Participant-Independent Spatiotemporal Neural Modeling of Graded Occupational Risk Judgments from EEG: A Leakage-Controlled Comparative Framework

Self-reported perceived-risk responses form an ordered behavioral outcome in which classification errors differ in both direction and magnitude. Existing EEG-based safety studies have mainly considered hazard categories or binary states, with limited attention to participant-independent modeling of graded perceived-risk responses. A leakage-controlled comparative modeling framework was evaluated for four self-reported response levels—high risk, medium risk, low risk, and safe—using 2634 non-overlapping 3 s stimulus-locked EEG epochs from 31 workers with complete usable recordings. The signals contained 16 channels sampled at 250 Hz. Because participants could respond during the 3 s analysis interval, the epochs typically contained the behavioral response and some post-response activity; the task should therefore be interpreted as decoding response-associated perceived-risk states rather than prospective pre-response prediction of objective workplace risk. Four spatiotemporal neural architectures, EEGNet, DeepConvNet, ShallowConvNet+SE, and EEG-TCN, were evaluated against SVM-RBF, gradient-boosted decision trees, and random forest models trained on 80 full-window power spectral density features. Participant identities were separated in an outer three-fold cross-validation scheme, with inner-participant validation for early stopping, and the procedure was repeated across three participant permutations. One out-of-fold probability vector was obtained for each trial in each permutation, and the three vectors were averaged to generate a consensus prediction. EEGNet achieved the highest trial-pooled balanced accuracy of 0.7447 (95% CI, 0.6796–0.7769), together with a Macro-F1 of 0.7494, Macro-AUROC of 0.9322, and Macro-AUPRC of 0.8295. EEG-TCN reached a balanced accuracy of 0.7333 and the highest participant-equal mean balanced accuracy (0.6644 ± 0.1090); its participant-level difference from EEGNet was not significant after Holm correction. Medium-risk responses remained the least separable category. Under the shared-scene, response-inclusive setting examined here, compact spatiotemporal neural models provided the strongest participant-independent decoding performance, while the change in model ranking between trial-pooled and participant-equal evaluation highlighted the influence of the selected estimand on model interpretation.