Surface Electromyography-Based Gesture Recognition via Optimizer Selection and Multi-Criteria Performance Analysis

Multichannel surface electromyography-based hand gesture recognition plays an important role in the development of intuitive prosthetic and human-machine interfaces. It requires high accuracy, minimal noise degradation, and electrode drift tolerance for real-world deployment. The goal is to find the best tuned classifier configurations that balance accuracy and real-world robustness, comparing grid search, random search, and Bayesian optimization for the four classical classifiers: K-Nearest Neighbors (KNN), Decision Tree (DT), Random Forest (RF), and Support Vector Machine (SVM). Their performance is evaluated by repeated stratified 5 × 5 cross-validation with 25 runs and leave-one-subject-out testing. Furthermore, we use a multi-objective Pareto front analysis considering three factors: (1) accuracy, (2) adaptation to simulated drift, and (3) relative accuracy degradation under noise. KNN reaches stable 98.4% accuracy on all optimizers, but Bayesian optimization performs better than other optimization techniques in terms of accuracy and other performance metrics when using SVM, DT, and RF classifiers. Additionally, Bayesian optimization gets KNN into the Pareto front with 98.44% accuracy, 98.9% adaptability, and a minimal 0.10% decrease in noise, which likewise clearly improves SVM and RF. These results provide guidance about how to choose effective, real-time electromyography classifiers not only for accuracy but also through trade-offs of accuracy, adaptability, and noise tolerance in prosthetic and rehabilitation systems.