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

Acute sleep deprivation significantly impacts cognitive function, contributes to accidents, and increases the risk of chronic illnesses, underscoring the need for reliable and objective diagnosis. Our work aims to develop a machine learning-based approach to discriminate between EEG recordings from acutely sleep-deprived individuals and those that are well-rested, facilitating the objective detection of acute sleep deprivation and enabling timely intervention to mitigate its adverse effects.

Sixty-one-channel eyes-open resting-state electroencephalography (EEG) data from a publicly available dataset of 71 participants were analyzed. Following preprocessing, EEG recordings were segmented into contiguous, non-overlapping 20-second epochs. For each epoch, a comprehensive set of features was extracted, including statistical descriptors, spectral measures, functional connectivity indices, and graph-theoretic metrics. Four machine learning classifiers - Light Gradient-Boosting Machine (LightGBM), eXtreme Gradient Boosting (XGBoost), Random Forest (RF), and Support Vector Classifier (SVC) - were trained on these features using nested stratified cross-validation to ensure unbiased performance evaluation. In parallel, three deep learning models-a Convolutional Neural Network (CNN), Long Short-Term Memory network (LSTM), and Transformer-were trained directly on the raw multi-channel EEG time-series data. All models were evaluated under two conditions: (i) without subject-level separation, allowing the same participant to contribute to both training and test sets, and (ii) with subject-level separation, where models were tested exclusively on unseen participants. Model performance was assessed using accuracy, F1-score, and area under the receiver operating characteristic curve (AUC).

Without subject-level separation, CNN achieved the highest accuracy (95.72%), followed by XGBoost (95.42%), LightGBM (94.83%), RF (94.53%), and SVC (85.25%), with the Transformer (77.39%) and LSTM (66.75%) models achieving lower accuracies. Under subject-level separation, RF achieved the highest accuracy (68.23%), followed by XGBoost (66.36%), LightGBM (66.21%), CNN (65.35%), and SVC (65.08%), while the Transformer (63.35%) and LSTM (61.70%) models achieved the lowest accuracies.

This study demonstrates the potential of EEG-based machine learning for detecting acute sleep deprivation, while underscoring the challenges of achieving robust subject-level generalization. Despite reduced accuracy under cross-subject evaluation, these findings support the feasibility of developing scalable, non-invasive tools for sleep deprivation detection using EEG and advanced ML techniques.