Detection of Cortical Arousals in Sleep Using Multimodal Wearable Sensors and Machine Learning

Detection of Cortical Arousals in Sleep Using Multimodal Wearable Sensors and Machine Learning | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Detection of Cortical Arousals in Sleep Using Multimodal Wearable Sensors and Machine Learning Murat Kucukosmanoglu, Sarah Conklin, Kanika Bansal, Sena Kaya, and 6 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6574148/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 18 Dec, 2025 Read the published version in Scientific Reports → Version 1 posted 12 You are reading this latest preprint version Abstract Cortical arousals are brief brain activations that disrupt sleep continuity and contribute to cardiovascular, cognitive, and behavioral impairments. Although polysomnography is the gold standard for arousal detection, its cost and complexity limit use in long-term or home-based monitoring. This study presents a noninvasive machine learning based framework for detecting cortical arousals using the RestEaze™ system, a leg-worn wearable that records multimodal physiological signals including accelerometry, gyroscope, photoplethysmography (PPG), and temperature. Across multiple methods tested, including logistic regression, XGBoost, and Random Forest classifiers, we found that features related to movement intensity were the most effective in identifying cortical arousals, while heart rate variability had a comparatively lower impact. The framework was evaluated in 14 children with attention-deficit/hyperactivity disorder (ADHD) who were being assessed for possible restless leg syndrome related sleep disruption. The Random Forest model achieved the best performance, with a ROC AUC of 0.94. For the arousal class specifically, it reached a precision of 0.57, recall of 0.78, and F1-score of 0.65. These findings support the feasibility of wearable-based machine learning for real-world arousal detection, demonstrated here in a pediatric ADHD cohort with sleep-related behavioral concerns. Biological sciences/Neuroscience Biological sciences/Physiology Health sciences/Biomarkers Health sciences/Medical research cortical arousals RestEaze ADHD wearables machine learning sleep monitoring Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Introduction Cortical arousals are brief interruptions in electroencephalographic (EEG) activity that fragment sleep without full awakening. Although transient, these arousals contribute to autonomic activation and disrupted sleep pattern, with growing evidence linking them to hypertension, cognitive decline, and elevated cardiovascular risk 1 – 3 . Total sleep duration less than 5 hours per night is considered high-risk for cardiovascular morbidity and mortality 4 . Disrupted or insufficient sleep has also been associated with systemic inflammation, metabolic dysfunction, and increased all-cause mortality 5 . Elevated rates of sleep disturbances, including cortical and autonomic arousals, have also been observed in children with attention-deficit/hyperactivity disorder (ADHD) 6 – 8 . Early and accurate detection of these arousals may offer clinical insights into the relationship between poor sleep quality and daytime behavioral symptoms that may reveal patterns that differ by clinical subtype. Polysomnography remains the gold standard for detecting cortical arousals 9 , 10 , yet its high cost, complexity, and requirement for overnight clinical supervision limit its use for large-scale or long-term monitoring 11 . Consumer sleep technologies, such as sleep trackers, offer a non-invasive, scalable approach to sleep monitoring, with the potential to support early identification of sleep fragmentation in home environments. While these devices offer greater accessibility, they often suffer from poor agreement with polysomnography, particularly in detecting brief or motionless arousals 12 . A multicenter validation study involving 11 wearable, nearable, and airable consumer sleep trackers confirmed substantial variation in performance across devices, with some showing macro F1-scores as low as 0.26 when compared to Polysomnography 13 . However, the growing integration of wearable sleep technologies into daily life offers a valuable opportunity to develop advanced frameworks that can effectively use these technologies to detect clinically relevant features of sleep. One promising solution involves tracking leg movements during sleep, which frequently occur alongside cortical arousals, especially in populations with conditions like restless leg syndrome, periodic limb movement disorder, or ADHD 14 – 17 . Recent studies using wearable leg sensors have shown that leg movements during sleep features can effectively distinguish arousals, and that leg-EEG signal coupling may reflect deeper physiological mechanisms of sleep disruption 18 , 19 . In this study, we evaluate multimodal sensor data from a leg-worn wearable, RestEaze™, to detect cortical arousals using interpretable machine learning models, with the aim of advancing practical and reliable sleep health monitoring solutions outside of traditional clinical settings. The RestEaze™ system integrates accelerometry, gyroscope, photoplethysmography (PPG), and temperature sensors, offering a comprehensive view of movement and physiological dynamics during sleep. In a prior pilot study using a similar platform, we introduced neuro-extremity analysis, a novel approach that employed Granger causal modeling to assess the temporal and directional relationships between cortical arousals and leg movements 20 . That study revealed that textile-based capacitive sensors showed stronger temporal and spectral coupling with EEG-theta oscillations than inertial sensors, and more accurately identified expert-labeled cortical arousals. These findings support the hypothesis that leg movements and cortical arousals are driven by coordinated activity within a shared central arousal system. The current study builds upon this work by incorporating PPG and temperature sensors into the previously studied system and focusing exclusively on inertial sensors for movement detection, as they were found to reliably capture arousal-related leg movements while avoiding the redundancy and implementation challenges associated with textile-based capacitive sensors. This setup allows extraction of heart rate (HR) and heart rate variability (HRV) features that may offer additional insight into autonomic activation during sleep 21 – 23 . Results Sleep is composed of two main states: rapid eye movement (REM) sleep and non-rapid eye movement (NREM) sleep. NREM includes three stages: N1, N2, and N3, which progress from light to deep sleep. These stages repeat in cycles throughout the night 24 . We began by examining the distribution of cortical arousals across sleep stages to establish a physiological context for the classification task. Arousals occurred most frequently during N2 sleep, with a mean proportion of 56.77% (95% confidence interval [CI]: 46.14–67.40%), followed by N1 at 17.47% (95% CI: 8.15–26.79%), REM at 13.17% (95% CI: 4.43–21.90%), and N3 at 12.60% (95% CI: 7.17–18.02%), averaged across subjects. This distribution aligns with established sleep physiology: N2 sleep not only comprises a larger portion of total sleep time but also has a lower arousal threshold, making it more prone to cortical arousals due to its transitional nature between wakefulness and deeper sleep stages 24 . Similarly, the elevated rate of arousals during N1 reflects its light sleep status and proximity to wakefulness. Interestingly, we also observed notable levels of arousals during N3 and REM sleep, suggesting increased cortical arousal beyond the lighter stages. This pattern may support prior findings showing that adolescents with ADHD and learning disorders exhibit increased cortical arousal during N2 and N3 sleep, particularly in central and frontal brain regions 25 . To enable real-time detection of these arousal events using wearable data, we implemented and evaluated machine learning models designed to classify arousals from multimodal physiological signals. We evaluated the performance of three machine learning classifiers: Logistic Regression, XGBoost, and Random Forest for detecting cortical arousals based on multimodal physiological data from a leg-worn wearable device on full cohort of 14 children with ADHD, a population known to experience elevated levels of sleep fragmentation and frequent cortical arousals 6 . We chose these models to represent different levels of complexity and explainability: Logistic Regression as a simple linear baseline, Random Forest as a robust ensemble method, and XGBoost as a state-of-the-art gradient boosting algorithm. All models were trained using a leave-one-subject-out cross-validation (LOOCV) approach to ensure robust subject-independent evaluation. The classification task involved identifying arousal events versus non-arousal periods. Evaluation metrics included class-wise precision, recall, F1-score, and overall Receiver Operating Characteristic - Area Under the Curve (ROC-AUC). Model training and performance The performance of each model is summarized in Table 1 . While all three classifiers showed high accuracy in detecting non-arousal periods (Class 0), their ability to detect arousal events (Class 1) varied considerably. Logistic Regression achieved a Class 1 F1-score of 0.57 and a ROC-AUC of 0.90. XGBoost improved precision but had lower recall for Class 1, with a resulting F1-score of 0.61 and a ROC-AUC of 0.93. Random Forest achieved the best balance, with a Class 1 F1-score of 0.65 and the highest ROC-AUC of 0.94. Based on these results, the Random Forest model was selected for further analysis. Table 1 summarizes the performance of each model. Table 1 Model Performance Summary Model Class Precision Recall F1-Score ROC-AUC Logistic Regression 0 0.99 0.94 0.96 0.90 1 0.45 0.84 0.57 XGBoost 0 0.99 0.95 0.97 0.93 1 0.50 0.82 0.61 Random Forest 0 0.99 0.96 0.98 0.94 1 0.57 0.78 0.65 Feature Importance Figure 1 presents the ranked list of the most important features contributing to cortical arousal classification, as determined by the Random Forest model. These features were predominantly derived from accelerometer and gyroscope signals, with a smaller contribution from HR and HRV