BCIwiki (bciwiki.com) — According to arXiv, Huan Luo, Na Lu, Xiaopeng Wang, and Ziyang Shang uploaded a preprint on October 12, 2026, proposing adaptive time-window optimization for motor imagery (MI) brain-computer interfaces (BCIs). The preprint has not been peer reviewed. The authors note that online systems typically decode continuous electroencephalography (EEG) through fixed-length sliding windows, which may be misaligned with neural events. Evaluating windows independently can also overlook temporal dependencies and produce fragmented or physiologically implausible selections.
Their Temporal Structure Embedded Reinforcement Learning (TSE-RL) framework incorporates prior knowledge of MI event structure into a constrained Markov decision process, a way to make sequential choices under rules. It assigns each window an explicit role through Discard, Onset, Sustain, or Offset actions. A finite-state machine masks invalid actions to enforce the order of MI phases. A Transformer-based history encoder captures dependencies across windows; the authors report improved training stability and computational efficiency compared with recurrent architectures.
The authors describe TSE-RL as plug-and-play and compatible with diverse MI classifiers. They report consistent decoding improvements across multiple benchmark datasets in both cross-session and cross-subject experiments. They also interpret visualization analyses as suggesting that the learned window-selection policy aligns with neurophysiological principles. That interpretation does not directly validate the corresponding neural events, and dataset-level improvements do not establish gains in online use or clinical benefit.