Transformer-Based Context-Informed Incremental Learning With sDTW Alignment Unlocks Fast and Precise Regression-Based...

Regression-based myoelectric interfaces hold the promise of enabling intuitive proportional and simultaneous control but remain limited by calibration sensitivity, unpredictable dynamics, and inconsistent user behaviours. Temporal neural architectures have the potential to substantially improve these controllers by capturing the temporal structure of user behaviours, provided they are trained using dynamics that are sufficiently representative of closed-loop use. Context-informed incremental learning (CIIL) offers a mechanism for acquiring such data online; however, its reliance on environment-derived pseudo-labels makes it vulnerable to temporal deviations between assumed and true user intent. This study introduces T-sDTW-CIIL, a transformer-based incremental learning framework that integrates temporal modelling, closed-loop learning, and soft dynamic time warping (sDTW) to enable tolerant label alignment. Twelve participants completed an adaptive regression-based cursor-control task using four pipelines: static and CIIL variants of both MLP and transformer models. T-sDTW-CIIL achieved significantly higher success rates, throughputs, efficiencies, and simultaneity gains when evaluated in a high precision ISO-Fitts' environment. T-sDTW-CIIL achieved throughputs of $2.0\times $ , $2.4\times $ , and $3.7\times $ those of an MLP trained using conventional screen-guided training when acquiring large, medium, and small targets, respectively. Perhaps more importantly, it maintained success rates of 98.4% for small targets, whereas the static MLP degraded to only 23.4% success. T-sDTW-CIIL-based adaptation also reduced overall contraction intensities by ~10%. These results demonstrate the powerful combination of temporal learning with context-informed co-adaptation. T-sDTW-CIIL overcomes key limitations of existing regression-based myoelectric controllers, enabling robust, low-intensity human-computer interaction.