Análise comparativa de modelos de aprendizado de máquina para a classificação de movimentos finos de mão com sinais de eletromiografia de superfície

Abstract

The intuitive control of multifunctional hand prostheses represents a significant challenge in rehabilitation engineering, especially in the classification of fine hand movements from surface electromyography (sEMG) signals. This work presents a systematic comparative analysis of three machine learning paradigms: Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), and Convolutional Neural Network (CNN). For this purpose, three-channel sEMG signals were acquired from five healthy volunteers during the execution of seven fine hand movements. For the classical models (LDA and SVM), a set of time-domain features was extracted, whereas the CNN was trained end-to-end with the raw sEMG signal. The models were evaluated and compared in terms of classification accuracy, inter-subject robustness, the impact of the number of classes, and computational cost-benefit. The results demonstrated a clear performance hierarchy, with the CNN achieving the highest mean validation accuracy of 96.46%, followed by the SVM at 88.58% and the LDA at 75.12%. The CNN also proved to be the most robust model, showing the lowest performance variability among participants. The cost-benefit analysis quantified the fundamental trade-off, where the CNN's superior performance contrasts with its high training cost compared to the efficiency of classical models. It is concluded that the CNN represents the best-performing paradigm for the classification of fine finger movements, and this study provides a quantitative evidence base to guide the selection of algorithms in the development of future myoelectric interfaces.