Characterization of Knee and Gait Features From a Wearable Tele-Health Monitoring System

Abstract

Mobility is crucial for healthy aging. Any disruption to mobility can affect mental, physical and social health, and socio-economic independence. Therefore, studies in gait and lower-joint functionality with respect to different demographic features will play a vital role in maintaining good mobility. In this study, we analyzed a gait database from 70 healthy subjects (18–86 years) constructed using our custom-built multi-sensor-based wearable tele-health monitoring system. The purpose was to extract and use the most informative features for classifying knee joint and gait characteristics of the subjects with respect to their age, body mass index – BMI, and sex. Four supervised machine learning algorithms: partial least square-discriminant analysis (PLS-DA), support vector machine (SVM), random forest (RF), and artificial neural network (ANN) were used to classify the subjects. The features that significantly contributed to all classifications are knee angle, quadriceps muscle pressure adjacent to the knee joint, rotational energy (mediolateral and vertical), acceleration energy (mediolateral), cross-sample entropy (anteroposterior-mediolateral), knee angle variability, symmetry of swing and stance phase, and walk ratio. Classification accuracies of all four methods were ~89%, 83%, 81%, 86% for age, 90%, 80%, 83%, 86% for BMI, and 97%, 97%, 96%, 97% for sex, respectively. PLS-DA had the best classification performance for all three categories which makes it preferable for these kinds of analyses. Thus, our knee and gait monitoring system coupled with an efficient machine learning tool can be exploited for real-time evaluation and early diagnoses of mobility disabilities, health assessment, and monitoring the need for interventions.