On Tiny Feature Engineering: Towards an Embedded EMG-Based Hand Gesture Recognition Model

Robotic-based therapy is becoming a very popular treatment for the rehabilitation of hand function impairment as a consequence of strokes, due to the high-dimensional and less interpretable nature of superficial electromyography (sEMG) signals. In this context, feature engineering becomes particularly important to estimate the intention of upper limb movements by utilizing machine learning models, specially when a hardware embedded on-board implementation is expected, due to the strong computational, energy and latency constraints. Our work compares the performance achieved by implementing four state-of-the-art feature techniques (random forest, minimum redundancy maximum relevance (MRMR), Davies-Bouldin index, and t-tests), when these are evaluated on a sEMG dataset intended for training hand gesture classifiers. The results of three different machine learning algorithms (neuronal networks, k-nearest neighbors and bagged forest) are used as a reference to validate the analysis. This ongoing research has revealed valuable information on the potential and constraints of these 4 feature generation methods for real gesture recognition embedded applications.