EEG-Based Cross-Dataset Driver Drowsiness Recognition With an Entropy Optimization Network.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Driver drowsiness recognition using electroencephalography (EEG) is crucial for developing advanced driver-assistance systems.
• Calibration-free systems are highly desirable but face challenges due to distribution drift across datasets.
• Existing methods struggle to maintain accuracy when applied to new, unseen data distributions.

Purpose of the Study:

• To propose a novel model, the entropy optimization network (EON), for cross-dataset driver drowsiness recognition.
• To address the challenge of distribution drift in EEG-based drowsiness detection.
• To advance the development of calibration-free driver drowsiness recognition systems.

Main Methods:

• A two-step strategy is employed to effectively separate unlabeled target domain data.
• A modified entropy loss function is utilized to cluster source domain-aligned unlabeled samples.
• A self-training framework progressively refines sample separation by leveraging inherent target domain patterns.

Main Results:

• The proposed EON model achieved high 2-class recognition accuracies on domain adaptation tasks using two public datasets.
• The method significantly outperformed existing baseline approaches in cross-dataset driver drowsiness recognition.
• Demonstrated effectiveness in mitigating the impact of distribution drift on recognition performance.

Conclusions:

• The entropy optimization network (EON) presents a promising approach for robust, calibration-free driver drowsiness recognition.
• The proposed method effectively handles distribution shifts in EEG data, enhancing cross-dataset applicability.
• This research illuminates a viable path toward developing universally applicable driver monitoring systems without individual calibration.