metrics. The most important features included statistical, energy-based, and entropy-related measures. Importantly, standard deviation, root mean square (RMS), maximum, and range from the x-axis of the accelerometer appeared prominently in the ranking. This suggests that lateral leg movement (x-direction) plays a critical role in arousal episodes, consistent with biomechanical patterns observed during limb movement–related arousals. Entropy-based features such as spectral entropy from both accelerometer and gyroscope signals were also among the top-ranked predictors. These features reflect the signal complexity or irregularity during sleep and are useful for capturing subtle variations in movement associated with arousals. Similarly, RMS AUC (Root Mean Square Area Under the Curve) quantifies cumulative signal energy, which is often elevated during microarousals due to brief bursts of leg activity. Other contributing features included HRV-derived indices such as HRV Higuchi fractal dimension (HRV-HFD), HRV Cardiac Sympathetic Index (HRV-CSI), and HRV Fuzzy Entropy (HRV-FuzzyEn), all of which reflect beat-to-beat HRV complexity, physiological markers known to fluctuate during autonomic arousals 26 . However, they were less important than movement-based metrics, suggesting a stronger motor component to arousals in children with ADHD. Similarly, temperature-based features were not among the top-ranked predictors, indicating minimal relevance to arousal classification in this context. In addition to feature rankings, we analyzed PPG signal quality across arousal categories. The mean PPG quality score was 0.818 (95% CI: 0.738–0.899) during non-arousal periods and 0.488 (95% CI: 0.420–0.556) during arousal events. This significant decline in signal quality during arousals suggests increased motion artifacts or sensor dropout, which may explain the lower importance of PPG-derived features in the final model. Agreement with Ground Truth Figure 2 shows the model prediction of the arousal rates against the true arousal rates (ground truth). In this study, arousal rate refers to the number of 60-second windows that contain at least one cortical arousal event, normalized per hour of total sleep time. The predicted rates exhibited a strong correlation with the ground truth, yielding a Spearman’s rank correlation coefficient ρ = 0.89 (p = 2.00 × 10⁻⁵) and a Kendall’s τ = 0.76 (p = 3.95 × 10⁻⁵). These results show a strong relationship, suggesting that the model successfully preserves subject-wise ranking in arousal frequency, which is crucial for estimating severity and comparing individuals. The fitted linear regression line further supports the alignment between predicted and true values. The slope below 1.0 indicates underestimation at higher arousal rates, yet the close clustering of points around the line reflects consistency in the overall prediction trend. The regression slope was statistically significant ( p < 0.01), with a 95% CI of [0.383, 1.050]. To further assess agreement, a Bland–Altman analysis was conducted (Fig. 3 ). This plot shows the differences between predicted and true arousal rates as a function of their average, both expressed in arousals per hour. The mean difference was + 0.88 arousals/hour (Predicted − True), indicating a slight overall tendency of the model to overestimate arousal frequency. The 95% limits of agreement ranged from − 1.40 to + 3.17 arousals/hour. Temporal Prediction Patterns To evaluate model behavior across time, we visualized prediction sequences for three subjects who showed distinct arousal patterns. Figure 4 shows minute-by-minute comparisons between predicted and true arousals across the sleep duration. For Subject A (Fig. 4 a), who exhibited frequent and widely distributed arousals, the model effectively captured both isolated and clustered events throughout the night. Minute-by-minute inspection showed that most predictions were temporally aligned with ground truth, with several pre-arousal predictions appearing within one to two minutes of labeled events. In contrast, Subject B (Fig. 4 b) presented arousals that occurred in distinct temporal clusters during the early and late portions of the recording. The model maintained high temporal precision, correctly identifying contiguous arousal periods while avoiding false positives during quiescent intervals. Subject C (Fig. 4 c) exhibited a sparser distribution of arousals. The model's predictions closely matched the few true events, with overclassification toward the end. The agreement between predicted and true arousals is quantified using Arousals (Class 1) F1-scores: 0.62 (a), 0.68 (b), and 0.54 (c). These scores indicate strong model performance given the substantial class imbalance, where arousals make up only ~ 6% of the data. For context, random guessing would yield an F1-score near 0.06, making the observed values highly meaningful. These subject-level, minute-by-minute visualizations highlight the model’s adaptability to inter-individual variability in sleep and arousal patterns. Discussion This study demonstrates the feasibility of using multimodal wearable sensors and machine learning to detect cortical arousals during sleep, offering an accessible alternative to traditional in-clinic polysomnography. Among the classifiers tested, the Random Forest model achieved the best balance between recall and precision, yielding the highest ROC-AUC of 0.94. This result is consistent with Random Forest’s ability to handle complex patterns, feature interactions, and imbalanced data. Its ensemble-based architecture and embedded feature selection likely contributed to its robustness in this complex real-world dataset. Compared to Logistic Regression, which assumes linearity, and XGBoost, which can be sensitive to hyperparameter tuning in small datasets, the Random Forest model proved particularly effective at capturing subtle, subject-specific arousal signatures. Feature importance analysis further revealed that the most predictive signals were derived from accelerometry and gyroscope data, particularly features reflecting signal variability and complexity, such as root mean square amplitude, standard deviation, and spectral entropy. These findings are consistent with prior work suggesting that leg movements are linked with cortical arousals 14 , 16 , 17 . Entropy measures likely captured the fragmented nature of movement during arousals. In contrast, HR and HRV features extracted from PPG contributed less prominently to model performance. This was not entirely unexpected, as the original sampling rate of 25 Hz may be insufficient for accurate HRV estimation. Prior work has shown that HRV metrics like Standard Deviation of NN Intervals (SDNN) and Root Mean Square of Successive Differences (RMSSD) require significantly higher sampling rates to ensure reliability, at least 50 Hz for SDNN and 100 Hz or more for RMSSD without interpolation 27 . Additionally, signal quality issues further limited the reliability of PPG-derived features. These noises, primarily motion artifacts and high-frequency noise, are inevitable in wearable-based health and well-being monitoring systems and can significantly impact peak detection accuracy 28 . In our study, the average PPG signal quality declined from 0.818 during non-arousal periods to 0.488 during arousal. This indicates a consistent drop in signal quality during arousal events. Interestingly, the model predicted more arousals than were annotated by experts, particularly in subjects with sparse arousal profiles (Subject C). Rather than representing pure false positives, these predictions may reflect physiological events, such as sub-threshold arousals or autonomic activations, that were not captured by EEG-based criteria. This raises the possibility that wearable sensors may detect some physiological markers of sleep disruption that fall outside the boundaries of current clinical scoring systems. Indeed, prior research has shown that physiological changes surrounding arousal events can be significant, often extending beyond the boundaries of EEG-defined arousals 29 , 30 . These findings highlight how machine learning and wearables can improve sleep assessment beyond conventional methods. Additionally, the use of 60-second windows may have contributed to some discrepancy by grouping multiple arousals into a single event or capturing signal fluctuations surrounding true arousals. Lastly, our subject-independent and interpretable framework provides minute-level temporal precision, making it suitable for clinical applications that require generalizable detection. It shows promise for individuals with ADHD, a group often underserved by traditional sleep diagnostics. Pediatric restless legs syndrome, for example, can cause significant sleep disruption, behavioral issues, and impaired daytime functioning that mimic ADHD symptoms 31 , 32 . While ADHD's recognized subtypes (inattentive, hyperactive-impulsive, and combined) are well-described, their association with distinct sleep profiles remains unclear, highlighting the need for detailed pediatric sleep assessment 33 . Refined at-home monitoring could help identify specific sleep disorders and support more personalized, subtype-targeted treatments for pediatric ADHD. Building on these findings, this work presents multiple opportunities for future development. Priorities include expanding to larger and more diverse datasets, using deep learning to model long-range patterns, and incorporating continuous arousal scoring to reflect subtle physiological changes. Real-world feedback such as sleep staging, user experiences, and device usability will be vital for transforming this research into a practical home-based health solution. Ultimately, these efforts aim to bring clinical-quality sleep analytics into everyday environments through smart and accessible wearables. Conclusion This study presents a non-invasive, wearable-based framework for detecting cortical arousals using multimodal physiological signals from a leg-worn device. Among the classifiers evaluated, the Random Forest model performed best, achieving a ROC-AUC of 0.94 and showing strong alignment with expert-labeled EEG annotations. Key predictive features, such as leg movement variability and signal entropy, support the role of movement-related physiological signals as