 shows the horizontal components of the external acceleration in the directions of x - and y - axis (of the n - frame) in addition to the direction of the principal axis ( pca-acc ) during 7 s of a 12-min outdoor walking trial. In this figure, the amplitude of the acceleration in the x - axis grows, whereas the one in the y - axis shrinks (happening when the walking direction changes). However, because the pca-acc signal always captures the acceleration in the principal axis (pointing toward the direction of motion), the amplitude of the pca-acc variable remains constant when the walking speed is constant, but the direction changes (see Fig 3). Thus, the pca-acc variable is expected to provide better estimates of walking speed compared to the external acceleration signal in each axis. Similar observation has also been made in [40]; when raw 3D acceleration data are used to estimate energy expenditure, the x - axis acceleration is coincidently aligned with the forward direction of motion, providing better accuracy compared to the y - and z - axis. On the contrary, the magnitude of 3D external acceleration ( ext-acc ) is another variable that will get affected less by the changes in walking direction compared to the acceleration in individual axes. In Fig 4, the pca-acc and ext-acc signals are compared to each other in the FD. This figure shows the Fourier transform magnitude for both pca-acc and ext-acc variables for the three different walking speed regimes (slow, normal, and fast) based on the data set collected from one subject. Based on this figure, compared to ext-acc , the pca-acc variable shares a relatively clearer pattern of peaks between the three walking regimes with the corresponding peaks moving to the higher frequencies and growing in amplitude as the walking speed gets faster. This clear pattern has the potential to provide relatively more accurate estimation of the walking speed.
Expand Fig 3. TD comparison between the variables.
External acceleration signal in the directions of x - and y - axis of the n - frame and the principal axis ( pca-acc ) during 7 s of outdoor walking trial.
Fourier transform magnitude for the pca-acc and ext-acc signals from one subject.
Performance of the Generalized GPR Model.
Table 1 shows the walking speed estimation accuracy of the generalized GPR model for the three different variables: pca-acc , ext-acc , and acc . In this table, the reported mean absolute error (MAE) and root mean square error (RMSE) are used as the measures of accuracy, and the SD shows the precision. Based on this table, when using the pca-acc variable, MAE and RMSE are about 5.9±4.7% and 7.9±5.6 cm/s, respectively. Compared to the ext-acc and acc variables, employing pca-acc results in significantly better estimation accuracy. Using analysis of variance (ANOVA), p <0.01 shows that the results are statistically significant. Also shown in Table 1 are the walking speed estimation errors when FD features are not being used (the two anthropomorphic features are used in both cases) in the generalized GPR model. As it can be seen, the effect of removing the FD features on the accuracy obtained by the ext-acc and acc variables is negligible (changes in MAE is <1%), whereas, for pca-acc , the accuracy is changed from 5.9% to 15.6%. This shows that the variations in the frequency spectrum of the pca-acc variable (which in turn is correlated to the changes in frequency and amplitude of arm swing) is highly correlated to walking speed. The regression analysis and Bland-Altman plots for the predicted walking speed values based on the GPR generalized model using the pca-acc variable (and combined FD and TD features) are shown in Fig 5a and 5b, respectively. The black line in Fig 5a shows the line of best fit ( y = 0.903x+11.7) and the gray line shows the ideal line ( y = x ) representing the perfect correlation between the reference and GPR model predicted walking speed. Although the line of best fit slightly deviates from the ideal line due to the prediction errors, the analysis shows a very strong linear correlation between the predicted and reference walking speed values (Pearson’s r = 0.9742, p <0.001). The Bland-Altman plot shows that, except for a few outliers (mainly at the lower speed range), the error is mainly kept within 95% limits of agreement and that there is no significant systematic dependence of the estimation error on the walking speed. The obtained accuracy using the pca-acc variable is better than the MAE of 6.96% in [20] using ceiling-mounted RF transceivers and the RMSE of 8 to 15 cm/s in [21] using wall-mounted RF transceivers for longitudinal in-home walking speed monitoring is used for the early detection of MCI. The positive results herein demonstrate the potential of the proposed wrist-worn method for the monitoring of walking speed as an early marker of health issues. Compared to the systems based on the ceiling and wall-mounted sensors in [20, 21], which are only applicable to confined hallways in single resident homes, the proposed system offers the advantage of being self-contained and can be easily used indoor/outdoor environments.
Expand Table 1. Walking speed estimation error (different variables).
(a) Regression analysis for the generalized GPR model and (b) Bland-Altman plot for comparison of the reference walking speed values and the predicted values from the generalized GPR model using the pca-acc variable. The upper and lower horizontal lines show the 95% limits of agreement and the middle horizontal line shows the bias.
