Sleep deprivation induces a wide range of neurological and behavioral impairments that significantly undermine cognitive functioning. Sleep deprivation leads to a phenomenon known as ‘localized sleep’, wherein specific brain regions transition into a sleep-like state while the individual remains awake (Vyazovskiy et al., 2011) leading to measurable lapses in cognitive and motor performance (Hung et al., 2013). Moreover, sleep deprivation disrupts the functional connectivity of critical brain networks diminishing communication efficiency across brain regions (Bernardi et al., 2015). Notably, the cognitive effects of sleep deprivation are task-dependent, with executive functions, such as impulse control and visuomotor coordination, exhibiting heightened vulnerability (Bernardi et al., 2015). Moreover, these studies have demonstrated that the increasing occurrence of local neuronal off periods during extended wakefulness correlates with progressive declines in task performance (Hung et al., 2013; Bernardi et al., 2015). Collectively, these findings underscore the essential role of sleep in maintaining cognitive and behavioral stability while highlighting the intricate neural mechanisms through which sleep deprivation compromises brain function.

The real-world implications of these findings are profound, particularly given the high prevalence of acute sleep deprivation in modern society (Gohari et al., 2024). The combination of impaired impulse control, diminished visuomotor coordination and attention lapses significantly increases the likelihood of accidents and human errors (Kayser et al., 2022). These risks are particularly critical in professions requiring sustained attention and precise motor skills, such as healthcare, transportation, and manufacturing. Furthermore, cognitive deficits associated with acute sleep deprivation, including reduced concentration and impaired decision-making (Killgore et al., 2006), negatively affect workplace efficiency and productivity (Brossoit et al., 2019). The link between acute sleep deprivation and diminished impulse control has also been associated with behavioral impulsivity, disinhibition, and increased aggression, contributing to broader implications for mental health and social behavior (Killgore, 2010). Despite the adverse effects of acute sleep deprivation, current detection methodologies primarily rely on self-reported sleep metrics or laboratory-based assessments. While laboratory tests are often resource-intensive and impractical for real-time monitoring (Mitler and Miller, 1996), self-reported metrics are biased or influenced by financial and peer pressure considerations. Therefore, there is a critical need for reliable, objective, and portable tools capable of detecting acute sleep deprivation in real-world settings and mitigating its impact.

Electroencephalography (EEG) has been validated as an effective modality for capturing neural alterations associated with sleep deprivation (Lian et al., 2023). Spectral analysis of EEG signals consistently reveals shifts in brain oscillatory activity that correspond to cognitive and behavioral impairments (Tramonti Fantozzi et al., 2022; Gorgoni et al., 2014; Tassi et al., 2006; Forest and Godbout, 2000). Sleep deprivation is characterized by increased power in low-frequency bands (delta, theta) and reduced power in higher-frequency bands (alpha, beta), particularly in cortical regions implicated in attention and visual processing (Lian et al., 2023; Liu et al., 2025; Hung et al., 2013).

Regional differences in neural activity have been consistently documented. In frontal regions, total sleep deprivation has been shown to lead to a marked loss of functional connectivity in prefrontal cortical areas, characterized by reductions in clustering coefficient in the alpha band and increases in path length in the theta band (Verweij et al., 2014). Additionally, prolonged wakefulness produces a robust increase in frontal low EEG activity (1–7 Hz), reflecting heightened sleep pressure and cortical fatigue (Cajochen et al., 2001). In contrast, in parietal and occipital regions, significant reductions in alpha-band power and increases in delta-band power in the precuneus, inferior parietal lobule, and superior parietal lobule have been reported following 24 h of total sleep deprivation (Lian et al., 2023). Furthermore, theta-band power in centro-parieto-occipital areas increases substantially after prolonged sleep deprivation, with greater elevations correlating with more severe vigilance impairments (Liu et al., 2025). Collectively, these studies highlight the sensitivity of both frontal and posterior cortical regions to sleep loss, demonstrating the effectiveness of EEG in detecting neural alterations relevant to cognitive decline and underscoring its utility for machine learning-based acute sleep deprivation classification.

Machine learning has been widely applied in EEG-based sleep research, with ensemble methods (Mienye and Sun, 2022) and traditional classifiers frequently demonstrating robust performance. Ensemble techniques, such as Random Forest (RF) (Zhao et al., 2022; Monowar et al., 2025), eXtreme Gradient Boosting (XGBoost) (Wang et al., 2024; Monowar et al., 2025), and Light Gradient-Boosting Machine (LightGBM) (Wang et al., 2023; Jain and Ganesan, 2025), have been successfully deployed to model complex, nonlinear patterns in heterogeneous EEG datasets, enhancing generalizability and mitigating overfitting. Support Vector Machines (SVMs) (Kumari et al., 2020) also remain competitive, especially in sleep quality (Wen, 2021) and disorder detection tasks (Falih et al., 2024; Djemal et al., 2025; Bansal et al., 2025; Kim et al., 2025), due to their capacity to accommodate high-dimensional spaces and nonlinear separations.

Deep learning (DL) approaches, including convolutional neural networks (CNNs) (Zhang et al., 2023; Tanci and Hekim, 2025), hybrid CNN–LSTM (Long Short-Term Memory) architectures (Zhuang et al., 2022), and more recently Transformer-based models (Wan et al., 2025), have emerged as powerful tools by learning spatial and temporal representations directly from raw EEG signals. However, studies indicate that, particularly when dataset size is limited, traditional machine learning models, such as SVMs and decision trees, can offer greater robustness and interpretability compared to deep learning approaches, which often require extensive data and computational resources (Rahul et al., 2024).

Despite these advancements, the application of machine learning for the detection of acute sleep deprivation remains underexplored. Existing work in this domain is scarce, with only one notable study to date (Baygin, 2025). Although it achieved high classification accuracy, the study lacked a focus on interpretability, which reduces its relevance for clinical application. Without physiologically grounded explanations, model predictions remain difficult to translate into actionable neurophysiological insights.

The present study aims to address these shortcomings by (i) extracting a comprehensive set of statistical, spectral, functional connectivity, and graph-theoretic features with clear neurophysiological relevance, (ii) implementing a nested stratified cross-validation framework with subject-level separation to derive realistic generalization estimates for unseen individuals. (iii) conducting feature importance analyses to enhance interpretability and lay the groundwork for clinical integration.

Table 2 summarizes the classification performance of the seven evaluated models in terms of mean accuracy, F1-score, and AUC across five test folds, reported as mean ± standard deviation. Results are presented separately for evaluations conducted without and with subject-level separation.

In the evaluation without subject-level separation, CNN achieved the highest mean accuracy (0.9572 ± 0.0134), F1-score (0.9558 ± 0.0148), and AUC (0.9923 ± 0.0039). XGBoost followed closely, with an accuracy of 0.9542 ± 0.0151, F1-score of 0.9536 ± 0.0150, and AUC of 0.9862 ± 0.0081. LightGBM and Random Forest exhibited comparable performance, with accuracies of 0.9483 ± 0.0101 and 0.9453 ± 0.0092, respectively. SVC attained an accuracy of 0.8525 ± 0.0217, F1-score of 0.8480 ± 0.0223, and AUC of 0.9261 ± 0.0148. The Transformer yielded an accuracy of 0.7739 ± 0.0745, F1-score of 0.7944 ± 0.0589, and AUC of 0.8728 ± 0.0508, while LSTM recorded the lowest performance among all models in this setting, with an accuracy of 0.6675 ± 0.0262, F1-score of 0.6546 ± 0.0277, and AUC of 0.7289 ± 0.0281.

