Answers are grounded in this paper and cite their sources.
Objective. This study evaluates explicit adaptive observation–history fusion for continuous EEG-based tracking of PERCLOS-anchored drowsiness and tests whether reward-driven proximal policy optimisation (PPO) provides predictive benefit over supervised fusion and temporal alternatives. Approach. SEED-VIG data from 21 participants and 23 recording sessions were evaluated using strict 21-fold subject-wise leave-one-subject-out (LOSO) validation. In each fold, an EEG encoder generated a PERCLOS-supervised observation proxy, which was recursively combined with the previous estimate through an adaptive observation–history fusion coefficient. Eleven primary methods spanned raw-proxy prediction, conventional filters, supervised and PPO-based fusion, and neural temporal models. Percentage of eyelid closure over time (PERCLOS) provided the ocular-behavioural reference for supervision, PPO reward construction, model selection, and evaluation. Main results. Among methods available in all 21 folds, S4D had the lowest descriptive mean root-mean-square error (RMSE; 0.240746), followed by Gate-7D (0.241490). PPO-7D was available in 20/21 folds. In the prespecified matched comparison (n = 20), the mean RMSE difference (Gate-7D minus PPO-7D) was -0.0022941 (95% bootstrap CI -0.00813595 to 0.00361267; Wilcoxon p = 0.348810), which did not establish a predictive advantage for PPO-7D. None of the 18 broader prespecified PPO-versus-alternative RMSE comparisons met its rejection criterion after Holm correction. Validation-side controlled perturbation altered the EEG-derived proxy, but the prespecified stronger reliability-response criterion was not satisfied. Significance. Explicit observation–history fusion makes the relative contributions of the current EEG proxy and recursive history inspectable. PPO yielded descriptively slower, more persistent fusion-coefficient dynamics than Gate-7D, but its predictive advantage and necessity were not established. The coefficient describes recursive weighting, not calibrated observation reliability.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: