Latent Space Variational Autoencoder for Myoelectric Hand Gesture Recognition

Gesture recognition based on surface electromyography (sEMG) and motion signals have been the primary method to decode human motor intent for applications in prosthetics,wearable robots, and neurorehabilitation. However, the reliability of sEMG signals is often challenged by factors such as motion artifacts, electrode displacement, and signal degradation over time, making robust machine learning and deep learning\u2013based classification methods essential. This study presents a gesture recognition framework that leverages multimodal biosignals to create compact and discriminative latent representations using a contrastive variational autoencoder (VAE) with an integrated classifier. The model was trained on hand gesture datasets and validated both at the individual and group levels. Subject-specific analyses revealed distinct and separable latent spaces with classification accuracies ranging from 75.1% to 91.7%, which were further enhanced through hyperparameter optimization to reach 87.45% to 97.49%. The proposed VAE demonstrated strong generalizability across users, with compressed latent spaces enabling robust and high-performance gesture classification. Beyond methodological contributions, this work highlights the clinical relevance of reliable gesture recognition in neurorehabilitation, where accurate decoding of motor intentions from sEMG and motion data can support personalized therapy, adaptive assistive control, and scalable monitoring of motor recovery