Transferring Gait Predictors Across EMG Acquisition Systems with Domain Adaptation

Lower limb assistive technology (e.g. exoskeletons) can benefit significantly from higher resolution information related to physiological state. High-density electromyography (HD-EMG) grids offer valuable spatial information on muscle activity; however their hardware is impractical, and bipolar electrodes remain the standard in practice. Exploiting information rich HD-EMG datasets to train machine learning models could help overcome the spatial limitations of bipolar electrodes. Unfortunately, differences in signal characteristics across acquisition systems prevent the direct transfer of models without a drop in performance. This study investigated Domain Adaptation (DA) to render EMG-based models invariant to different acquisition systems. This approach was evaluated using a Temporal Convolutional Network (TCN) that mapped EMG signals to the subject's knee angle, using HD-EMG as source data and Delsys bipolar EMG as target data. Furthermore, the feature extraction learnt by the TCN was also applied across muscle groups, evaluating the transferability of the sensor agnostic features. The DA implementation shows promise in both scenarios, with an average increase in accuracy (angular error normalised by the range of motion) of 7.36% for the Rectus Femoris, Biceps Femoris and Tibialis Anterior, as well as a cross-muscle performance increase of up to 10.80%. However, when the domain discrepancy is severe, the model is currently unable to generate a reliable walking trajectory due to inherent limitations related to the applied regression scheme and the chosen Mean Squared Error loss function. Therefore, future research should focus on exploring advanced loss functions and classification-based DA models that prioritise restoring key features of the gait.