Event Title

Improving Wearable Activity Sensors using Hidden Markov Models for Outpatient Physical-Therapy Applications

Session Number

F04

Advisor(s)

Mark Albert, Loyola University
Jessie Pichleap, Loyola University

Location

A-115

Start Date

28-4-2016 9:50 AM

End Date

28-4-2016 10:15 AM

Abstract

The purpose of this investigation is to improve the accuracy of the software used in recognizing patient activities in physical therapy in order to allow for the collection of objective and continuous activity data in a manner which is convenient for the subject. Patients’ activities are traditionally predicted simply using static classifiers such as Support Vector Machines (SVMs). Studies have shown the static classifiers are prone to mistakes which an additional model can correct, a Hidden Markov Model (HMM). In our experiments, the accuracy of the activity recognition software was determined using data containing patient’s movements-such as walking, running, sitting, etc. Our results have confirmed this hypothesis, with an increase to 86% accuracy with the HMM vs. 84% accuracy without. Due to the success of the model, doctors will be able to access unbiased and continuous data regarding how patients go about their prescribed activities without having to meet with them regularly or rely on inaccurate patient reports. By using HMMs to correct mistakes made by static classifiers analyzing activity recognition data doctors will have more accurate profiles of patients’ activities as well as the patient’s reactions to treatments while increasing their comfort and freedom.


Share

COinS
 
Apr 28th, 9:50 AM Apr 28th, 10:15 AM

Improving Wearable Activity Sensors using Hidden Markov Models for Outpatient Physical-Therapy Applications

A-115

The purpose of this investigation is to improve the accuracy of the software used in recognizing patient activities in physical therapy in order to allow for the collection of objective and continuous activity data in a manner which is convenient for the subject. Patients’ activities are traditionally predicted simply using static classifiers such as Support Vector Machines (SVMs). Studies have shown the static classifiers are prone to mistakes which an additional model can correct, a Hidden Markov Model (HMM). In our experiments, the accuracy of the activity recognition software was determined using data containing patient’s movements-such as walking, running, sitting, etc. Our results have confirmed this hypothesis, with an increase to 86% accuracy with the HMM vs. 84% accuracy without. Due to the success of the model, doctors will be able to access unbiased and continuous data regarding how patients go about their prescribed activities without having to meet with them regularly or rely on inaccurate patient reports. By using HMMs to correct mistakes made by static classifiers analyzing activity recognition data doctors will have more accurate profiles of patients’ activities as well as the patient’s reactions to treatments while increasing their comfort and freedom.