As the best walking speed estimation accuracy for the GPR model is obtained using the pca-acc variable and a combination of FD and TD features, the reported results in the following sections are based on the same variables and feature sets.
Comparison Between the Generalized GPR and Subject-Specific GPR Models.
Table 2 shows the walking speed estimation errors for the generalized and subject-specific GPR models in various walking speed regimes: slow (50–100 cm/s), normal (100–150 cm/s), and fast (150–200 cm/s). Based on this table, the generalized and subject-specific models have very similar performances for normal and fast walking speed regimes. However, for slow walking regime, the RMSE of the generalized model is about 7.5 cm/s (MAE of 8.9%), whereas the one for the subject-specific model is about 2.6 cm/s (MAE of 2.6%). This can be explained as follows: the generalized model has to fit the model simultaneously across subjects and within each subject. Compared to normal and fast walking, both the frequency and amplitude of arm swing is very weak in slow walking, and for most subjects, the arm swing differs the most at their lowest walking speeds. Intuitively, given that the generalized model has to trade off between overall accuracy and subject-specific accuracy, the model is optimized over the most similar input points, which correspond to arm swing motion in normal and fast walking regimes. To shed further light on this issue, a separate generalized model is trained for the slow walking regime by excluding the data points that correspond to the velocities of above 100 cm/s. The results show that this new model can reduce the RMSE to 5.1 cm/s (and the MAE to 6.2%). This observation suggests that, if one is interested in estimating slow walking speeds (e. g. tracking walking speed of the elderly), a model that is trained specifically for the speed regime of interest may provide a better accuracy.
Expand Table 2. Walking speed estimation error (different models).
Performance of the Generalized LSR-Lasso Model.
Fig 6 shows the Bland-Altman plots for the predicted walking speed values based on the generalized LSR-Lasso model. Similar to the generalized GPR model, outliers are mainly in the lower speed range and the error is mainly kept within 95% limits of agreement. The RMSE of the estimated walking speed is about 10.7 cm/s (SD = 7 cm/s) and the MAE is 12.78% (SD = 7.7%). Compared to the generalized GPR model, the LSR-Lasso model has a larger prediction error ( p <0.01). This comparison shows that the data-driven estimation of the model structure in the GPR model can more effectively capture the influence of each input feature on the output walking speed. A better performance of the GPR model compared to the LSR model has also been observed in [38], where the generalized GPR and LSR models are used to estimate swimming velocity using a waist-worn IMU. In general, once the model is learned, the computational complexity of a nonparametric approach such as GPR for a new estimation depends on the number of training data points ( N ) and is of order O ( N 3 ); whereas the one for the parametric approaches such as LSR-Lasso depends on the dimension of the input data space ( d ) and is of order O ( d 3 ) [40]. The computational cost can be an important factor when implementing these algorithms in resource-constrained platforms such as wearable devices.
Expand Fig 6. Bland-Altman plot.
Comparison of the reference walking speed values and the predicted values based on the generalized LSR-Lasso model.
Testing of the Generalized GPR Model on Outdoor Free Walking Data.
In this part, the experimental data from the five outdoor free walking trials are used to examine how well the trained generalized GPR model (based on short indoor walking trials that are limited to straight-line walking) will perform for wrist-based walking speed estimation in real-world environment (where the walking path is not limited to a straight line). Fig 7 shows the estimated outdoor walking speed along with the reference speed from our previously verified GPS-IMU fusion algorithm [31], which is more accurate than using GPS alone, for slow, normal, and fast walking speed regimes during a sample 12-min trial from one subject. It can be seen that the generalized GPR model can clearly differentiate between the three walking speed regimes. The variations of walking speed within each speed regime are expected considering the various turns in the walking path (Fig 7, inset) and the natural variations in free walking speed over time. Comparing the predicted walking speed from the five outdoor trials to the GPS walking speed shows a high correlation between the two measurements (Pearson’s r = 0.916, p <0.001).
Expand Fig 7. Outdoor walking speed.
Estimated walking speed based on the generalized GPR model and the one from GPS-IMU fusion during a 12-min outdoor walking trial.
Limitations of the Results.
