Optimized echo state networks with leaky integrator neurons for EEG-based microsleep detection.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Microsleep detection is crucial for safety-critical applications.
• Electroencephalography (EEG) provides valuable spectral features for sleep state analysis.
• Previous single classifier models showed limited performance in detecting microsleep.

Purpose of the Study:

• To develop an improved microsleep detection system.
• To evaluate the performance of ensemble learning compared to single classifiers.
• To leverage spectral features from EEG for accurate microsleep identification.

Main Methods:

• Utilized 16-channel EEG data sampled at 256 Hz.
• Extracted spectral features for microsleep state detection.
• Implemented and compared single classifier models: Echo State Network (ESN) with leaky integrator neurons, ESN with sigmoidal inputs, and Linear Discriminant Analysis (LDA).
• Applied ensemble learning via stacked generalization to combine multiple classifier outputs.

Main Results:

• The best single classifier (ESN with leaky integrator neurons) achieved a phi correlation (φ) of 0.38 and 67.3% accuracy.
• Linear Discriminant Analysis (LDA) achieved φ of 0.31 and 53.6% accuracy.
• Ensemble learning using stacked generalization significantly improved performance, reaching φ of 0.51 and 81.2% accuracy.
• This represents a substantial improvement over previous results.

Conclusions:

• Ensemble learning, specifically stacked generalization of ESN models, offers superior performance for microsleep detection.
• The developed system demonstrates a significant advancement in accurately identifying microsleep states from EEG.
• This approach holds promise for enhancing safety systems reliant on real-time sleep state monitoring.