A Multi-Dataset Characterization of Window-based Hyperparameters for Deep CNN-driven sEMG Pattern Recognition

The control performance of myoelectric prostheses would not only depend on the feature extraction and classification algorithms but also on interactions of dynamic window-based hyper-parameters (WBHP) used to construct input signals. However, the relationship between these hyper-parameters and how they influence the performance of the convolutional neural networks (CNNs) during motor intent decoding has not been studied. Therefore, we investigated the impact of various combinations of WBHP (window length and overlap) employed for the construction of raw 2-dimensional (2D) surface electromyogram signals on the performance of CNNs when used for motion intent decoding. Moreover, we examined the relationship between the window length of the 2D sEMG and three commonly used CNN kernel sizes. To ensure high confidence in the findings, we implemented three CNNs which are variants of the existing models, and a newly proposed CNN model. Experimental analysis was conducted using three distinct benchmark databases, two from upper limb amputees and one from able-bodied subjects. The results demonstrate that the performance of the CNNs improved as the overlap between consecutively generated 2D signals increased, with 75% overlap yielding the optimal improvement by 12.62% accuracy and 39.60% F1-score compared to no overlap. Moreover, the CNNs performance was better for kernel size of seven than three and five across the databases. For the first time, we have established with multiple evidence that WBHP would substantially impact the decoding outcome and computational complexity of deep neural networks, and we anticipate that this may spur positive advancement in myoelectric control and related fields.