With subject-level separation, performance values were lower for all models. RF classifier obtained the highest mean accuracy (0.6823 ± 0.0217), F1-score (0.6702 ± 0.0451) and AUC (0.7290 ± 0.0391). LightGBM and XGBoost yielded similar results, with accuracies of 0.6621 ± 0.0352 and 0.6636 ± 0.0274, respectively. CNN achieved an accuracy of 0.6635 ± 0.0529, F1-score of 0.6415 ± 0.0806, and AUC of 0.7196 ± 0.0477. SVC recorded an accuracy of 0.6508 ± 0.0572, F1-score of 0.6276 ± 0.1068, and AUC of 0.6951 ± 0.0630. The transformer model yielded an accuracy of 0.6335 ± 0.0278, an F1-score of 0.6367 ± 0.0431 and an AUC of 0.6981 ± 0.0529, while LSTM reported the lowest accuracy of 0.6170 ± 0.0320, F1-score of 0.6059 ± 0.0455, and AUC of 0.6662 ± 0.0461. Confusion matrices for the two best-performing models with subject-level separation are presented in Figure 2.

At the epoch level (Figure 3A), Friedman tests revealed significant overall differences in model performance for accuracy (p = 0.0003), F1-score (p = 0.0003), and AUC (p = 0.0002). Post hoc Dunn’s tests with multiplicity-adjusted p-values showed that, for accuracy, the LSTM model performed significantly worse than both XGBoost (rank-sum difference = +25.0, p = 0.0053) and CNN (+24.5, p = 0.0070). For F1-score, LSTM again underperformed relative to LightGBM (+22.0, p = 0.0269), XGBoost (+24.0, p = 0.0093), and CNN (+25.0, p = 0.0053). For AUC, LSTM achieved significantly lower scores than LightGBM (+24.0, p = 0.0093) and CNN (+27.0, p = 0.0016), while CNN also outperformed the Transformer (+21.0, p = 0.0443). No other pairwise comparisons reached statistical significance after correction. Here, the rank-sum difference represents the difference between the sums of the within-fold ranks assigned to each model in the Friedman procedure, with positive values indicating that the first-listed model achieved higher (better) ranks.

With subject-level separation (Figure 3B), Friedman tests showed no overall differences among model performance (accuracy: p = 0.0823, F1-score: p = 0.4137, AUC: p = 0.7076). Consistent with this, Dunn’s multiplicity-adjusted pairwise comparisons found no significant contrasts. A modest trend was noted for accuracy, with RF ranking above LSTM (rank-sum difference = +20.0, adjusted p = 0.0717). All other adjusted p-values were ≥ 0.40 .

SHAP analysis was conducted to identify the most influential features for the two best-performing models under subject-level separation: RF and XGBoost. In the RF model, the highest mean absolute SHAP value was observed for Fp2_theta_meanAM (mean amplitude modulation in the theta band at the right frontal site), followed by Coh_beta (average magnitude-squared coherence between all channel pairs in the beta band), Fp1_theta_meanAM (mean amplitude modulation in the theta band at the left frontal site), and AF3_beta_meanPeakAmplitude (mean peak amplitude within the beta band, indicating the strength of the strongest oscillations) (Figure 4A). The XGBoost model ranked Coh_beta as the most significant feature, followed by Fp2_theta_meanAM, Fp1_beta_meanAM, and T7_beta_peakFrequency (single frequency within the beta band at the left temporal site T7 with the highest total energy over the epoch) (Figure 4B). While both models identified Fp2_theta_meanAM and Coh_beta among their top contributors, the differences in ranking suggest that each algorithm prioritizes distinct but partially overlapping neural markers.

Beeswarm plots (Figures 4C,D) illustrated both the magnitude and directionality of each feature’s influence on the predicted class. Higher Fp2_theta_meanAM values consistently increased the likelihood of a sleep deprivation prediction in both models, indicating that elevated frontal theta modulation is a key discriminator. In contrast, Coh_beta exhibited a more variable influence, with both high and low values contributing in different directions, suggesting interaction effects with other features.

Correlation heatmaps (Figures 4E,F) indicated moderate-to-strong positive correlations among spectral features within neighboring frontal sites. In the RF model, Fp1_theta_meanAM and Fp2_theta_meanAM were strongly correlated. In the XGBoost model, a particularly high correlation was observed between Fpz_theta_meanAM (midline frontal theta modulation) and Fp2_theta_meanAM, suggesting shared frontal theta activity across midline and lateral sites. Across both models, Coh_beta showed relatively low correlations with amplitude or frequency-based features, highlighting its distinct connectivity-related contribution.

