Improved drowsiness detection in drivers through optimum pairing of EEG features using an optimal EEG channel comparable to a multichannel EEG system

Abstract

Abstract Multichannel electroencephalography (EEG)-based drowsiness detection (DD) offers higher coverage but comes with increased computational demands, hardware requirements, and user discomfort, whereas single-channel devices are cost-effective and user-friendly but provide lower coverage. We hypothesized that an optimal channel with optimum paired EEG features could achieve coverage comparable to a multichannel system. Subject-specific, EEG-feature-specific thresholding techniques were introduced to classify 927 EEG epochs, derived from visual-based scoring through image processing of fifty drivers’ facial expressions during a 50-min driving simulation, using six individual EEG channels with paired features. Ten normalized EEG features were extracted per epoch using discrete wavelet transform (DWT), and seven thresholding techniques were applied to identify the most consistent method across subjects. Epochs were classified as drowsy or wakeful based on whether their normalized values exceeded or fell below a specific threshold. We then assessed the coverage of each channel by comparing EEG patterns with visual-based scoring. To determine the optimal feature pair for classifying each epoch in alignment with visual-based scoring, 45 feature combinations were evaluated. The pairing of power spectral density (PSD) alpha and PSD theta in channels Frontal4 (F4) and Occipital2 (O2) yielded the highest coverage, achieving 96.1% and 95% with corresponding accuracies of 95.4% and 94.7%, respectively. These results slightly surpassed the coverage achieved using six channels with a single feature, with increases of 1.47% for F4 and 0.32% for O2. Our study demonstrates that an optimal EEG channel with optimum paired EEG features can reduce channels from six to one, lowering computational demands for wearable DD devices. Graphical abstract