Abstract
Driver drowsiness is an important contributor to road accidents, but conventional multichannel electroencephalography (EEG) systems are difficult to deploy in everyday driving contexts. This study investigates whether band-power features obtained from a low-cost, single-channel commercial EEG device (NeuroSky MindWave Mobile 2, Fp1 position) can discriminate between the alert and drowsy labels provided in a public dataset. Frequency-band features were combined with physiologically motivated ratios, and several machine learning (ML) classifiers were evaluated using repeated nested stratified cross-validation. Because the dataset does not provide participant identifiers, the analysis estimates within-dataset discrimination and cannot establish generalization to unseen individuals. Random Forest achieved the highest mean area under the receiver operating characteristic curve (AUC = 0.869), whereas the Soft-Voting Ensemble provided a compromise between accuracy (0.78), drowsiness recall (0.77), and AUC (0.863). SHapley Additive exPlanations (SHAP) analysis identified delta, highBeta, and highGamma as the most influential variables. A reduced model using these three features retained an AUC of 0.831, and the ensemble required approximately 7.52 ms per prediction on the evaluated computing platform. These findings support the computational feasibility of explainable drowsiness classification from consumer-grade single-channel EEG features, while subject-independent and embedded-hardware validation remain necessary.