Fair channel ablation and cost-sensitive classification for EEG-based driver fatigue detection: engineering guidelines from real-world driving

Abstract

Wearable EEG fatigue monitoring requires principled channel selection and robust classification under class imbalance. We introduce a fair per-channel-only ablation protocol for binary fatigue detection—alert (KSS ≤ 6) vs. fatigue (KSS ≥ 7)— applied to 64-channel EEG from 99 real-driving participants (534 trials). Each channel configuration uses only per-channel spectral features (15 per channel), with SelectKBest k scaling linearly with channel count, isolating regional information content from feature-engineering confounds. A cost-sensitive Random Forest with SMOTE oversampling addresses class imbalance. Three findings emerge: (1) a 7-channel prefrontal set (PF-7: Fp1, Fp2, AF3, AF4, AF7, AF8, Fz) achieves 75.0% balanced accuracy—outperforming central channels and retaining 99% of frontal-region performance; (2) theta relative power is the most robust spectral biomarker (76% of channels significant, peak H = 55.4 at Fz), with Spearman correlations confirming alpha increase (ρ = +0.13) and theta decrease (ρ = −0.12) with fatigue; and (3) the cost-sensitive pipeline achieves balanced accuracy of 75.2% with macro-F1 = 0.75 (alert F1 = 0.72, fatigue F1 = 0.78), demonstrating clinically meaningful fatigue detection from real-world driving EEG with minimal preprocessing.