Re-purposing EEG Artifacts: Eye Blink Artifact Features as Predictors of Drowsiness Using Deep Learning

Drowsiness has emerged as a significant factor to traffic accidents, workplace accidents and fatalities. In efforts to detect drowsiness in drivers and site workers, researchers have increasingly utilized eye/eyelid images or spectral information from electroencephalograms (EEG). Traditional drowsiness detection systems rely on EEG spectral analysis or camera-based eye tracking, both of which require either clean EEG data or complex hardware setups. In contrast, this study rethinks conventional EEG pre-processing by repurposing eye blink artifacts, typically discarded as noise, as meaningful indicators of drowsiness. This can complement spectral analysis or external cameras/EOG, thereby simplifying the hardware configurations. Eye blink artifact features extracted from two public EEG datasets (nearly 16,000 eye blink events) are smoothed and averaged to enhance the quality of the feature. Seven deep learning models, 1D CNN, MLP, RNN, LSTM, BiLSTM, GRU, and BiGRU, were trained, validated, tested on unseen data, and then optimized using Keras Tuner algorithms to enhance performance. Bi-LSTM achieved the best performance in both datasets, with accuracy exceeding 95%, and precision, recall, AUC-ROC, and F1 score exceeding 96%. These findings suggest that the features of eye blink artifacts offer a viable alternative or can complement conventional EEG spectral analysis to improve drowsiness detection.