markers of central arousals. These findings demonstrate the potential of systems like RestEaze™ for clinically meaningful, at-home sleep monitoring. Future work should include larger, more diverse populations and explore continuous arousal scoring to enhance clinical relevance. Methods Participants and Data Acquisition Physiological and movement data were collected from 14 children diagnosed with ADHD using the RestEaze™ Movement Analyzer, a wireless, leg-worn wearable designed for non-intrusive sleep monitoring and arousal detection. More details about the RestEaze™ can be found in previous publication 18 . As illustrated in Fig. 5 , the RestEaze™ device integrates multiple synchronized sensors: A 3-D accelerometer and 3-D gyroscope embedded within an inertial measurement unit (IMU) for leg movement and orientation tracking, A PPG sensor for capturing cardiovascular dynamics, and Object and ambient temperature sensors for thermal signature during sleep. The accelerometer (X, Y, Z axes), gyroscope (X, Y, Z axes), and PPG channels (IR, red, green LEDs) were all sampled at 25 Hz, providing high-resolution capture of biomechanical and cardiovascular signals. Temperature data was sampled at 0.2 Hz, appropriate for monitoring slow-changing thermal conditions. This setup enables continuous, multimodal recording throughout the night, capturing both fine-grained leg movements and physiological fluctuations associated with cortical arousals. Across the 14 participants, the average total sleep time was approximately 7.59 hours per subject, totaling 106.32 hours of recorded sleep data. Data collection was conducted during natural sleep in a home or clinical setting. All study procedures were approved by the Institutional Review Board of Johns Hopkins University. Research was conducted in accordance with the Declaration of Helsinki and all relevant ethical guidelines and regulations, including obtaining informed consent from all participants and/or their legal guardians. Cortical Arousals Rate Cortical arousals (ground truth) were identified and scored according to the guidelines set by the American Academy of Sleep Medicine (AASM) 34 , which define arousals as abrupt shifts in EEG frequency, including alpha, theta, or activity exceeding 16 Hz, that last for at least 3 seconds and occur after a minimum of 10 seconds of uninterrupted sleep 21 . Arousal rate was calculated as the number of 60-second windows labeled with at least one cortical arousal event, normalized per hour of total sleep time. Specifically, if any arousal occurred within a given 60-second segment, the entire window was labeled as an arousal window (Class 1). The resulting arousal rate, expressed in arousal windows per hour, provides a temporally consistent metric for comparing arousal frequency across individuals. In addition to cortical arousals, sleep stages, and limb movements were scored manually by trained technicians according to the AASM guidelines 34 . Bilateral limb movement events were also manually annotated, whereas leg movement channels were scored using an automated algorithm via the Sleepware G3 platform (Philips Respironics, US). Final scoring was reviewed and confirmed by a board-certified sleep physician and AASM fellow. Preprocessing and Feature Generation All raw sensor signals were processed using a unified preprocessing pipeline (see Fig. 5 ), which included filtering, segmentation into 60-second non-overlapping windows, and modality-specific feature extraction. The choice of a 60-second window was guided by the need to balance temporal resolution with physiological interpretability. Each one-minute segment contains sufficient cardiac cycles (typically 60–100 beats) to allow reliable estimation of HR and HRV, while also being short enough to detect changes in physiological state over time. For the PPG signal, the preprocessing began with upsampling to 200 Hz using linear interpolation. This step was essential for achieving the temporal resolution required for accurate peak detection and compatibility with feature extraction functions that assume higher sampling rates. Several methods did not perform at the native 25 Hz resolution, especially those involving frequency-domain HRV metrics. The upsampled signal was then bandpass filtered between 0.2 and 5 Hz using a Butterworth filter to remove baseline drift and suppress motion artifacts. The filter was implemented in Python 3.11 using the butter and filtfilt functions from the scipy.signal module, which apply zero-phase forward and reverse filtering to avoid phase distortion 35 . Following filtering, we evaluated several peak detection strategies to identify heartbeats from the PPG waveform. Among these, the ppg-findpeaks function from the NeuroKit2 library 36 provided reliable results in terms of peak timing consistency and robustness to signal noise. Figure 6 shows the effects of preprocessing: the top panel displays the raw PPG signal with notable baseline fluctuations (Fig. 6 a), the middle panel shows the filtered waveform with clearly resolved peaks (Fig. 6 b), and the bottom panel plots the computed PPG signal quality over time (Fig. 6 c). This quality metric, ranging from 0 to 1, reflects the reliability of the signal for physiological analysis. Once peaks were detected, HR and HRV features were extracted from each 60-second window. HR metrics included minimum, maximum, and mean HR. HRV features encompassed time-domain measures (e.g., RMSSD, SDNN), frequency-domain indices (e.g., low-frequency/high-frequency ratio), and nonlinear metrics such as entropy, coefficient of signal irregularity, coefficient of variation of intervals, and fractal complexity (e.g., Higuchi fractal dimension). Signals from the 3-D accelerometer and 3-D gyroscope were high-pass filtered with a cutoff frequency of 0.2 Hz to reduce low-frequency drift and artifacts. Each axis (X, Y, Z) was segmented into non-overlapping 60-second windows and processed to extract statistical features (mean, standard deviation, variance, skewness, kurtosis, minimum, maximum, and range), signal energy features (RMS and AUC), and spectral characteristics (dominant frequency and spectral entropy). Object and ambient temperature signals were not filtered but were similarly segmented into 60-second windows and processed to extract basic descriptive statistics, including mean, median, standard deviation, minimum, maximum, and range. All features across modalities were combined into a unified feature matrix indexed by timestamp and subject ID. Arousal labels were resampled into 60-second non-overlapping windows to match the feature segmentation. A window was labeled as an arousal event if it contained any arousal occurrence within its duration, ensuring sensitivity to even brief arousal activity. This binary labeling approach allowed the model to learn from both isolated and clustered arousal events, supporting robust temporal prediction. The dataset was imbalanced, with arousal windows (Class 1) comprising 6.6% of the data and non-arousal windows (Class 0) accounting for 93.4%, reflecting the rarity of cortical arousals during sleep. While this approach simplifies the classification task, it introduces a limitation: multiple arousals occurring within the same 60-second window are treated as a single event. This may underestimate the actual number of arousals in windows with dense activity. We initially experimented with shorter windows (e.g., 30 seconds) to capture finer temporal dynamics. However, this led to increased false positives, likely because pre- and post-arousal changes over the signals extended beyond the arousal itself. Thus, the 60-second window length was selected as an optimal trade-off between capturing relevant signal changes and maintaining specificity. Additionally, arousals that spanned multiple windows, a potential source of edge effects, were observed in approximately 10% of cases. Given that most arousals lasted 8 to 12 seconds, this level of boundary overlap was considered acceptable within the 60-second segmentation framework. Machine Learning Framework and Feature Selection We evaluated and compared the performance of three classifiers: Logistic Regression As a baseline, we trained a Logistic Regression model with L2 regularization (Ridge penalty), which helps prevent overfitting and handles multicollinearity. The model was trained with subject-level z-scored features, class balancing, and LOOCV. Hyperparameters, including the regularization strength, were tuned using RandomizedSearchCV with 50 randomized iterations. While it offers greater interpretability, it lacks the capacity to model nonlinear interactions present in physiological time-series data. Gradient-Boosted Decision Tree Model (XGBoost) We also implemented XGBoost, a high-performance gradient-boosted decision tree model that incorporates both first- and second-order gradients. We tuned hyperparameters including learning rate, tree depth, subsampling rate, and L1/L2 penalties using RandomizedSearchCV with 50 randomized iterations. All training followed the same LOOCV protocol as the previous model. Bagged Tree Ensemble Model (Random Forest) We used a Random Forest classifier, known for its robustness to noise, ability to model nonlinear relationships and embedded feature importance analysis. Hyperparameters were optimized using RandomizedSearchCV with 50 randomized iterations. Tuned parameters included the number of trees, maximum depth, minimum samples per split and leaf node, and feature subsampling ratio. All training followed the same LOOCV protocol as the other models. The best-performing hyperparameters for each model, selected based on cross-validation performance across folds, are summarized in Table 2 . To account for inter-individual variability in physiological signals, all features were standardized per subject using z-score normalization. Columns with excessive missingness were removed, and the remaining missing values were imputed using subject-level k-nearest neighbors 37 . This method estimates missing values by averaging the feature values from the most similar observations in the dataset. Dimensionality reduction and feature selection were performed using Recursive Feature Elimination 38 within the training folds to retain only the most informative features for classification. Table 2 Best-performing