Evaluating Bilateral Surface EMG Features for Automatic Identification of Gait Phase Transitions in Ground Walking Conditions

Abstract

The identification of gait events is essential to achieve a synchronization of muscular recruitment during human walking. The current study aims to introduce a machine learning approach to identify the transition events between the four main gait phases, i.e., Heel Strike (HS), Counter Lateral Toe Off (CTO), Heel Rise (HR), and Toe Off (TO). As far as we know, this is the first study trying to pursue this objective. To this purpose, a multilayer perceptron (MLP) model with 1 hidden layers composed of 32 neurons was adopted to interpret surface electromyographic signals (sEMG) collected bilaterally in five muscle of 31 healthy subjects during walking. Standard metrics as F1-score and mean absolute error (MAE) were adopted to evaluate model performance in detecting gait events. Results showed as the proposed approach was able to provide encouraging performances in prediction of HS, HR and TO (mean F1-score ≈86% and MAE < 30 ms). Specifically, performances of HS and TO identification were comparable with those achieved in the simpler task of binary classification. This is a significant result, since these two events play a key role in the characterization of the gait cycle. Only CTO prediction was affected by a reduction of F1-score (≈ 80%) and an increase of the error (MAE = 38.2 ms). Despite the encouraging results, further studies should be projected to improve the prediction performances, especially for CTO, by adopting new machine learning architectures, model parameter tuning, and post-processing algorithms to clean the model output.