Predicting Microsleep States Using EEG Inter-Channel Relationships.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Microsleeps are brief, involuntary losses of consciousness.
• Accurate detection of microsleeps is crucial for safety-critical operations.
• Electroencephalography (EEG) is a common tool for monitoring brain activity.

Purpose of the Study:

• To evaluate the effectiveness of various inter-channel EEG relationships for predicting microsleep states.
• To compare the performance of Linear Discriminant Analysis (LDA) and Linear Support Vector Machine (SVM) classifiers.
• To identify the most informative features for microsleep prediction.

Main Methods:

• Seven pairwise inter-channel EEG relationships were analyzed: covariance, Pearson's correlation, wavelet cross-spectral power, wavelet coherence, joint entropy, mutual information, and phase synchronization index.
• Features were extracted from 5-second EEG windows and processed using LDA and SVM classifiers with incorporated priors for class imbalance.
• The Sequential Forward Selection (SFS) method was employed for feature selection.
• Leave-one-subject-out cross-validation was used to evaluate prediction performance on 16-channel EEG data.

Main Results:

• Non-normalized EEG features outperformed normalized features.
• Demeaning the non-normalized features significantly improved prediction performance.
• The LDA classifier using joint entropy features achieved the best average prediction performance (phi=0.47, AUC_PR=0.50, AUC_ROC=0.95).
• Joint entropy between the O1 and O2 electrodes in the theta frequency band was identified as the most informative feature.

Conclusions:

• Continuous prediction of microsleep states from EEG is feasible using machine learning classifiers.
• Joint entropy analysis of EEG signals, particularly in the theta band between O1 and O2, is a highly effective method for detecting microsleeps.
• Feature engineering, including demeaning non-normalized data, plays a critical role in enhancing prediction accuracy.