Hyperparameters for Each Classifier Classifier Best Hyperparameters Logistic Regression C = 0.1336, penalty = 'l2', solver = 'liblinear', fit_intercept = True, max_iter = 1000, tol = 0.0039 XGBoost colsample_bytree = 0.7547, gamma = 4.6836, learning_rate = 0.0513, max_depth = 6, n_estimators = 91, reg_alpha = 0.2579, reg_lambda = 2.9800, scale_pos_weight = 2, subsample = 0.9929 Random Forest ccp_alpha = 0.0036, criterion = 'entropy', max_depth = 15, max_features = 0.3986, min_samples_leaf = 3, min_samples_split = 6, min_weight_fraction_leaf = 0.0037, n_estimators = 120 A LOOCV scheme was used, where each subject was held out in turn as the test fold while the remaining subjects were used for training. This approach ensured strict subject-level separation and prevented data leakage, supporting robust evaluation of model generalizability. To address the natural class imbalance between arousal and non-arousal events, a two-step resampling strategy was applied within each training fold. First, Tomek Links 39 were removed to clean the decision boundary, followed by Random Undersampling 40 to balance the class distribution during model fitting. Importantly, the held-out test subject was never undersampled, preserving the original data distribution for evaluation. Thresholds for classification were selected based on the precision-recall curve computed on the raw (non-resampled) version of the training data, ensuring that decision thresholds reflected realistic class ratios. The selected threshold was then applied to the test fold. Together, these classifiers enabled direct performance comparisons. The outputs were evaluated using window-based overlap metrics and correlation analyses, described in the next section. Model Comparison and Evaluation Model performance was assessed using both classification-based metrics and agreement-based statistical analyses, with careful consideration given to subject-level separation through LOOCV. For each model, the area under the ROC-AUC was computed to quantify overall discriminative ability. In addition, precision, recall, and F1-score, defined in equations (1) through (3), were calculated separately for arousal (Class 1) and non-arousal (Class 0) classes on a per-window basis. These equations quantify the performance of the model in different aspects: \(\:Precision\:=\frac{True\:Positives}{True\:Positives+False\:Positives}\) \(\:\left(1\right)\) \(\:Recall\:=\frac{True\:Positives}{True\:Positives\:+\:False\:Negatives}\) \(\:\left(2\right)\) \(\:{F}_{1}\:=2\frac{\left(Precision\:x\:Recall\right)}{(Precision\:+Recall)}\) \(\:\left(3\right)\) To ensure equal contribution from each subject and prevent performance estimates from being skewed by subjects with longer recordings or more events, all metrics (precision, recall, F1-score) were first computed individually for each left-out subject in the LOOCV framework. The final reported values (Table 1 ) represent the mean of per-subject metrics, formalized as: $$\:\stackrel{-}{M}\:=\frac{1}{S}{\sum\:}_{s=1}^{S}{M}^{\left(s\right)}$$ (4) Where: \(\:\stackrel{-}{M}\) Subject-averaged metric (e.g., precision, recall, F1-score) \(\:S\) Total number of subjects $$\:{M}^{\left(s\right)}:Metric\:value\:(e.g.,\:{Precision}^{\left(s\right)}=\:\frac{{True\:Positives}^{\left(s\right)}}{{True\:Positives}^{\left(s\right)}+\:{False\:\:Positives}^{\left(s\right)}}$$ In addition to discrete classification metrics, we evaluated the agreement between predicted arousals and ground truth arousals across subjects. The predicted arousal rate for each subject, defined as the number of arousal events per hour of total sleep time, was compared with the true arousal rate using Spearman’s rank correlation coefficient (ρ) and Kendall’s tau (τ) to assess monotonic relationships. Agreement between predicted and true arousal rate were further examined using Bland–Altman analysis 41 , which visualizes the bias and limits of agreement between model estimates and expert-scored references. Feature Importance Analysis After model training and evaluation, we analyzed feature importances using the Random Forest model trained on the entire dataset to capture generalizable patterns across all subjects. Random Forest determines feature importance by evaluating the total decrease in node impurity, such as Gini impurity, each feature contributes across all decision trees in the ensemble. Features that result in larger impurity reductions when used for splitting are considered more important 42 . This approach allows the model to naturally account for nonlinear relationships and feature interactions. To enhance interpretability and reduce noise from low-importance variables, we selected the top ranked features for post hoc analysis. This number was chosen empirically: including more than 30 features resulted in only marginal improvements in classification performance while increasing model complexity and risk of overfitting. The selected features represented a balanced trade-off between performance and interpretability and were used in downstream visualizations and interpretation. Abbreviations AASM American Academy of Sleep Medicine ADHD Attention-Deficit/Hyperactivity Disorder AUC Area Under the Curve CSI EEG Cardiac Sympathetic Index (HRV-derived) Electroencephalographic FuzzyEn Fuzzy Entropy (HRV-derived) HRV Heart Rate Variability HFD Higuchi fractal dimension (HRV-derived) LOOCV Leave-One-Subject-Out Cross-Validation PPG Photoplethysmography REM NREM RMS RMS AUC RMSSD ROC Rapid Eye Movement Non-Rapid Eye Movement Root Mean Square Root Mean Square Area Under the Curve Root Mean Square of Successive Differences Receiver Operating Characteristic SDNN Standard Deviation of NN Intervals Declarations Competing Interests JB, CF, and NB are shareholders of Tanzen Medical Inc. All other authors have no competing interests. Funding This study was supported by the National Institutes of Health under award number 1R43MH133495-01A1 (NIH SBIR Phase I). The funding agency was not involved in the study design, data collection, data analysis, decision to publish, or preparation of the manuscript. Author Contribution NB and CF conceptualized the research. MK and SK performed the research. NB supervised the research. MK, SC, and KB prepared the figures and wrote the manuscript. YA and QD contributed to data processing and supported manuscript revision. GK and JB contributed to the discussion section and manuscript revision. All authors reviewed and approved the final manuscript. Acknowledgement This research was supported by the National Institutes of Health under grant #1R43MH133495-01A1 (NIH SBIR Phase I). The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the National Institutes of Health or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein. Data Availability All data generated or analyzed during the current study are not publicly available due to institutional restrictions but are available from the senior author (NB) upon reasonable request and with approval from the University of Maryland, Baltimore County. References Morgan, B. J. et al. Neurocirculatory consequences of abrupt change in sleep state in humans. J. Appl. Physiol. 80 , 1627–1636 (1996). Xue, Y. et al. Durative sleep fragmentation with or without hypertension suppress rapid eye movement sleep and generate cerebrovascular dysfunction. Neurobiol. Dis. 184 , 106222 (2023). Chouchou, F. et al. Sympathetic overactivity due to sleep fragmentation is associated with elevated diurnal systolic blood pressure in healthy elderly subjects: the PROOF-SYNAPSE study. Eur. Heart J. 34 , 2122–2131 (2013). Cappuccio, F. P., Cooper, D., D’Elia, L., Strazzullo, P. & Miller, M. A. Sleep duration predicts cardiovascular outcomes: a systematic review and meta-analysis of prospective studies. Eur. Heart J. 32 , 1484–1492 (2011). Duan, D., Kim, L. J., Jun, J. C. & Polotsky, V. Y. Connecting insufficient sleep and insomnia with metabolic dysfunction. Ann. N Y Acad. Sci. 1519 , 94–117 (2023). Wajszilber, D., Santiseban, J. A. & Gruber, R. Sleep disorders in patients with ADHD: impact and management challenges. Nat. Sci. Sleep. 10 , 453–480 (2018). Lal, C., Strange, C. & Bachman, D. Neurocognitive impairment in obstructive sleep apnea. Chest 141 , 1601–1610 (2012). Owens, J. A. A clinical overview of sleep and attention-deficit/hyperactivity disorder in children and adolescents. J. Can. Acad. Child. Adolesc. Psychiatry J. Acad. Can. Psychiatr Enfant Adolesc. 18 , 92–102 (2009). Kushida, C. A. et al. Practice parameters for the indications for polysomnography and related procedures: an update for 2005. Sleep 28 , 499–521 (2005). Dement, W. & Kleitman, N. Cyclic variations in EEG during sleep and their relation to eye movements, body motility, and dreaming. Electroencephalogr. Clin. Neurophysiol. 9 , 673–690 (1957). Gerstenslager, B. & Slowik, J. M. Sleep Study. in StatPearls (StatPearls Publishing, 2025). Lee, Y. J., Lee, J. Y., Cho, J. H., Kang, Y. J. & Choi, J. H. Performance of consumer wrist-worn sleep tracking devices compared to polysomnography: a meta-analysis. J. Clin. Sleep. Med. 21 , 573–582 (2025). Lee, T. et al. Accuracy of 11 Wearable, Nearable, and Airable Consumer Sleep Trackers: Prospective Multicenter Validation Study. JMIR MHealth UHealth . 11 , e50983 (2023). Bogan, R. K. Effects of restless legs syndrome (RLS) on sleep. Neuropsychiatr Dis. Treat. 2 , 513–519 (2006). Bansal, K. et al. A pilot study to understand the relationship between cortical arousals and leg movements during sleep. Sci. Rep. 12 , 12685 (2022). Cortese, S. et al. Restless legs syndrome and attention-deficit/hyperactivity disorder: a review of the literature. Sleep 28 , 1007–1013 (2005). Ferri, R. et al. Heart rate and spectral EEG changes accompanying periodic and non-periodic leg movements during sleep. Clin. Neurophysiol. 118 , 438–448 (2007). Bobovych, S. et al. Low-power accurate sleep monitoring using a wearable multi-sensor ankle band. Smart Health . 16 , 100113 (2020). Jha, A. et al. Pilot study: can machine learning analyses of movement discriminate between leg movements in sleep (LMS) with vs. without cortical arousals? Sleep. Breath. 