Utilizing machine learning techniques in estimating ankle prosthesis power output using wearable IMU sensors

Ankle power output can be used to identify abnormalities and uneven weight distribution in gait. The acquring of the ankle power output usually involves using invasive motion capture systems along with force plates to capture the exerted power on the ground during stance phase. When clinicians are fitting prosthesis on to amputees it can be hard to produce real data that could be used to assist with the fitting and prosthesis choosing process. Motion capture is not used for this process due to the invasiveness of the technique. This research explores the possibility of using machine learning techniques to estimate the prosthesis power output using IMU sensors attached to the prosthesis. IMU data was collected from 9 able body subjects and 2 TT amputees using 2 different prostheses. This research explores a way of bypassing the need for collecting motion capture data from TT amputees by using able body subjects instead since getting the same amount of motion capture data from TT amputees is extremely hard. 2 machine learning algorithms were optimized, trained and tested, namely Random Forest quantile regression and Long short term memory Recurring neural network. The input data in these two machine learning algorithms was the IMU data, for each algorithm different preprocessing of the IMU data was used. This research focused on the model’s ability to predict the peak power output. The best performing LSTM models used to predict ankle power in able body subjects gave an accuracy of 55.9% that the prediction would be within a 0.25 W/kg range. For TT amputees using prosthesis 1, the best performing LSTM model gave an accuracy of 52.6% that the prediction would be within a 0.25 W/kg range. For TT amputee using prosthesis 2, the best performing LSTM model gave an accuracy of 47.9% that the predictions would be within a 0.25 W/Kg range. The trained LSTM model managed to achieve an average R^2 score of 76 for the 3 LSTM models.

Author dc:creator

• Halldór Reynisson 1998-

Contributors dc:contributor

• Háskólinn í Reykjavík

• Heilbrigðisverkfræði
• Meistaraprófsritgerðir
• Gerviliðir
• Prosthesis
• Hreyfifærni
• Vélrænt nám
• Biomedical engineering
• Motion
• Machine learning

Language dc:language.iso

en

Handle dc:identifier.uri

http://hdl.handle.net/1946/46270

OAI identifier oai:identifier

oai:skemman.is:1946/46270