Sound-side Limb Gait Analysis Using Machine Learning In Lower Limb Prosthesis

Unilateral lower limb amputees face a higher risk of joint degradation due to uneven loads while walking. Existing research often focuses on joint moment predictions but lacks comprehensive assessment of joint work, which is crucial for understanding energy dynamics and joint health. This thesis addresses this gap by developing and validating a machine learning pipeline to estimate joint work from onboard prosthesis sensors, eliminating the need for external devices. Our approach provides a joint estimation method that can be used to optimize prosthetic assistance and promote healthier gait patterns for real-time feedback. We hypothesize that joints proximal to the prosthesis, such as the prosthesis-side hip, will show higher accuracy in joint work estimation due to sensor proximity, while distal joints will be less accurate without sound-side data. We tested three machine learning models on seven subjects with varying sensor datasets. CNN models significantly outperformed TCN models in predicting joint work, with sound-side sensors improving accuracy for distal joints like the sound ankle (p < 0.05). Our results confirm that while a minimal sensor set achieves near-optimal accuracy, additional sound-side sensors enhance predictions for distal joints. This study introduces a novel method that reduces reliance on costly motion capture systems, providing a method for real-time feedback for prosthetic optimization and aiding in more balanced and efficient ambulation for amputees.

Date

2024-07-25

Resource Type

Text

Resource Subtype

Thesis

Full item page