Reservoir computing approaches to microsleep detection.

Area of Science:

• Neuroscience
• Computer Science
• Occupational Safety

Background:

• Microsleep detection is crucial for workplace safety, especially in high-risk occupations.
• Existing methods for microsleep detection require improvement.

Purpose of the Study:

• To present a novel microsleep classifier using reservoir computing (RC) methodology, specifically echo state networks (ESNs).
• To enhance benchmark performance in microsleep detection through an individualized, fine-grained ESN architecture.

Main Methods:

• A clustered design featuring a novel ESN-based leaky integrator with up to 8 neurons per cluster was implemented.
• Electroencephalography (EEG) data was analyzed using a cascaded ESN classifier with leaky-integrator neurons.
• 60 principal components from 544 power spectral features were utilized for detection.

Main Results:

• The ESN-based classifier achieved a leave-one-subject-out average detection performance of φ = 0.51 ± 0.07.
• Area Under the Receiver Operating Characteristic Curve (AUC-ROC) was 0.88 ± 0.03.
• Area Under the Precision-Recall Curve (AUC-PR) was 0.44 ± 0.09.

Conclusions:

• This refined method using RC models establishes a new benchmark for EEG-based microsleep detection.
• Despite modest overall performance, the approach shows significant promise for improving workplace safety.