25 , 373–379 (2021). Bansal, K. et al. A pilot study to understand the relationship between cortical arousals and leg movements during sleep. Sci. Rep. 12 , 12685 (2022). Li, A., Chen, S., Quan, S. F., Powers, L. S. & Roveda, J. M. A deep learning-based algorithm for detection of cortical arousal during sleep. Sleep 43 , zsaa120 (2020). Pitson, D. J. & Stradling, J. R. Autonomic markers of arousal during sleep in patients undergoing investigation for obstructive sleep apnoea, their relationship to EEG arousals, respiratory events and subjective sleepiness. J. Sleep. Res. 7 , 53–59 (1998). Somers, V. K., Dyken, M. E., Mark, A. L. & Abboud, F. M. Sympathetic-nerve activity during sleep in normal subjects. N Engl. J. Med. 328 , 303–307 (1993). Patel, A. K., Reddy, V., Shumway, K. R., Araujo, J. F. & Physiology Sleep Stages. in StatPearls (StatPearls Publishing, 2025). Ricci, A. et al. Association of a novel EEG metric of sleep depth/intensity with attention-deficit/hyperactivity, learning, and internalizing disorders and their pharmacotherapy in adolescence. Sleep 45 , zsab287 (2022). Olsen, M. et al. Automatic, electrocardiographic-based detection of autonomic arousals and their association with cortical arousals, leg movements, and respiratory events in sleep. Sleep 41 , zsy006 (2018). Béres, S. & Hejjel, L. The minimal sampling frequency of the photoplethysmogram for accurate pulse rate variability parameters in healthy volunteers. Biomed. Signal. Process. Control . 68 , 102589 (2021). Kazemi, K., Laitala, J., Azimi, I., Liljeberg, P. & Rahmani, A. M. Robust PPG Peak Detection Using Dilated Convolutional Neural Networks. Sensors 22 , 6054 (2022). Bonnet, M. H. et al. The scoring of arousal in sleep: reliability, validity, and alternatives. J. Clin. Sleep. Med. JCSM Off Publ Am. Acad. Sleep. Med. 3 , 133–145 (2007). Davies, R. J., Belt, P. J., Roberts, S. J., Ali, N. J. & Stradling, J. R. Arterial blood pressure responses to graded transient arousal from sleep in normal humans. J. Appl. Physiol. Bethesda Md. 1985 74 , 1123–1130 (1993). Cameli, N. et al. Restless Sleep Disorder and the Role of Iron in Other Sleep-Related Movement Disorders and ADHD. Clin. Transl Neurosci. 7 , 18 (2023). Martins, R. et al. Sleep disturbance in children with attention-deficit hyperactivity disorder: A systematic review. Sleep. Sci. Sao Paulo Braz . 12 , 295–301 (2019). Lazzaro, G., Galassi, P., Bacaro, V., Vicari, S. & Menghini, D. Clinical characterization of children and adolescents with ADHD and sleep disturbances. Eur. Arch. Psychiatry Clin. Neurosci. 10.1007/s00406-024-01921-w (2024). Berry, R. B. et al. AASM Scoring Manual Updates for 2017 (Version 2.4). J. Clin. Sleep. Med. JCSM Off Publ Am. Acad. Sleep. Med. 13 , 665–666 (2017). Gommers, R. et al. scipy/scipy: SciPy 1.9.0. Zenodo (2022). 10.5281/zenodo.6940349 Makowski, D. et al. NeuroKit2: A Python toolbox for neurophysiological signal processing. Behav. Res. Methods . 53 , 1689–1696 (2021). Pedregosa, F. Scikit-learn: Machine learning in Python. (2011). Guyon, I., Weston, J., Barnhill, S. & Vapnik, V. Gene Selection for Cancer Classification using Support Vector Machines. Mach. Learn. 46 , 389–422 (2002). TOMEK, I. TWO MODIFICATIONS OF CNN. TWO Modif. CNN (1976). He, H. & Garcia, E. A. Learning from Imbalanced Data. IEEE Trans. Knowl. Data Eng. 21 , 1263–1284 (2009). Martin Bland, J., Altman, D. G., STATISTICAL METHODS FOR & ASSESSING AGREEMENT BETWEEN TWO METHODS OF CLINICAL MEASUREMENT. Lancet 327 , 307–310 (1986). Breiman, L. Random Forests. Mach. Learn. 45 , 5–32 (2001). Additional Declarations Competing interest reported. JB, CF, and NB are shareholders of Tanzen Medical Inc. All other authors have no competing interests. Cite Share Download PDF Status: Published Journal Publication published 18 Dec, 2025 Read the published version in Scientific Reports → Version 1 posted Editorial decision: Revision requested 23 Sep, 2025 Reviews received at journal 22 Sep, 2025 Reviewers agreed at journal 12 Sep, 2025 Reviewers agreed at journal 11 Sep, 2025 Reviewers agreed at journal 11 Sep, 2025 Reviews received at journal 10 Sep, 2025 Reviewers agreed at journal 01 Sep, 2025 Reviewers agreed at journal 25 Aug, 2025 Reviewers invited by journal 25 Aug, 2025 Editor assigned by journal 06 May, 2025 Submission checks completed at journal 04 May, 2025 First submitted to journal 01 May, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6574148","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":455036329,"identity":"f664bd32-b49b-430e-b7de-4c99e4893ccf","order_by":0,"name":"Murat Kucukosmanoglu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABKUlEQVRIiWNgGAWjYPCCBCBmg3GYGxg+sMH4Btg1HEDVwtjAOINkLcw8bNhVgoDujNyHnz8wpCX2z0hLfPgzxy6Pn39hm7RN2Z18PrHDBxh+FGBoMbuRbixxgCEnccaNtMPGvNuSiyVnPGyTzjn3zLJNOi2BsQfTYWY30hiAWioSN0ikt0kzbmNO3HDjYJt0btthAzbpHAMGHqxamH/AtEj+3FYP0WIJ1pL/gfEPVi1sYIdtkEg7JsG77XDihvONQOsgtjAwY7PlzDM2izMGacYzzjxLBvrleOLMGYzNlj3nngG1pBkclsGi5Xga842KimTZ/vY0w4c/t1Un9vMfPnjjR9kdA/nZyQ8fvvmDI6hRjJJIAJEHGBAkQcB/gATFo2AUjIJRMBIAANRKcZwNaCxmAAAAAElFTkSuQmCC","orcid":"","institution":"D-Prime LLC","correspondingAuthor":true,"prefix":"","firstName":"Murat","middleName":"","lastName":"Kucukosmanoglu","suffix":""},{"id":455036330,"identity":"b41e9956-de2d-4e14-822e-9891ddaed492","order_by":1,"name":"Sarah Conklin","email":"","orcid":"","institution":"D-Prime LLC","correspondingAuthor":false,"prefix":"","firstName":"Sarah","middleName":"","lastName":"Conklin","suffix":""},{"id":455036331,"identity":"f345002f-7118-4234-b234-d071bd3332c1","order_by":2,"name":"Kanika Bansal","email":"","orcid":"","institution":"University of Maryland","correspondingAuthor":false,"prefix":"","firstName":"Kanika","middleName":"","lastName":"Bansal","suffix":""},{"id":455036332,"identity":"35cd6925-1bb7-4035-94ae-e9b5dd9a6d12","order_by":3,"name":"Sena Kaya","email":"","orcid":"","institution":"University of Maryland","correspondingAuthor":false,"prefix":"","firstName":"Sena","middleName":"","lastName":"Kaya","suffix":""},{"id":455036333,"identity":"8b027ae5-3c4a-4f33-9d28-7f5db8095aa4","order_by":4,"name":"Yumna Anwar","email":"","orcid":"","institution":"University of Maryland","correspondingAuthor":false,"prefix":"","firstName":"Yumna","middleName":"","lastName":"Anwar","suffix":""},{"id":455036334,"identity":"9501931b-b3d1-48e5-98bf-589d23231485","order_by":5,"name":"Quang Dang","email":"","orcid":"","institution":"University of Maryland","correspondingAuthor":false,"prefix":"","firstName":"Quang","middleName":"","lastName":"Dang","suffix":""},{"id":455036335,"identity":"9156027c-e9a2-4162-9fba-c039e6dcac5d","order_by":6,"name":"Golshan Kargosha","email":"","orcid":"","institution":"D-Prime LLC","correspondingAuthor":false,"prefix":"","firstName":"Golshan","middleName":"","lastName":"Kargosha","suffix":""},{"id":455036336,"identity":"4d9b2236-cb81-4012-a94e-30fbf954d7ff","order_by":7,"name":"Justin Brooks","email":"","orcid":"","institution":"D-Prime LLC","correspondingAuthor":false,"prefix":"","firstName":"Justin","middleName":"","lastName":"Brooks","suffix":""},{"id":455036337,"identity":"42d1cba7-0c73-4c88-9475-455e5905c865","order_by":8,"name":"Cody Feltch","email":"","orcid":"","institution":"Tanzen Medical Inc","correspondingAuthor":false,"prefix":"","firstName":"Cody","middleName":"","lastName":"Feltch","suffix":""},{"id":455036338,"identity":"f90b6e74-427c-4e28-aa0f-8e571f4ac802","order_by":9,"name":"Nilanjan Banerjee","email":"","orcid":"","institution":"University of Maryland","correspondingAuthor":false,"prefix":"","firstName":"Nilanjan","middleName":"","lastName":"Banerjee","suffix":""}],"badges":[],"createdAt":"2025-05-01 20:38:06","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6574148/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6574148/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s41598-025-27739-7","type":"published","date":"2025-12-18T15:58:35+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":82892501,"identity":"346acf84-68bd-49cb-a3de-727ab1e78f8e","added_by":"auto","created_at":"2025-05-16 12:19:41","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":349292,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTop 30 Features for cortical arousal classification. \u003c/strong\u003eTop features ranked by importance using a Random Forest model. Feature importance was determined based on the mean decrease in impurity.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-6574148/v1/9e9e91854be8ab8fbf54ec31.png"},{"id":82892497,"identity":"35e94d6e-2232-4a14-bb99-935fc7634d76","added_by":"auto","created_at":"2025-05-16 12:19:41","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":241436,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eArousal rate correlation.\u003c/strong\u003eCorrelation between predicted and true arousal rates (n = 14). Strong positive correlations were observed (Spearman’s ρ = 0.89, \u003cem\u003ep\u003c/em\u003e = 2.00 × 10⁻⁵; Kendall’s τ = 0.76, \u003cem\u003ep\u003c/em\u003e = 3.95 × 10⁻⁵). The solid line represents the best-fit linear regression: \u003cem\u003ey\u003c/em\u003e = 0.72\u003cem\u003ex\u003c/em\u003e + 0.32.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-6574148/v1/c222db4a09d8c4f239fe3b83.png"},{"id":82892498,"identity":"59ced699-50c5-4b63-95a8-61222cade900","added_by":"auto","created_at":"2025-05-16 12:19:41","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":199374,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eBland–Altman plot for arousal rates\u003c/strong\u003e. Bland–Altman plot comparing predicted and true (expert-labeled) arousal rates. The mean difference was +0.88 arousals per hour (Predicted − True), with 95% limits of agreement ranging from −1.40 to +3.17.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-6574148/v1/4e05a470b91931c8e3aab984.png"},{"id":82894225,"identity":"45e8962e-22f0-4a74-90d1-139a66faba53","added_by":"auto","created_at":"2025-05-16 12:35:41","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":530947,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTemporal prediction of cortical arousals.\u003c/strong\u003e Predicted versus true cortical arousal events for three ADHD participants. Each subplot shows 1-minute window predictions across the sleep period (x-axis in hours). Blue crosses represent model-predicted arousals, and red circles indicate ground truth events.