Neuromechanical-Driven Ankle Angular Position Control During Gait Using Minimal Setup and LSTM Model

Data-driven decoding of lower limb muscles surface electromyography (sEMG) and joints kinematics is a crucial approach for enhancing prosthetic control and assistive reha-bilitation. In this study, three different neuromechanical-driven ankle angle estimation strategies were investigated, i.e. Least Squares - Support Vector Machine (LS-SVM) fed with Time Domain (TD) features, single hidden layer Long Short Term Memory (1HL-LSTM) deep learning model fed within the same features, and 2 hidden layers LSTM (2HL-LSTM) fed with a sequence of raw data windows. The above mentioned schemes were tested with three myoelectric-mechanical combinations, i.e. three thigh muscles (3M), their fusion with hip joint angle (3M+Hip), and Biceps Femoris muscle with the hip (BF+Hip). A combined TD features with 1HL-LSTM has outperformed the other two models in all three input combinations, with an RMSE equal to 2.18 ± 0.44 deg in the case of BF+Hip as input, that appeared to be the best muscles configuration. The results of this study provide a robust ankle angle estimation strategy while preventing the computational cost from being drastically increased. They also highlight the reliability of the BF + Hip combination, thus representing a further step in the advancement of effective control strategies for active ankle prosthesis and rehabilitative devices.