Single-channel gamma burst dynamics and adaptive deep learning for robust EEG motor state decoding

Reliable motor decoding from single-channel EEG remains challenging due to low signal quality and strong inter-session variability. Traditional ERD/ERS features overlook transient high-frequency events that reflect key motor processes. This paper introduces a compact decoding framework that models gamma burst dynamics-including burst onset, duration, amplitude, and recurrence-as informative biomarkers of motor preparation and execution. Bursts are extracted using an analytic wavelet transform with adaptive thresholding, and combined with classical spectral features through a multi-stream adaptive deep learning model equipped with instance normalization and few-shot recalibration for session robustness. Experiments on the WAY-EEG-GAL dataset show that the proposed approach achieves competitive performance under a strict single-channel constraint and significantly improves cross-session accuracy by more than 10% over non-adaptive baselines. These findings highlight the functional relevance of gamma bursts and demonstrate the feasibility of practical, lightweight EEG-based motor BCIs.