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-6574148/v1/5b93baebb5edd55afdd03fcd.png"},{"id":82893717,"identity":"9dbe40b4-0d1b-4927-8a6f-d03554c4a942","added_by":"auto","created_at":"2025-05-16 12:27:41","extension":"jpeg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":284790,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eMultimodal data preprocessing pipeline for arousal classification\u003c/strong\u003e. Raw data from the RestEaze™ wearable system included PPG, 3-D accelerometer, 3-D gyroscope, and temperature sensors\u003c/p\u003e","description":"","filename":"floatimage5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6574148/v1/22bbf26347e6db965565adb8.jpeg"},{"id":82893720,"identity":"337b58fb-fcf0-46d9-8abd-17e576d09f62","added_by":"auto","created_at":"2025-05-16 12:27:41","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":561937,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePPG signal preprocessing and peak detection\u003c/strong\u003e. The top panel (a) shows the raw LED green PPG signal, which contains low-frequency drift and movement-related noise. The middle panel (b) displays the same signal after linear interpolation to 200 Hz and bandpass filtering (0.2–5 Hz). The bottom panel (c) shows the corresponding PPG signal quality over time, with values closer to 1 indicating cleaner, more reliable signal segments.\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-6574148/v1/878ca6f3511c3f536cc3eef2.png"},{"id":82893723,"identity":"cc971c39-a82a-400f-a388-fec5c81233c8","added_by":"auto","created_at":"2025-05-16 12:27:41","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":298313,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAccelerometer and gyroscope feature trends across sleep.\u003c/strong\u003e The top panel (a) shows the standard deviation of the X-axis accelerometer signal, reflecting variability in leg movement amplitude. The bottom panel (b) displays the spectral entropy of the Z-axis gyroscope signal, which quantifies the irregularity or complexity of rotational motion. Red markers indicate windows labeled as arousals, while blue markers denote non-arousal periods.\u003c/p\u003e","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-6574148/v1/11adf1516f2e3c644d753a50.png"},{"id":98814370,"identity":"faa8d1e2-9e2f-4ac3-916b-a4f4e2861b46","added_by":"auto","created_at":"2025-12-22 16:12:27","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2803633,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6574148/v1/5050207f-2261-47dc-8095-04045f905f69.pdf"}],"financialInterests":"Competing interest reported. JB, CF, and NB are shareholders of Tanzen Medical Inc. All other authors have no competing interests.","formattedTitle":"Detection of Cortical Arousals in Sleep Using Multimodal Wearable Sensors and Machine Learning","fulltext":[{"header":"Introduction","content":"\u003cp\u003eCortical arousals are brief interruptions in electroencephalographic (EEG) activity that fragment sleep without full awakening. Although transient, these arousals contribute to autonomic activation and disrupted sleep pattern, with growing evidence linking them to hypertension, cognitive decline, and elevated cardiovascular risk \u003csup\u003e\u003cspan additionalcitationids=\"CR2\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. Total sleep duration less than 5 hours per night is considered high-risk for cardiovascular morbidity and mortality \u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e. Disrupted or insufficient sleep has also been associated with systemic inflammation, metabolic dysfunction, and increased all-cause mortality \u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. Elevated rates of sleep disturbances, including cortical and autonomic arousals, have also been observed in children with attention-deficit/hyperactivity disorder (ADHD) \u003csup\u003e\u003cspan additionalcitationids=\"CR7\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e. Early and accurate detection of these arousals may offer clinical insights into the relationship between poor sleep quality and daytime behavioral symptoms that may reveal patterns that differ by clinical subtype.\u003c/p\u003e \u003cp\u003ePolysomnography remains the gold standard for detecting cortical arousals \u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e,\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e, yet its high cost, complexity, and requirement for overnight clinical supervision limit its use for large-scale or long-term monitoring \u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e. Consumer sleep technologies, such as sleep trackers, offer a non-invasive, scalable approach to sleep monitoring, with the potential to support early identification of sleep fragmentation in home environments. While these devices offer greater accessibility, they often suffer from poor agreement with polysomnography, particularly in detecting brief or motionless arousals \u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e. A multicenter validation study involving 11 wearable, nearable, and airable consumer sleep trackers confirmed substantial variation in performance across devices, with some showing macro F1-scores as low as 0.26 when compared to Polysomnography \u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e. However, the growing integration of wearable sleep technologies into daily life offers a valuable opportunity to develop advanced frameworks that can effectively use these technologies to detect clinically relevant features of sleep.\u003c/p\u003e \u003cp\u003eOne promising solution involves tracking leg movements during sleep, which frequently occur alongside cortical arousals, especially in populations with conditions like restless leg syndrome, periodic limb movement disorder, or ADHD \u003csup\u003e\u003cspan additionalcitationids=\"CR15 CR16\" citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e. Recent studies using wearable leg sensors have shown that leg movements during sleep features can effectively distinguish arousals, and that leg-EEG signal coupling may reflect deeper physiological mechanisms of sleep disruption \u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e,\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e. In this study, we evaluate multimodal sensor data from a leg-worn wearable, RestEaze\u0026trade;, to detect cortical arousals using interpretable machine learning models, with the aim of advancing practical and reliable sleep health monitoring solutions outside of traditional clinical settings.\u003c/p\u003e \u003cp\u003eThe RestEaze\u0026trade; system integrates accelerometry, gyroscope, photoplethysmography (PPG), and temperature sensors, offering a comprehensive view of movement and physiological dynamics during sleep. In a prior pilot study using a similar platform, we introduced neuro-extremity analysis, a novel approach that employed Granger causal modeling to assess the temporal and directional relationships between cortical arousals and leg movements \u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. That study revealed that textile-based capacitive sensors showed stronger temporal and spectral coupling with EEG-theta oscillations than inertial sensors, and more accurately identified expert-labeled cortical arousals. These findings support the hypothesis that leg movements and cortical arousals are driven by coordinated activity within a shared central arousal system. The current study builds upon this work by incorporating PPG and temperature sensors into the previously studied system and focusing exclusively on inertial sensors for movement detection, as they were found to reliably capture arousal-related leg movements while avoiding the redundancy and implementation challenges associated with textile-based capacitive sensors. This setup allows extraction of heart rate (HR) and heart rate variability (HRV) features that may offer additional insight into autonomic activation during sleep \u003csup\u003e\u003cspan additionalcitationids=\"CR22\" citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eSleep is composed of two main states: rapid eye movement (REM) sleep and non-rapid eye movement (NREM) sleep. NREM includes three stages: N1, N2, and N3, which progress from light to deep sleep. These stages repeat in cycles throughout the night \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e. We began by examining the distribution of cortical arousals across sleep stages to establish a physiological context for the classification task. Arousals occurred most frequently during N2 sleep, with a mean proportion of 56.77% (95% confidence interval [CI]: 46.14\u0026ndash;67.40%), followed by N1 at 17.47% (95% CI: 8.15\u0026ndash;26.79%), REM at 13.17% (95% CI: 4.43\u0026ndash;21.90%), and N3 at 12.60% (95% CI: 7.17\u0026ndash;18.02%), averaged across subjects. This distribution aligns with established sleep physiology: N2 sleep not only comprises a larger portion of total sleep time but also has a lower arousal threshold, making it more prone to cortical arousals due to its transitional nature between wakefulness and deeper sleep stages\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e. Similarly, the elevated rate of arousals during N1 reflects its light sleep status and proximity to wakefulness. Interestingly, we also observed notable levels of arousals during N3 and REM sleep, suggesting increased cortical arousal beyond the lighter stages. This pattern may support prior findings showing that adolescents with ADHD and learning disorders exhibit increased cortical arousal during N2 and N3 sleep, particularly in central and frontal brain regions \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eTo enable real-time detection of these arousal events using wearable data, we implemented and evaluated machine learning models designed to classify arousals from multimodal physiological signals. We evaluated the performance of three machine learning classifiers: Logistic Regression, XGBoost, and Random Forest for detecting cortical arousals based on multimodal physiological data from a leg-worn wearable device on full cohort of 14 children with ADHD, a population known to experience elevated levels of sleep fragmentation and frequent cortical arousals \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. We chose these models to represent different levels of complexity and explainability: Logistic Regression as a simple linear baseline, Random Forest as a robust ensemble method, and XGBoost as a state-of-the-art gradient boosting algorithm.