An EEG Channel Selection Framework for Driver Drowsiness Detection via Interpretability Guidance.

Area of Science:

• Neuroscience
• Transportation Safety
• Machine Learning

Background:

• Drowsy driving poses significant risks to road safety, necessitating effective driver monitoring systems.
• Electroencephalogram (EEG) signals offer a reliable measure of mental fatigue for drowsiness detection.
• Existing methods often overlook EEG data noise and redundancy, limiting drowsiness detection performance.

Purpose of the Study:

• To introduce an Interpretability-guided Channel Selection (ICS) framework for enhanced driver drowsiness detection.
• To address the limitations of using raw or full-head EEG data in current drowsiness monitoring models.
• To improve the accuracy and applicability of drowsiness detection models, particularly in cross-subject scenarios.

Main Methods:

• A two-stage training strategy employing interpretability guidance for channel selection.
• Training a teacher network on full-head EEG data, followed by Class Activation Mapping (CAM) for channel importance analysis.
• A channel voting scheme to select top contributing EEG channels, followed by training a student network on selected channels.

Main Results:

• The proposed ICS framework effectively identifies key EEG channels for drowsiness detection.
• The method significantly improves the performance of driver drowsiness detection models.
• Experimental results on a public dataset validate the framework's applicability and effectiveness, especially for cross-subject detection.

Conclusions:

• The ICS framework offers a novel and effective approach to optimize EEG channel usage for driver drowsiness detection.
• By focusing on interpretable, high-contributing channels, the method enhances model performance and robustness.
• This research contributes to safer driving through advanced, data-driven fatigue monitoring techniques.