Brain-Muscle Atlas: A novel framework for motor brain-computer interfaces.

Although substantial progress has been achieved in the field of brain-computer interfaces (BCIs), conventional approaches remain limited in decoding stability and physiological interpretability, particularly in continuous motor control tasks. Most existing studies adopt an end-to-end mapping paradigm from neural signals directly to device commands, without explicitly modelling the intermediate muscle-level transformation underlying natural movement. This omission constrains both the structural plausibility and the explanatory capacity of the resulting models. To address this limitation, we propose a Brain-Muscle-Elbow Interface framework that reconstructs muscle-level electromyographic (EMG) activity from electroencephalography (EEG), thereby introducing an explicit myoelectric intermediate representation within the decoding pipeline. Specifically, we develop a Brain-Muscle Atlas (BMA) model incorporating a temporal self-attention mechanism to reconstruct integrated EMG (iEMG) waveforms from EEG signals. In an offline experiment with 49 participants performing elbow flexion-extension tasks, the iEMG signals reconstructed by the BMA model showed a moderate correlation with the corresponding ground-truth iEMG signals, achieving a maximum correlation coefficient of 0.83. Furthermore, in a real-time online control experiment with 10 participants, the proposed method enabled stable and continuous control of a virtual elbow joint, demonstrating its feasibility for online application. Cortical region masking analysis revealed that the sensorimotor cortex consistently played a critical role in model prediction, supporting the physiological plausibility of the learned mapping. Collectively, these findings indicate that explicit modelling of the muscle-level transformation enhances both the interpretability and the structured motor decoding capability of BCI systems.