Activity Recognition for Trans-Tibial Prosthesis Users

Abstract

Prosthetists and medical researchers often find it difficult to assess the results of specific prosthesis prescriptions or therapies, which causes difficulties in medical decisions regarding an amputee. Currently, subject self-reporting is often used to determine how well a particular prosthesis is performing. Subjects may also be asked to perform certain actions in the clinic to assess the effectiveness of a given prosthesis. These methods are limited because self-reporting may be unreliable and experiments in a lab or doctor's office do not adequately approximate behavior and activities in a subject's daily life. To determine how subjects use their prostheses outside of the clinic, it has become common to use accelerometers to enable long-term monitoring of a subject's activities. Accelerometers can offer information about how often the prosthesis is worn, how active the subject is over time, and what body positions subjects take. Currently available systems tend to only measure activity level or to require subjects to wear an accelerometer (or multiple accelerometers) at inconvenient positions on their body. The use of a single accelerometer mounted on a prosthesis may avoid the high compliance demands of similar methods while still allowing for accurate classification. We propose and test three methods of classifying the activities and body postures of prosthesis users using only data from a single prosthesis mounted accelerometer. The classification methods tested were a rule-based system, a K-Nearest Neighbor system, and a Hidden Markov Model. Personalized and non-personalized classifiers were tested for the K-Nearest Neighbor and Hidden Markov Model systems. All three classification methods were tested on data collected from eleven subjects with trans-tibial amputation. Data was collected while the subjects performed an hour long semi-structured activity protocol. The activity protocol included donning and doffing the prosthesis, sitting in various positions, standing, walking, and using the stairs. The predicted activity from each classification method was compared to the ground truth record and the overall accuracy for each method was calculated. All three methods were able to classify activities with approximately 90% accuracy. Personalization improved the accuracy of both K-Nearest Neighbor and Hidden Markov Model classifiers. All classifiers were able to accurately differentiate periods of motion, but periods where the prosthesis was doffed, the subjects were sitting, or the subjects were standing were sometimes confused. These methods may be applicable for clinical use, and may improve the quality of care available to prosthesis users.