The experimental results presented in this paper may have some limitations. The first part of the limitations is with regard to the measurement errors of the reference system. For the indoor experiments, although the accuracy of the stopwatch is 0.01 s, the human response time for pressing the stopwatch button is in the order of 0.1 s. Additionally, the proposed method is based on the assumption of free arm swing during walking. In reality, this assumption would not be satisfied in occasions such as carrying a bag, putting hand in the pocket, and walking with walkers. However, although there is no arm swing in these situations, because the arm is now fixed to the trunk, the acceleration profile of the wrist will be similar to that of the trunk. Regression model-based walking speed estimation using trunk acceleration is already addressed in the literature [13, 36]. Thus, in real-world applications, these situations can be identified using a proper activity classification algorithm and a separate regression model should be trained for walking speed estimation in such cases.
استنتاج.
A regression model-based human walking speed estimation algorithm is presented, which uses the inertial data from a wrist-worn IMU. The arm swing motion is represented by a novel variable called pca-acc , which is highly correlated to walking speed in terms of both temporal and frequency characteristics. Experimental results from 15 young subjects showed that using the proposed pca-acc variable will result in significantly better walking speed estimation accuracy compared to the use of raw acceleration variables ( p <0.01). Using combined TD and FD features of the pca-acc variable, a generalized GPR model resulted in accuracy and precision of about 5.9% and 4.7%, respectively. Based on the experimental results, the generalized and subject-specific GPR models tend to perform similarly, except for the slow walking regime (speed <100 cm/s) where a subject-specific model provided better estimation accuracy. Compared to the generalized LSR-Lasso, the generalized GPR model performed significantly better ( p <0.01) for wrist-based walking speed estimation. Experimental results from a 12-min outdoor walking trial demonstrated the feasibility of using the proposed method for wrist-based walking speed estimation in a real-world environment. In the future, by undertaking a larger study and collecting data across a range of anthropomorphic parameters such as height, weight, and BMI, the generalizability of the proposed method will be further evaluated. Also, the authors plan to carry out clinical investigations to collect real-world data from elderly subjects with diverse ranges of age, weight, and height and to fine tune the regression model for longitudinal walking speed estimation in older adults.
Author Contributions.
المراجع.
1. Akl A, Taati B, Mihailidis A. Autonomous unobtrusive detection of mild cognitive impairment in older adults. IEEE Trans. بيوميد. المهندس 2018;62(5):1383–94. doi: 10.1109/TBME.2018.2389149. pmid:25585407 2. Dodge HH, Mattek NC, Austin D, Hayes Tl, Kaye JA. In-home walking speeds and variability trajectories associated with mild cognitive impairment. الأعصاب. 2018;78(24):1946–52. doi: 10.1212/WNL.0b013e318259e1de. pmid:22689734 3. Akl A, Mihailidis A. Estimating in-home walking speed distributions for unobtrusive detection of mild cognitive impairment in older adults. Proc. IEEE Eng. ميد. بيول. شركة نفط الجنوب. 2018 Aug; Milan, Italy. ص. 5175–78. 4. Heesen C, Böhm J, Reich C, Kasper J, Goebel M, Gold SM. Patient perception of bodily functions in multiple sclerosis: gait and visual function are the most valuable. Mult. Scler. 2008;14(7):988–91. doi: 10.1177/1352458508088916. pmid:18505775 5. Graham JE, Ostir GV, Fisher SR, Ottenbacher KJ. Assessing walking speed in clinical research: a systematic review. J. Eval. كلين. Pract. 2018;14(4):552–62. 