ABSTRACT
The quality of extracted traditional electromyogram (EMG) features has recently been identified in the literature as a limiting factor preventing translation from laboratory to clinical settings. Deep learning was implemented to overcome the limitation. While deep learning models produce promising results on raw EMG data, their clinical deployment is frequently hampered by their relatively high computational costs (a substantially high quantity of parameters for the models and a large quantity of information required for training). A shift in attention from traditional feature extraction approaches to deep learning models was observed to solve this limitation. This chapter focuses on efficiently extracting the spatial-temporal dynamics of EMG signals by integrating the simplicity and low computational characteristics of traditional feature extraction methods with memory principles from deep learning models. The novelty of the proposed method can be summarised as follows: (i) the invention of a new novel cardinality-based feature extraction method; (ii) the memory concept leveraged from deep learning structures, which captures long- and short-term temporal dependencies of the EMG signals; and (iii) the use of cardinality to generate logical combinations of spatially distinct EMG signals. The accuracy of the proposed method is shown to be significantly improved reaching as high as 98%.
