Enhancing Driver Drowsiness Detection: A Compact and Interpretable LSTM Model with Attention Mechanism for Cross-Subject EEG Analysis

Abstract

Driver drowsiness is a significant contributor to road accidents, leading to numerous fatalities and hazards globally. Electroencephalography (EEG) has become a reliable physiological signal for identifying drowsiness because it directly monitors brain activity. However, creating a precise and calibration-free method for identifying driver drowsiness is still difficult because EEG is affected by individual variations and changes in subjects. In study proposed concise and easy-to-understand Long Short-Term Memory (LSTM) model equipped with an attention mechanism to efficiently capture common EEG characteristics among various individuals for detecting driver drowsiness. The attention mechanism improved the model's interpretability by pinpointing crucial time steps in EEG signals that have the greatest impact on classification. By conducting leave-one-subjectout (LOSO) cross-validation on 11 subjects, the model obtained an average accuracy of 79.1%, showcasing better results when compared to conventional techniques. Power Spectral Density (PSD) analysis also showed that the model successfully captured biologically significant characteristics, such as variations in Delta and Theta waves, that are associated with drowsiness. The proposed LSTM model with attention mechanism shows potential for practical use in improving road safety through EEG-based drowsiness detection. Future research will concentrate on enlarging the dataset, integrating multi-channel EEG data, and investigating real-time application for practical utilization in the transportation sector.