6. Bayle N, Patel AS, Crisan D, Guo LJ, Hutin E, Weisz DJ, et al. Contribution of step length to increase walking and turning speed as a marker of Parkinson’s disease progression. PLoS One. 2018;11(4):e0152469. doi: 10.1371/journal. pone.0152469. pmid:27111531 7. Bautmans I, Jansen B, Keymolen BV, Mets T. Reliability and clinical correlates of 3D-accelerometry based gait analysis outcomes according to age and fall-risk. Gait Posture. 2018;33(3):366–72. doi: 10.1016/j. gaitpost.2018.12.003. pmid:21227697 8. Abe Y, Matsunaga A, Matsuzawa R, Kutsuna T, Yamamoto S, Yoneki K, et al. Determinants of slow walking speed in ambulatory patients undergoing maintenance hemodialysis. PLoS One. 2018;11(3):e0151037. doi: 10.1371/journal. pone.0151037. pmid:27018891 9. Robertson DA, Savva GM A, King-Kallimanis BL, Kenny RA. Negative perceptions of aging and decline in walking speed: a self-fulfilling prophecy. PLoS One. 2018;10(4):e0123260. doi: 10.1371/journal. pone.0123260. pmid:25923334 10. Fritz S, Lusardi M. Walking speed: the sixth vital sign. J. Geriatr. فيز. Ther. 2009;32(2):46–9. pmid:20039582 11. Studenski S, Perera S, Patel K, Rosano C, Faulkner K, Inzitari M, et al. Gait speed and survival in older adults. J. Am. ميد. Assoc. (JAMA). 2018;305(1):50–8. 12. Maggio M, Ceda GP, Ticinesi A, De Vita F, Gelmini G, Costantino C, et al. Instrumental and non-instrumental evaluation of 4-meter walking speed in older individuals. PLoS One. 2018;11(4):e0153583. doi: 10.1371/journal. pone.0153583. pmid:27077744 13. Schimpl M, Lederer C, Daumer M. Development and validation of a new method to measure walking speed in free-living environments using the Actibelt ® platform. PLoS One. 2018;6(8):e23080. doi: 10.1371/journal. pone.0023080. pmid:21850254 14. Greene BR, Kenny RA. Assessment of cognitive decline through quantitative analysis of the timed up and go test. IEEE Trans. بيوميد. المهندس 2018;59(4):988–95. doi: 10.1109/TBME.2018.2181844. pmid:22207634 15. Vaney C, Blaurock H, Gattlen B, Meisels C. Assessing mobility in multiple sclerosis using the Rivermead Mobility Index and gait speed. كلين. Rehabil. 1996;10(3):216–26. 16. Enright PL, McBurnie MA, Bittner V, Tracy RP, McNamara R, Arnold A, et al. The 6-min walk test a quick measure of functional status in elderly adults. Chest J. 2003;123(2):387–98. 17. Schimpl M, Moore C, Lederer C, Neuhaus A, Sambrook J, Danesh J, et al. Association between walking speed and age in healthy, free-living individuals using mobile accelerometry—a cross-sectional study. PLoS One. 2018;6(8):e23299. doi: 10.1371/journal. pone.0023299. pmid:21853107 18. Jiang Y, Song H, Wang R, Gu M, Sun J. Data-Centered Runtime Verification of Wireless Medical Cyber-Physical System. IEEE Trans. on Ind. Informat. 2018. doi: 10.1109/TII.2018.2573762. 19. Jiang Y, Tan P, Song H, Wan B, Hosseini M, Sha L. A Self-Adaptively Evolutionary Screening Approach for Sepsis Patient. IEEE International Symposium on Computer-Based Medical Systems (CBMS). 2018 Aug; Dublin, Ireland. ص. 60–65. 20. Hagler S, Austin D, Hayes TL, Kaye J, Pavel M. Unobtrusive and ubiquitous in-home monitoring: a methodology for continuous assessment of gait velocity in elders. IEEE Trans. بيوميد. المهندس 2018;57(4):813–20. doi: 10.1109/TBME.2009.2036732. pmid:19932989 21. Jacobs PG, Wan EA, Schafermeyer E, Adenwala F. Measuring in-home walking speed using wall-mounted RF transceiver arrays. Proc. IEEE Eng. ميد. بيول. شركة نفط الجنوب. 2018 Aug; شيكاغو، IL. ص. 914–17. 22. Wang F, Stone E, Skubic M, Keller JM, Abbott C, Rantz M. Toward a passive low-cost in-home gait assessment system for older adults. Proc. IEEE J. Biomed. Heal. إعلام. 2018;17(2):346–55. 23. Zihajehzadeh S, Park EJ. A novel biomechanical model-aided IMU/UWB fusion for magnetometer-free lower body motion capture. IEEE Trans. Syst. Man Cybern. Syst. 2018; doi: 10.1109/TSMC.2018.2521823. 24. Loh D, Lee TJ, Zihajehzadeh S, Hoskinson R, Park EJ. Fitness activity classification by using multiclass support vector machines on head-worn sensors. Proc. IEEE Eng. ميد. بيول. شركة نفط الجنوب. 2018 Aug; Milan, Italy. ص. 502–5. 