Causal Sequence Modeling of Electromyography Signals Using Autoregressive Transformer With Patch‐Based Tokenization for Hand Gesture Recognition

Utilizing electromyographic signals as a method of recognizing hand gestures has created an intuitive interface for human–computer interaction, rehabilitation, and prosthetic control. The use of convolutional, recurrent, and attention‐based deep learning models for identifying hand gestures is widespread, and they represent electromyography (EMG) signals as fixed or aggregated features. Relatively, deep learning models that capture causal temporal dependencies, context evolution, and fine‐grained muscle activation dynamics would be more beneficial, but they are explored less. This research presents a generative transformer framework for effectively recognizing hand gestures using EMG signals by employing the generative pretrained transformer decoder‐only architecture using autoregressive learning. The generative transformer framework uses preprocessing methods for the extraction of EMG signals, including band‐pass filtering, rectifying, and interquartile range normalization, utilizing a transformer model to encode the time‐dependent nature of EMG signals in a tokenized manner and incorporating positional information. This study was conducted on 24,000 EMG samples obtained from the publicly available EMG‐EPN‐612 dataset, which consists of eight channels of EMG signals collected from 612 participants performing six different hand gestures. In addition, two more datasets are used for the validation of the proposed model. The novelty of this work lies in the exploration of autoregressive, decoder‐only sequence modeling for EMG‐based hand gesture recognition, framing muscle activity as a generative temporal process rather than a static classification problem. With a test accuracy of 96.81%, this study shows that transformer‐based learning methods can accurately model complex spatiotemporal muscle patterns.