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

Area of Science:

• Neuroscience
• Cognitive Science
• Biomedical Engineering

Background:

• Electroencephalography (EEG) offers noninvasive monitoring of attentional and working-memory processes, providing physiological markers for mental workload.
• Previous research consistently shows increased frontal-midline theta and beta power, alongside suppressed posterior alpha activity, during demanding cognitive tasks.
• Existing studies often use group-level statistics or machine learning (ML) classification independently, sometimes overlooking the neurophysiological validity of ML-identified features.

Purpose of the Study:

• To reanalyze an existing EEG dataset to investigate cognitive workload during a mental arithmetic task.
• To examine the convergence of traditional statistical analyses and ML classification in identifying EEG markers of mental workload.
• To validate established neurophysiological markers of cognitive engagement using both statistical inference and predictive modeling.

Main Methods:

• Reanalyzed EEG data from 36 young adults performing a mental arithmetic task.
• Quantified EEG activity using power spectral density (PSD) across delta, theta, alpha, beta, and gamma frequency bands.
• Trained three ML classifiers (Logistic Regression, SVM, Random Forest) using spectral features to differentiate resting and task conditions, employing subject-level cross-validation.

Main Results:

• ERP analysis revealed early (20-50 ms) stimulus-locked modulations, indicating rapid sensory engagement.
• Significant workload-related spectral changes included increased theta (η² = 0.49) and beta power (η² = 0.18), and decreased alpha (η² = 0.37) and delta power (η² = 0.38).
• The Random Forest model achieved the highest classification performance (accuracy = 0.92 ± 0.03, AUC = 0.94), with theta and alpha band powers being the most important features.

Conclusions:

• The findings replicate known EEG signatures of cognitive workload and demonstrate consistency between statistical and ML approaches.
• Frontal theta enhancement and posterior alpha suppression are confirmed as reliable indicators of cognitive engagement.
• EEG-based workload monitoring shows potential for applications in healthcare and neuroscience, such as early cognitive decline detection and neurorehabilitation, though further validation is needed.