Validation of EEG mental workload markers using integrated statistical and machine learning analyses

Attentional and working-memory processes can be monitored noninvasively using electroencephalography (EEG), which provides physiological indices of mental workload. Prior studies consistently report increased frontal-midline theta and beta power together with suppression of posterior alpha activity during cognitively demanding tasks. However, most investigations rely either on group-level statistical analyses or on machine-learning (ML) classification alone, often without examining whether the predictive features identified by ML models correspond to established neurophysiological markers.

This study reanalyzed an open EEG dataset comprising 36 young adults performing a mental arithmetic task. EEG activity was quantified using power spectral density (PSD) estimation based on Welch's method (1-second Hamming windows with 50% overlap) across canonical frequency bands: delta (1-4 Hz), theta (4-8 Hz), alpha (8-12 Hz), beta (13-30 Hz), and gamma (30-50 Hz). Event-related potentials (ERPs) time locked to arithmetic stimulus onset were also examined. Spectral features were subsequently used to train three ML classifiers Logistic Regression, Support Vector Machine (SVM), and Random Forest using subject-level cross-validation to distinguish resting and task conditions.

ERP analysis revealed early stimulus-locked modulations within the 20-50 ms latency range, reflecting rapid sensory engagement during task performance. Spectral analysis demonstrates significant workload-related changes: theta power increased (η² = 0.49), alpha power decreased (η² = 0.37), beta power increased (η² = 0.18), and delta power decreased (η² = 0.38), whereas gamma-band differences did not remain significant after correction for multiple comparisons. Among the ML models, Random Forest achieved the highest classification performance (accuracy = 0.92 ± 0.03, AUC = 0.94). Feature-importance analysis indicated that theta and alpha band powers contributed most strongly to classification, consistent with the statistical findings.

The results replicate well-established EEG signatures of cognitive workload and demonstrate convergence between statistical inference and machine-learning prediction. The alignment between physiological interpretation and predictive modeling supports frontal theta enhancement and posterior alpha suppression as reliable indicators of cognitive engagement. These findings highlight the potential of EEG-based workload monitoring for healthcare and applied neuroscience applications, including early detection of cognitive decline and neurorehabilitation monitoring. Nevertheless, the modest sample size and single dataset design warrant cautious interpretation and future validation in larger and independent cohorts.