A compact and interpretable convolutional neural network for cross-subject driver drowsiness detection from...

Area of Science:

• Neuroscience
• Transportation Safety
• Machine Learning

Background:

• Driver drowsiness is a major cause of road accidents.
• Electroencephalography (EEG) directly measures brain activity, making it suitable for drowsiness detection.
• Calibration-free EEG-based drowsiness detection is challenging due to individual signal variations.

Purpose of the Study:

• To develop a compact and interpretable Convolutional Neural Network (CNN) for calibration-free, cross-subject driver drowsiness detection using EEG.
• To discover shared EEG features indicative of drowsiness across different individuals.
• To enhance road safety by providing a reliable method for monitoring driver alertness.

Main Methods:

• A Convolutional Neural Network (CNN) model incorporating a Global Average Pooling (GAP) layer was designed.
• Class Activation Map (CAM) was utilized for visualizing and localizing key EEG signal regions.
• The model was evaluated on cross-subject EEG signal classification for drowsiness detection.

Main Results:

• The proposed CNN model achieved an average accuracy of 73.22% in 2-class cross-subject EEG classification.
• Performance surpassed conventional machine learning and other deep learning methods.
• Visualization revealed the model learned biologically relevant features (e.g., Alpha spindles, Theta bursts) for drowsiness and artifacts for alertness.

Conclusions:

• The developed CNN model demonstrates effectiveness in discovering shared EEG features for driver drowsiness detection across subjects.
• The interpretability of the model aids in understanding the neurophysiological basis of drowsiness.
• This approach offers a promising direction for utilizing CNNs in analyzing mental states from EEG signals for improved transportation safety.