Optimal Dataset Size for Fine-Tuning sEMG-Based Hand Gesture Recognition in Rehabilitation Prosthesis

Surface electromyography (sEMG) has become a vital tool for controlling prostheses and rehabilitation using hand gesture recognition. However, the process of fine-tuning machine learning models to individual users often requires considerable amounts of data, which can be challenging to obtain due to user fatigue and discomfort. This work investigates the optimal dataset size needed for fine-tuning a pretrained Convolutional Neural Network (CNN) model for hand gesture recognition, using the NinaPro DB2 dataset. Our results show that training on just a third of the dataset achieves over 90% accuracy, highlighting a significant reduction in the data requirements compared to traditional methods. This approach can minimize the burden of data collection on users, making sEMG-based rehabilitation devices more practical and accessible.