EEG-fest: few-shot based attention network for driver's drowsiness estimation with EEG signals

Area of Science:

• Neuroscience and Transportation Safety
• Biomedical Engineering
• Machine Learning in Healthcare

Background:

• Driver inattentiveness is a primary cause of vehicular accidents.
• Electroencephalography (EEG) is a reliable method for assessing driver drowsiness.
• Existing drowsiness detection algorithms face challenges with limited data, anomalous signals, and subject-independent classification.

Purpose of the Study:

• To introduce EEG-Fest, a generalized few-shot learning model for driver drowsiness detection.
• To address limitations of previous algorithms, including small training sets and subject variability.
• To develop a robust system for accurate and reliable drowsiness assessment.

Main Methods:

• Proposed EEG-Fest: a generalized few-shot learning model.
• Few-shot classification for drowsiness level with minimal samples.
• Anomaly detection for identifying aberrant EEG signals.
• Subject-independent classification framework.

Main Results:

• EEG-Fest successfully classifies drowsiness using few support samples.
• The model effectively identifies anomalous EEG signals.
• Achieved subject-independent classification, outperforming conventional EEG algorithms in cross-subject validation.
• Demonstrated superior performance in cross-subject validation compared to two conventional EEG algorithms.

Conclusions:

• EEG-Fest offers a novel and effective solution for driver drowsiness detection.
• The few-shot learning approach overcomes data limitations in EEG analysis.
• The model's ability to perform subject-independent classification enhances its practical applicability for road safety.