Gait Change Detection: Classification and Clustering of Wearable Sensor Data

Abstract:

Inexpensive, wearable sensors that can alert health care workers about a patient’s change in health in real time have clear benefits in medical care. Pressure sensors that can be put in the soles of a shoe produce a noisy, time-series of data that can be used to both learn predictable patterns of behavior and also detect anomalous behavior. A possible application is identifying a negative reaction to a change in medication in Alzheimer’s patients hours or days after administering the new dose. This research presents an approach to the challenges of data cleaning, separation, and feature synthesis of the sensor signal; and shows the results of machine learning techniques used for both identifying normal patterns of behavior, and detecting anomalous activity. Classification algorithms of labeled data recorded from students using a treadmill will be used to show that the signal has predictive power; and dimensionality reduction and clustering algorithms will be used to predict anomalous activity.