Early Mental Fatigue Detection Using Multimodal Physiological Signals: From Subject-Dependent to Subject-Independent Validation

Mental fatigue a gradual cognitive decrement due to extensive cognitive engagement accounts for 13 % of occupational accidents in safety critical industries, but traditional subjective evaluation cannot guarantee real time observations. Although some recent machine learning solutions show promising evaluation results on classifications of bio indices, they mostly opt for subject dependent evaluation protocols that validate model generalization to novel subjects, but that is an essential aspect for real world deployments. In this study, we bridge this gap by evaluating both subject dependent and subject independent evaluation protocols based on 923,298 instances of multimodal bio indices. Experimental findings show a precipitous drop of 26.59 % over a$\mathbf{2 6. 5 9 \%}$accuracy difference between subject dependent (99.48 %) and subject independent (72.89 %) evaluation settings for Random Forest Classifier. Features associated with attention, theta, gamma, and alpha bands of EEG wavelet dominance outstand as cross subject consistent targets, jointly accounting for 36 % of accuracy, thus demanding an imperative need for subject dependent personalization strategies within real world deployable bio fatigue detection systems.