Comprehensive Evaluation of EEG Features for Microsleep Detection: Statistical Validation and Methodological Improvement

Microsleeps, short lapses in attention, are critical to detect in safety-sensitive settings such as driving and healthcare. Electroencephalography (EEG) provides a basis for detection, but the effectiveness of features and models remains uncertain. This study conducts a comprehensive evaluation of multiple EEG feature categories, including time-domain, frequency-domain, bispectrum, nonlinear, time-frequency, and intensity-weighted descriptors. These features are systematically assessed using both classical machine learning methods and neural models (CNN, LSTM) under a unified framework with consistent dataset, preprocessing, and metrics, ensuring fair comparison. Results show that intensity-weighted and nonlinear features achieve balanced trade-offs between sensitivity and specificity, while time-frequency features provide the highest accuracy but lower sensitivity. Tree-based models (Random Forest, Gradient Boost) and neural networks outperform simpler baselines. Statistical validation with the Friedman test confirmed significant differences across classifiers (Accuracy: p = 0.0011, Sensitivity: p = 0.0117, Specificity: p = 0.0017), and Nemenyi post-hoc analysis verified the superiority of tree-based and neural approaches.This work contributes by establishing reliable benchmarks for EEG-based microsleep detection, highlighting which features and models are most effective, and providing statistically validated evidence to guide future research.