In summary, both models relied heavily on frontal theta-band amplitude modulation and beta-band connectivity, underscoring the complementary roles of spectral dynamics and large-scale functional coupling in distinguishing sleep-deprived from well-rested states.

In this study, using an open-source eyes-open resting-state EEG dataset, we investigated the feasibility of detecting acute sleep deprivation with machine learning models. We initially report performance at the epoch level, where data from the same participant may be present in both training and test sets. While this evaluation yielded notably high accuracies, particularly for the CNN, these results are likely inflated due to subject-specific data leakage and temporal dependencies. To address these limitations and assess real-world generalizability, we conducted a more rigorous subject-level evaluation, ensuring that no participant contributed data to both training and test sets. Under this more realistic setting, performance declined across all models, underscoring the difficulty of generalizing to previously unseen individuals. Notably, the RF classifier achieved the highest accuracy in this configuration and demonstrated greater robustness to inter-subject variability compared to the other models. These findings highlight the importance of subject-level validation in the development of EEG-based sleep deprivation detection systems.

The superior performance of RF under subject-level separation may be attributed to the tendency of deep learning models to overfit when trained on relatively small datasets, as in the present study. Deep architectures, while powerful in capturing complex nonlinear patterns, require large volumes of diverse training data to generalize effectively. In contrast, traditional ensemble-based methods such as RF and XGBoost are less susceptible to overfitting in low-data regimes due to their inherent regularization mechanisms, bootstrap aggregation, and feature subspace sampling. These properties enable them to maintain more stable performance when confronted with limited training examples. We hypothesize, however, that with access to substantially larger and more diverse acute sleep deprivation-related EEG datasets, deep learning models may surpass traditional approaches in both accuracy and generalizability.

Feature importance derived from SHAP analysis revealed that the most influential predictors for the best-performing models (with subject-level separation) aligned closely with established neurophysiological effects of acute sleep deprivation (Figure 4). Neurophysiologically, sleep deprivation is associated with increased slow-wave activity, particularly in the theta band (4–7 Hz), reflecting drowsiness and the transition from beta to theta rhythms during the onset of sleep (Iber et al., 2007). Our analysis identified frontal theta-band amplitude modulation, most prominently at Fp2, Fp1, and Fpz, as consistently important across both Random Forest and XGBoost models, in line with evidence that frontal theta activity increases during sustained wakefulness, accompanied by significantly higher subjective sleepiness and decreased alertness (De Gennaro et al., 2007). In addition, beta-band connectivity (Coh_beta) emerged as a key contributor, representing large-scale functional coupling between cortical regions within the 12–30 Hz range. Intermittent increases in beta activity during sleep deprivation have been linked to micro-awakenings or brief periods of heightened cortical activation, potentially reflecting compensatory mechanisms to maintain alertness (Craig et al., 2012). The identification of beta-band peak amplitude and peak frequency features, particularly in frontal and temporal regions such as AF3 and T7, further supports the view that oscillatory dynamics in these bands play complementary roles in distinguishing between well-rested and sleep-deprived states. However, contrary to earlier findings indicating elevated delta-band power as a marker of sleep deprivation (De Gennaro et al., 2007), our models did not identify delta-related features among the top predictors. One possible explanation is methodological: the use of 20-second epochs, while stabilizing spectral estimates, may reduce sensitivity to transient delta fluctuations. In addition, substantial inter-subject variability in delta power likely diminishes its discriminative value for multivariate classification. These biological and methodological considerations may account for why delta features did not emerge as dominant predictors in our study. Collectively, the SHAP results suggest that the discriminative features leveraged by the ensemble models are neurophysiologically meaningful, capturing both the spectral slowing characteristic of drowsiness and the transient beta synchrony associated with compensatory arousal mechanisms.

