sEMG-based ankle position and moment prediction in silico: Neural network approach and muscle selection

Lower limb amputation is the partial or complete removal of a limb, and powered prostheses are the best solution for restoring amputees' locomotion abilities. Although recent advancements have enhanced their hardware, autonomous adaptation is required to achieve natural ambulation. The utilization of surface electromyogram (sEMG) holds promise, whereas real-time analysis is challenging. Also, a systematic analysis should be conducted for muscle selection to ensure compatibility with different levels of amputations. Therefore, the feature extraction was implemented for non-normalized sEMG amplitudes, and an economic algorithm minimizing sEMG input was sought. For the sake of different amputation level compatibility, a practical algorithm was aimed to limit the use of lower leg muscles. In this context, neural network-based algorithms with timing-based approaches utilizing sEMG amplitudes as inputs have been developed to (1) predict sagittal ankle position and moment during ground-level walking and (2) rank all muscle combinations based on success. Eight leg muscles were studied: tibialis anterior (TA), soleus (SO), medial gastrocnemius (MG), peroneus longus (PL), rectus femoris (RF), vastus medialis (VM), biceps femoris (BF) and gluteus maximus (GMax). The results showed the best-performing muscle variation was MG+RF+VM whereas, PL and GMax+VM were distinguished as the economic and practical variations (rposition>0.90, rmoment>0.97), respectively. The analysis regarding the effect of window size selection for feature extraction on prediction accuracy revealed that a window size of 150 ms demonstrated the best performance for the proposed neural network architecture. The cross-validation results supported the repeatability of the structure and methodology.