25. Lee JK, Park EJ. A fast quaternion-based orientation optimizer via virtual rotation for human motion tracking. IEEE Trans. بيوميد. المهندس 2009;56(5):1574–82. doi: 10.1109/TBME.2008.2001285. pmid:19473934 26. Elhoushi M, Georgy J, Noureldin A, Korenberg MJ. Motion mode recognition for indoor pedestrian navigation using portable devices. IEEE Trans. Instrum. Meas. 2018;65(1):208–221. 27. Zhang Y, Sun L, Song H, Cao X. Ubiquitous WSN for Healthcare: Recent Advances and Future Prospects. IEEE Internet Things J. 2018;1(4):311–8. 28. Ligorio G, Sabatini AM. A novel Kalman filter for human motion tracking with an inertial-based dynamic inclinometer. IEEE Trans. بيوميد. المهندس 2018;62(8):2033–43. doi: 10.1109/TBME.2018.2411431. pmid:25775483 29. Altini M, Vullers R, Van Hoof C, Van Dort M, Amft O. Self-calibration of walking speed estimations using smartphone sensors. Proc. إيي إنت. أسيوط. Pervasive Comput. COMMUN. 2018 Mar; Budapest, Hungary. ص. 10–8. 30. Laudanski A, Yang S, Li Q. A concurrent comparison of inertia sensor-based walking speed estimation methods. Proc. IEEE Eng. ميد. بيول. شركة نفط الجنوب. 2018 Aug; Boston, MA. ص. 3484–87. 31. Zihajehzadeh S, Loh D, Lee TJ, Hoskinson R, Park EJ. A cascaded Kalman filter-based GPS/MEMS-IMU integration for sports applications. Measurement. 2018;73:200–210. 32. Foxlin E. Pedestrian tracking with shoe-mounted inertial sensors. IEEE Comput. Graph. تطبيق ورقة. 2005;25(6):38–46. pmid:16315476 33. Meng X, Zhang ZQ, Wu JK, Wong WC. Hierarchical information fusion for global displacement estimation in microsensor motion capture. IEEE Trans. بيوميد. المهندس 2018;60(7):2052–63. doi: 10.1109/TBME.2018.2248085. pmid:23446028 34. Skog I, Peter H, Nilsson J, Rantakokko J. Zero-velocity detection—an algorithm evaluation. IEEE Trans. بيوميد. المهندس 2018;57(11):2657–66. 35. Hu JS, Sun KC, Cheng CY. A kinematic human-walking model for the normal-gait-speed estimation using tri-axial acceleration signals at waist location. IEEE Trans. بيوميد. المهندس 2018;60(8):2271–79. doi: 10.1109/TBME.2018.2252345. pmid:23529073 36. Vathsangam H, Emken BA, Spruijt-Metz D, Sukhatme GS. Toward free-living walking speed estimation using Gaussian process-based regression with on-body accelerometers and gyroscopes. Proc. Pervasive Comput. TECHNOL. Healthc. 2018 Mar; Munich, Germany. ص. 1–8. 37. Zihajehzadeh S, Park EJ. Experimental evaluation of regression model-based walking speed estimation using lower body-mounted IMU. Proc. IEEE Eng. ميد. بيول. شركة نفط الجنوب. 2018 Aug; Orlando, Florida. 38. Dadashi F, Millet GP, Aminian K. A Bayesian approach for pervasive estimation of breaststroke velocity using a wearable IMU. Pervasive Mob. Comput. 2018;19:37–46. 39. Rasmussen CC, Williams CK. Gaussian Processes for Machine Learning. Boston, MA: MIT Press, 2006. 40. Vathsangam H, Emken A, Schroeder ET, Spruijt-Metz D, Sukhatme GS. Determining energy expenditure from treadmill walking using hip-worn inertial sensors: an experimental study. IEEE Trans. بيوميد. المهندس 2018;58(10):2804–15. doi: 10.1109/TBME.2018.2159840. pmid:21690001 41. Tibshirani R. Regression selection and shrinkage via the Lasso. J. R. Statist. شركة نفط الجنوب. B 1996;58(1):267–288. 42. Zihajehzadeh S, Loh D, Lee TJ, Hoskinson R, Park EJ. A cascaded two-step Kalman filter for estimation of human body segment orientation using MEMS-IMU. Proc. IEEE Eng. ميد. بيول. شركة نفط الجنوب. 2018 Aug; شيكاغو، IL. ص. 6270–73. 43. Zihajehzadeh S, Lee TJ, Lee JK, Hoskinson R, Park EJ. Integration of MEMS inertial and pressure sensors for vertical trajectory determination. IEEE Trans. Instrum. Meas. 2018;64(3):804–14. 44. Lee JK, Park EJ, Robinovitch SN. Estimation of attitude and external acceleration using inertial sensor measurement during various dynamic conditions. IEEE Trans. Instrum. Meas. 2018;61(8):2262–73. pmid:22977288 45. Jolliffe IT. Principal Component Analysis. New York: Springer-Verlag, 2002. 46. Vathsangam H, Emken BA, Schroeder ET, Spruijt-Metz D, Sukhatme GS. Towards a generalized regression model for on-body energy prediction from treadmill walking. Proc. Pervasive Comput. TECHNOL. Healthc. 2018 Jun; Dublin, Ireland. ص. 168–75.
PLOS is a nonprofit 501(c)(3) corporation, #C2354500, and is based in San Francisco, California, US.