sEMGCareHCI: sEMG-based finger position prediction for amputee-centric human-computer interaction

sEMGCareHCI presents a low-cost, non-invasive surface electromyography (sEMG)-based system for recognizing finger positions and gestures from forearm muscle activity. The proposed "Spatio-Temporal Attention Model (STAM)" combines handcrafted time-domain features, autoencoder-derived latent representations, and continuous wavelet transform (CWT) scalograms to predict accuracy and robustness, demonstrating strong potential for prosthetic control, rehabilitation, and human-computer interaction applications. These heterogeneous representations are projected into a shared feature space, where graph convolutional layers model inter-electrode spatial relationships, CNN-TCN modules capture temporal muscle activation patterns, and an attention-based fusion mechanism adaptively weights spatial, temporal, and time-frequency information for gesture-specific classification. To establish this design, the first benchmark five classical classifiers (SVM, KNN, Decision Tree, Random Forest, MLP) trained on handcrafted features, identifying SVM with the Waveform Length feature as the strongest baseline (84.1% ± 1.9 offline cross-validation accuracy). STAM is evaluated against this baseline and against several deep-learning alternatives (CNN-BiLSTM, TCN, Transformer Encoder, GNN, Ensemble Fusion). The framework enhances signal quality and analog-to-digital conversion accuracy, with the observed features conceptually explained using a Lagrangian dynamics-based biomechanical model. The proposed STAM architecture produced continuous gesture predictions with an accuracy of 90.4%, demonstrating its potential for amputee-centric real-time human-computer interaction and assistive control applications.