Calibration‐Free Electromyography Motor Intent Decoding Using Large‐Scale Supervised Pretraining

Machine learning algorithms for myoelectric pattern recognition require substantial user‐specific training data, limiting broader applications of electromyography (EMG) in human–computer interfacing. Here, we present a framework for EMG‐mediated motor intent decoding designed to function for new users without collecting user‐specific training data. We introduce a Transformer‐based architecture, termed the Spatially Aware Feature‐learning Transformer (SAFT), which processes EMG time windows with variable numbers of channels from arbitrary spatial electrode configurations by combining channel‐wise temporal feature extraction with learned spatial encoding of electrode positions and attention across channels. This enables training of a single model across heterogeneous EMG datasets. In the present study, large‐scale supervised pretraining refers to pretraining on a pooled corpus of 29 public EMG databases comprising 506 subjects, 108 movement classes, and ≈9.9 million nonrest EMG windows after preprocessing. A pretrained SAFT model was fine‐tuned on a held‐out database and evaluated for cross‐user performance. On the 3DC benchmark, the pretrained‐only model achieved 28.7% balanced accuracy (vs. 10% chance), while pretrained and fine‐tuned cross‐user SAFT models achieved 81.8% balanced accuracy, comparable to conventional user‐specific linear discriminant analysis (LDA) models (82.9%). These findings indicate the feasibility of EMG intent decoding models that work “out of the box” without end‐user calibration.