An analysis of human movement accelerometery data for stroke rehabilitation assessment

Abstract

Human Activity Recognition (HAR) is concerned with the automated inference of what a person is doing at any given time. Recently, small unobtrusive wrist-worn accelerometer sensors have become affordable. Since these sensors are worn by the user, data can be collected, and inference performed, no matter where the user may be. This makes for a more flexible activity recognition method compared to other modalities such as in-home video analysis, lab-based observation, etc. This thesis is concerned with both recognizing subjects activities as well as recovery levels from movement-related disorders such as stroke. In order to perform activity recognition or to assess the degree to which a subject is affected by a movement-related disease (such as stroke), we need to create predictive models. These models output either the inferred activity (e.g. running or walking) in a classification model, or else the inferred disease recovery level using either classification or regression (e.g. inferred Chedoke Arm and Hand Activity Inventory Score for stroke rehabilitation assessment). These models use preprocessed data as inputs, a review of preprocessing methods for accelerometer data is given. In this thesis, we provide a systematic exploration of deep learning models for HAR, testing the feasibility of recurrent neural network models for this task. We also discuss modelling recovery levels from stroke based on the number of occurrences of events (based on mixture model components) on each side of the body. We also apply a MultiInstance Learning model to model stroke rehabilitation using accelerometer data, which has both visualization advantages and the potential to also be applicable to other diseases.