\u003c/p\u003e\n\u003cp\u003eAll models were trained using a leave-one-subject-out cross-validation (LOOCV) approach to ensure robust subject-independent evaluation. The classification task involved identifying arousal events versus non-arousal periods. Evaluation metrics included class-wise precision, recall, F1-score, and overall Receiver Operating Characteristic - Area Under the Curve (ROC-AUC).\u003c/p\u003e\n\u003cp\u003eModel training and performance\u003c/p\u003e\n\u003cp\u003eThe performance of each model is summarized in Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e. While all three classifiers showed high accuracy in detecting non-arousal periods (Class 0), their ability to detect arousal events (Class 1) varied considerably. Logistic Regression achieved a Class 1 F1-score of 0.57 and a ROC-AUC of 0.90. XGBoost improved precision but had lower recall for Class 1, with a resulting F1-score of 0.61 and a ROC-AUC of 0.93. Random Forest achieved the best balance, with a Class 1 F1-score of 0.65 and the highest ROC-AUC of 0.94. Based on these results, the Random Forest model was selected for further analysis. Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e summarizes the performance of each model.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n \u003cdiv class=\"colspec\" align=\"left\"\u003e\u0026nbsp;\u003c/div\u003e\n \u003ctable id=\"Tab1\" border=\"1\"\u003e\n \u003ccaption\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eModel Performance Summary\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eClass\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eF1-Score\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eROC-AUC\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" align=\"left\"\u003e\n \u003cp\u003eLogistic Regression\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" align=\"left\"\u003e\n \u003cp\u003e0.90\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.45\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.84\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.57\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" align=\"left\"\u003e\n \u003cp\u003eXGBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" align=\"left\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.61\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" align=\"left\"\u003e\n \u003cp\u003eRandom Forest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" align=\"left\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.57\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.65\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003ch2\u003eFeature Importance\u003c/h2\u003e\n\u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e presents the ranked list of the most important features contributing to cortical arousal classification, as determined by the Random Forest model. These features were predominantly derived from accelerometer and gyroscope signals, with a smaller contribution from HR and HRV metrics. The most important features included statistical, energy-based, and entropy-related measures. Importantly, standard deviation, root mean square (RMS), maximum, and range from the x-axis of the accelerometer appeared prominently in the ranking. This suggests that lateral leg movement (x-direction) plays a critical role in arousal episodes, consistent with biomechanical patterns observed during limb movement\u0026ndash;related arousals.\u003c/p\u003e\n\u003cp\u003eEntropy-based features such as spectral entropy from both accelerometer and gyroscope signals were also among the top-ranked predictors. These features reflect the signal complexity or irregularity during sleep and are useful for capturing subtle variations in movement associated with arousals. Similarly, RMS AUC (Root Mean Square Area Under the Curve) quantifies cumulative signal energy, which is often elevated during microarousals due to brief bursts of leg activity.\u003c/p\u003e\n\u003cp\u003eOther contributing features included HRV-derived indices such as HRV Higuchi fractal dimension (HRV-HFD), HRV Cardiac Sympathetic Index (HRV-CSI), and HRV Fuzzy Entropy (HRV-FuzzyEn), all of which reflect beat-to-beat HRV complexity, physiological markers known to fluctuate during autonomic arousals \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e. However, they were less important than movement-based metrics, suggesting a stronger motor component to arousals in children with ADHD. Similarly, temperature-based features were not among the top-ranked predictors, indicating minimal relevance to arousal classification in this context.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIn addition to feature rankings, we analyzed PPG signal quality across arousal categories. The mean PPG quality score was 0.818 (95% CI: 0.738\u0026ndash;0.899) during non-arousal periods and 0.488 (95% CI: 0.420\u0026ndash;0.556) during arousal events. This significant decline in signal quality during arousals suggests increased motion artifacts or sensor dropout, which may explain the lower importance of PPG-derived features in the final model.\u003c/p\u003e\n\u003cp\u003eAgreement with Ground Truth\u003c/p\u003e\n\u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e shows the model prediction of the arousal rates against the true arousal rates (ground truth). In this study, arousal rate refers to the number of 60-second windows that contain at least one cortical arousal event, normalized per hour of total sleep time. The predicted rates exhibited a strong correlation with the ground truth, yielding a Spearman\u0026rsquo;s rank correlation coefficient\u003c/p\u003e\n\u003cp\u003e\u0026rho;\u0026thinsp;=\u0026thinsp;0.89 (p\u0026thinsp;=\u0026thinsp;2.00 \u0026times; 10⁻⁵) and a Kendall\u0026rsquo;s \u0026tau;\u0026thinsp;=\u0026thinsp;0.76 (p\u0026thinsp;=\u0026thinsp;3.95 \u0026times; 10⁻⁵).\u003c/p\u003e\n\u003cp\u003eThese results show a strong relationship, suggesting that the model successfully preserves subject-wise ranking in arousal frequency, which is crucial for estimating severity and comparing individuals.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe fitted linear regression line further supports the alignment between predicted and true values. The slope below 1.0 indicates underestimation at higher arousal rates, yet the close clustering of points around the line reflects consistency in the overall prediction trend. The regression slope was statistically significant (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.01), with a 95% CI of [0.383, 1.050].\u003c/p\u003e\n\u003cp\u003eTo further assess agreement, a Bland\u0026ndash;Altman analysis was conducted (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e). This plot shows the differences between predicted and true arousal rates as a function of their average, both expressed in arousals per hour. The mean difference was +\u0026thinsp;0.88 arousals/hour (Predicted\u0026thinsp;\u0026minus;\u0026thinsp;True), indicating a slight overall tendency of the model to overestimate arousal frequency. The 95% limits of agreement ranged from \u0026minus;\u0026thinsp;1.40 to +\u0026thinsp;3.17 arousals/hour.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTemporal Prediction Patterns\u003c/p\u003e\n\u003cp\u003eTo evaluate model behavior across time, we visualized prediction sequences for three subjects who showed distinct arousal patterns. Figure\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e shows minute-by-minute comparisons between predicted and true arousals across the sleep duration.\u003c/p\u003e\n\u003cp\u003eFor Subject A (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003ea), who exhibited frequent and widely distributed arousals, the model effectively captured both isolated and clustered events throughout the night. Minute-by-minute inspection showed that most predictions were temporally aligned with ground truth, with several pre-arousal predictions appearing within one to two minutes of labeled events.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIn contrast, Subject B (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003eb) presented arousals that occurred in distinct temporal clusters during the early and late portions of the recording. The model maintained high temporal precision, correctly identifying contiguous arousal periods while avoiding false positives during quiescent intervals. Subject C (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003ec) exhibited a sparser distribution of arousals. The model\u0026apos;s predictions closely matched the few true events, with overclassification toward the end.\u003c/p\u003e\n\u003cp\u003eThe agreement between predicted and true arousals is quantified using Arousals (Class 1) F1-scores: 0.62 (a), 0.68 (b), and 0.54 (c). These scores indicate strong model performance given the substantial class imbalance, where arousals make up only\u0026thinsp;~\u0026thinsp;6% of the data. For context, random guessing would yield an F1-score near 0.06, making the observed values highly meaningful. These subject-level, minute-by-minute visualizations highlight the model\u0026rsquo;s adaptability to inter-individual variability in sleep and arousal patterns.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study demonstrates the feasibility of using multimodal wearable sensors and machine learning to detect cortical arousals during sleep, offering an accessible alternative to traditional in-clinic polysomnography. Among the classifiers tested, the Random Forest model achieved the best balance between recall and precision, yielding the highest ROC-AUC of 0.94. This result is consistent with Random Forest\u0026rsquo;s ability to handle complex patterns, feature interactions, and imbalanced data. Its ensemble-based architecture and embedded feature selection likely contributed to its robustness in this complex real-world dataset. Compared to Logistic Regression, which assumes linearity, and XGBoost, which can be sensitive to hyperparameter tuning in small datasets, the Random Forest model proved particularly effective at capturing subtle, subject-specific arousal signatures.