An EEG-based machine learning framework for diagnosing acute sleep deprivation.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Acute sleep deprivation impairs cognitive function and increases accident risk.
• Objective diagnosis of sleep deprivation is crucial for timely intervention.
• Electroencephalography (EEG) offers a potential non-invasive method for sleep status assessment.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for discriminating between sleep-deprived and well-rested individuals using EEG data.
• To assess the performance of various ML and deep learning models in detecting acute sleep deprivation.
• To investigate the impact of subject-level separation on model generalization.

Main Methods:

• Analysis of 61-channel resting-state EEG data from 71 participants.
• Feature extraction including statistical, spectral, connectivity, and graph-theoretic metrics.
• Training and evaluation of ML classifiers (LightGBM, XGBoost, RF, SVC) and deep learning models (CNN, LSTM, Transformer) using nested cross-validation.
• Model performance assessed with and without subject-level separation.

Main Results:

• Without subject separation, Convolutional Neural Network (CNN) achieved the highest accuracy (95.72%).
• With subject-level separation, Random Forest (RF) demonstrated the highest accuracy (68.23%), indicating challenges in cross-subject generalization.
• Deep learning models (Transformer, LSTM) generally underperformed compared to ML classifiers.

Conclusions:

• EEG-based ML shows promise for objective sleep deprivation detection.
• Robust subject-level generalization is a key challenge for these diagnostic tools.
• Findings support the development of scalable, non-invasive EEG tools for sleep deprivation monitoring.