Prediction of ankle kinematics and kinetics in stair ascent motion using surface EMG feature inputs of lower extremity muscle combinations.

Abstract

Recent studies show the relevance of using sEMG signals and extracted features in powered prosthesis control. However, amputation technique determines the availability of residual muscles, and minimizing sensor number is to be preferred. Therefore, (i) limiting the total number of sEMG muscle inputs, (ii) minimizing the inputs from the lower leg muscles, and (iii) completely excluding the lower leg muscles will make the control algorithm economical, flexible, and practical, respectively. Using healthy population data, the aim was to conduct a systematic analysis for ranking possible muscle combinations based on prediction success for ankle kinematics and kinetics during stair ascending motion. We used eight muscle inputs in a long short-term memory (LSTM): rectus femoris (RF), vastus medialis (VM), tibialis anterior (TA), peroneus longus (PL), gluteus maximus (Gmax), biceps femoris (BF), medial gastrocnemius (MG), and soleus (SOL). sEMG feature and muscle combinations were ranked based on Pearson's correlation coefficient (r > 0.90 indicates successful correlation) and root-mean-square-error. The best-performing involved several muscles: TA + SOL + MG + PL + VM + BF + Gmax (r_position = 0.93, r_moment = 0.95). However, a single muscle also performed successfully: SOL (r_position = 0.91, r_moment = 0.94) economical variation. SOL + BF (r_position = 0.92, r_moment = 0.94) was the flexible variation. The results confirm the successful use of sEMG also for a highly demanding motion but eliminate a practical variation.