Across-Day Lower Limb Pattern Recognition Performance of a Powered Knee-Ankle Prosthesis

Abstract

Powered lower limb prostheses have the capabilities to assist individuals with a lower limb amputation during ambulation. While these devices can generate power at the knee and/or ankle to assist with incline walking and stair climbing, it is difficult to control the transition between these ambulation modes in a seamless and natural way. Pattern recognition has been suggested as an alternative to using a key fob to switch between modes and recent results have shown reliable performance (less than 5% error rate) across five ambulation modes. In this study we investigated performance of a similar system across multiple sessions of use, a necessary step prior to clinical use. Two individuals with a transfemoral amputation used a powered knee-ankle for five ambulation activities including level-ground walking, ramp ascent, ramp descent, stair ascent, and stair descent over four sessions spaced out over at least two months. An intent recognition system was trained using embedded prosthesis mechanical sensors with varying amounts of data collected across the sessions to determine the effect of multi-session use and increased variation in the activities trained. Overall system error rate decreased from 1.45% [0.3%] when the system was trained with Session 1 data only and tested with Session 4 data to 0.60% [0.02%] when the system was trained with Sessions 1–3 data and tested with Session 4 data. These results demonstrate that a reliable intent recognition system can be created with multiple sessions of use, bringing lower limb intent recognition systems for powered prostheses one step closer to clinical viability.