Optimizing EEG-based Microsleep Detection: A Comparative Analysis of Time and Frequency Domain Features Utilizing SVM

Microsleeps are short, involuntary episodes of sleep that can have serious safety implications, especially in high-risk occupations. This study investigated optimal features for detecting microsleeps using electroencephalogram (EEG) signals. EEG data was collected from 15 participants using an 8- channel wearable headset. Six feature extraction techniques were evaluated: standard deviation, mean, and variance from the time domain; power spectral density, fast Fourier transform, and spectral centroid from the frequency domain. A Support Vector Machine (SVM) classifier with RBF kernel was utilized to classify the extracted features. Performance was assessed using metrics including accuracy, specificity, precision, recall, and F1-score. Results showed that FFT achieved the highest overall performance, with 93.50% accuracy, 93.55% specificity, 93.88% precision, 93.49% recall, and 93.68% F1- score. Standard deviation was the next best performing feature. These findings suggest that frequency domain features, particularly FFT, showed high performance for effective microsleep detection from EEG signals utilizing SVM. This approach could potentially be applied to develop systems for monitoring and preventing microsleep in safety-critical environments.