Focusing on studies that have utilized machine learning for the classification of acute sleep deprivation, a novel feature extraction method called MelPat was introduced in (Baygin, 2025) and applied to the same open-source EEG dataset used in this study, achieving 97% accuracy with an SVC classifier. While this approach demonstrated high performance, our study in contrast, employed a traditional, well-established EEG feature set that achieved comparable accuracy while enabling us to link discriminative features to established neurophysiological biomarkers, thereby enhancing interpretability. Furthermore, unlike MelPat, we also evaluated model performance under a subject-level separation scheme, providing a more rigorous assessment of cross-subject generalizability.

Our work offers several notable contributions. First, unlike prior studies (Baygin, 2025) in this domain, we explicitly evaluated both epoch-level and subject-level separation schemes, with the latter providing a more realistic estimate of model performance in real-world scenarios by preventing within-subject data leakage. Second, we assessed a broad spectrum of approaches, including both traditional machine learning models and deep learning architectures, enabling a comprehensive comparison of their relative strengths. While deep learning models demonstrated strong performance under the less stringent epoch-level setting, our analysis also highlighted their vulnerability to overfitting in limited-data contexts, underscoring the need for larger and more diverse datasets to realize their full potential. Third, our study placed strong emphasis on model interpretability by employing SHAP analysis, which offers a unified, model-agnostic framework for quantifying the contribution of individual features to model predictions. This approach not only facilitates transparency but also enables direct correspondence between the most influential features and established neurophysiological markers of sleep deprivation, thereby enhancing the clinical relevance and applicability of our findings.

This study has certain limitations that should be considered when interpreting the findings. The analysis is based on an open-source eyes-open resting-state EEG dataset obtained from a demographically homogeneous cohort of young, healthy adults, which may restrict the generalizability of the results to more diverse populations. Replication and benchmarking against independent datasets with comparable experimental protocols would enhance the robustness of our conclusions. However, this was not feasible due to the lack of suitable publicly available datasets. The present work conceptualized acute sleep deprivation as a binary state, whereas its neurocognitive and physiological effects likely manifest along a graded continuum. This simplification may obscure more nuanced patterns of impairment and recovery, and represents an inherent limitation of the current design. In the absence of a provided reference channel in the open-source dataset, we employed a common average reference (CAR). However, CAR can be sensitive to montage density and scalp coverage and may influence sensor-level topographies and connectivity (Yao et al., 2019). As we did not formally compare alternative re-referencing schemes, residual reference-related distortions cannot be excluded. While the use of non-overlapping 20-second epochs was chosen to balance temporal resolution with the stability of spectral estimates, this segmentation approach may limit sensitivity to transient or rapidly evolving neural dynamics.

This study represents an important step toward the development of EEG-based machine learning models for the objective detection of acute sleep deprivation and the characterization of its impact on cognitive readiness. While the best-performing model in this study achieved an accuracy of 68%, with subject-level separation, these findings demonstrate the overall viability of EEG-driven approaches for monitoring sleep deprivation in operational contexts. However, real-world deployment will require models with significantly higher performance, trained on larger and more diverse datasets. The present work lays the essential groundwork for such advancements by establishing a robust and interpretable pipeline that captures meaningful neurophysiological signatures of acute sleep loss. It is also important to recognize that the observed effects reflect transient, state-level changes associated with acute sleep deprivation, rather than stable, trait-like characteristics of individuals. Future research should examine how individual variability and longitudinal trajectories influence susceptibility to, and recovery from, acute sleep deprivation, thereby extending the present findings beyond transient state-level effects.

Future research should also consider adopting multi-class or regression-based frameworks to better capture the graded, impact of sleep deprivation on cognitive and physiological function. Expanding model inputs to include demographic variables such as age and sex will also be essential to improve generalizability across populations. Furthermore, exploring advanced EEG features, such as dynamic functional connectivity (dFC) (Chen et al., 2024) and microstate analysis (Khoo et al., 2024), may offer complementary insights that improve classification performance while deepening our understanding of the underlying neural mechanisms. Future work can also explore multi-scale or overlapping windowing strategies to better capture fine-grained temporal dynamics. Finally, integrating more accessible physiological modalities such as photoplethysmography (PPG) may help translate these models into real-world, wearable monitoring systems, ultimately supporting continuous, non-invasive assessment of sleep-related cognitive readiness.