\u003c/p\u003e \u003cp\u003eFeature importance analysis further revealed that the most predictive signals were derived from accelerometry and gyroscope data, particularly features reflecting signal variability and complexity, such as root mean square amplitude, standard deviation, and spectral entropy. These findings are consistent with prior work suggesting that leg movements are linked with cortical arousals \u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e,\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e,\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e. Entropy measures likely captured the fragmented nature of movement during arousals. In contrast, HR and HRV features extracted from PPG contributed less prominently to model performance. This was not entirely unexpected, as the original sampling rate of 25 Hz may be insufficient for accurate HRV estimation. Prior work has shown that HRV metrics like Standard Deviation of NN Intervals (SDNN) and Root Mean Square of Successive Differences (RMSSD) require significantly higher sampling rates to ensure reliability, at least 50 Hz for SDNN and 100 Hz or more for RMSSD without interpolation \u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e. Additionally, signal quality issues further limited the reliability of PPG-derived features. These noises, primarily motion artifacts and high-frequency noise, are inevitable in wearable-based health and well-being monitoring systems and can significantly impact peak detection accuracy \u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e. In our study, the average PPG signal quality declined from 0.818 during non-arousal periods to 0.488 during arousal. This indicates a consistent drop in signal quality during arousal events.\u003c/p\u003e \u003cp\u003eInterestingly, the model predicted more arousals than were annotated by experts, particularly in subjects with sparse arousal profiles (Subject C). Rather than representing pure false positives, these predictions may reflect physiological events, such as sub-threshold arousals or autonomic activations, that were not captured by EEG-based criteria. This raises the possibility that wearable sensors may detect some physiological markers of sleep disruption that fall outside the boundaries of current clinical scoring systems. Indeed, prior research has shown that physiological changes surrounding arousal events can be significant, often extending beyond the boundaries of EEG-defined arousals \u003csup\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e,\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e. These findings highlight how machine learning and wearables can improve sleep assessment beyond conventional methods. Additionally, the use of 60-second windows may have contributed to some discrepancy by grouping multiple arousals into a single event or capturing signal fluctuations surrounding true arousals.\u003c/p\u003e \u003cp\u003eLastly, our subject-independent and interpretable framework provides minute-level temporal precision, making it suitable for clinical applications that require generalizable detection. It shows promise for individuals with ADHD, a group often underserved by traditional sleep diagnostics. Pediatric restless legs syndrome, for example, can cause significant sleep disruption, behavioral issues, and impaired daytime functioning that mimic ADHD symptoms \u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e,\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e. While ADHD's recognized subtypes (inattentive, hyperactive-impulsive, and combined) are well-described, their association with distinct sleep profiles remains unclear, highlighting the need for detailed pediatric sleep assessment \u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e. Refined at-home monitoring could help identify specific sleep disorders and support more personalized, subtype-targeted treatments for pediatric ADHD. Building on these findings, this work presents multiple opportunities for future development. Priorities include expanding to larger and more diverse datasets, using deep learning to model long-range patterns, and incorporating continuous arousal scoring to reflect subtle physiological changes. Real-world feedback such as sleep staging, user experiences, and device usability will be vital for transforming this research into a practical home-based health solution. Ultimately, these efforts aim to bring clinical-quality sleep analytics into everyday environments through smart and accessible wearables.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis study presents a non-invasive, wearable-based framework for detecting cortical arousals using multimodal physiological signals from a leg-worn device. Among the classifiers evaluated, the Random Forest model performed best, achieving a ROC-AUC of 0.94 and showing strong alignment with expert-labeled EEG annotations. Key predictive features, such as leg movement variability and signal entropy, support the role of movement-related physiological signals as markers of central arousals. These findings demonstrate the potential of systems like RestEaze\u0026trade; for clinically meaningful, at-home sleep monitoring. Future work should include larger, more diverse populations and explore continuous arousal scoring to enhance clinical relevance.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003eParticipants and Data Acquisition\u003c/p\u003e\n\u003cp\u003ePhysiological and movement data were collected from 14 children diagnosed with ADHD using the RestEaze\u0026trade; Movement Analyzer, a wireless, leg-worn wearable designed for non-intrusive sleep monitoring and arousal detection. More details about the RestEaze\u0026trade; can be found in previous publication \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e. As illustrated in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e, the RestEaze\u0026trade; device integrates multiple synchronized sensors:\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cp\u003eA 3-D accelerometer and 3-D gyroscope embedded within an inertial measurement unit (IMU) for leg movement and orientation tracking,\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eA PPG sensor for capturing cardiovascular dynamics, and\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eObject and ambient temperature sensors for thermal signature during sleep.\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eThe accelerometer (X, Y, Z axes), gyroscope (X, Y, Z axes), and PPG channels (IR, red, green LEDs) were all sampled at 25 Hz, providing high-resolution capture of biomechanical and cardiovascular signals. Temperature data was sampled at 0.2 Hz, appropriate for monitoring slow-changing thermal conditions.\u003c/p\u003e\n\u003cp\u003eThis setup enables continuous, multimodal recording throughout the night, capturing both fine-grained leg movements and physiological fluctuations associated with cortical arousals. Across the 14 participants, the average total sleep time was approximately 7.59 hours per subject, totaling 106.32 hours of recorded sleep data. Data collection was conducted during natural sleep in a home or clinical setting.\u003c/p\u003e\n\u003cp\u003eAll study procedures were approved by the Institutional Review Board of Johns Hopkins University. Research was conducted in accordance with the Declaration of Helsinki and all relevant ethical guidelines and regulations, including obtaining informed consent from all participants and/or their legal guardians.\u003c/p\u003e\n\u003cp\u003eCortical Arousals Rate\u003c/p\u003e\n\u003cp\u003eCortical arousals (ground truth) were identified and scored according to the guidelines set by the American Academy of Sleep Medicine (AASM) \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e34\u003c/span\u003e\u003c/sup\u003e, which define arousals as abrupt shifts in EEG frequency, including alpha, theta, or activity exceeding 16 Hz, that last for at least 3 seconds and occur after a minimum of 10 seconds of uninterrupted sleep \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e. Arousal rate was calculated as the number of 60-second windows labeled with at least one cortical arousal event, normalized per hour of total sleep time. Specifically, if any arousal occurred within a given 60-second segment, the entire window was labeled as an arousal window (Class 1). The resulting arousal rate, expressed in arousal windows per hour, provides a temporally consistent metric for comparing arousal frequency across individuals.\u003c/p\u003e\n\u003cp\u003eIn addition to cortical arousals, sleep stages, and limb movements were scored manually by trained technicians according to the AASM guidelines \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e34\u003c/span\u003e\u003c/sup\u003e. Bilateral limb movement events were also manually annotated, whereas leg movement channels were scored using an automated algorithm via the Sleepware G3 platform (Philips Respironics, US). Final scoring was reviewed and confirmed by a board-certified sleep physician and AASM fellow.\u003c/p\u003e\n\u003cp\u003ePreprocessing and Feature Generation\u003c/p\u003e\n\u003cp\u003eAll raw sensor signals were processed using a unified preprocessing pipeline (see Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e), which included filtering, segmentation into 60-second non-overlapping windows, and modality-specific feature extraction. The choice of a 60-second window was guided by the need to balance temporal resolution with physiological interpretability. Each one-minute segment contains sufficient cardiac cycles (typically 60\u0026ndash;100 beats) to allow reliable estimation of HR and HRV, while also being short enough to detect changes in physiological state over time.\u003c/p\u003e\n\u003cp\u003eFor the PPG signal, the preprocessing began with upsampling to 200 Hz using linear interpolation. This step was essential for achieving the temporal resolution required for accurate peak detection and compatibility with feature extraction functions that assume higher sampling rates. Several methods did not perform at the native 25 Hz resolution, especially those involving frequency-domain HRV metrics. The upsampled signal was then bandpass filtered between 0.2 and 5 Hz using a Butterworth filter to remove baseline drift and suppress motion artifacts. The filter was implemented in Python 3.11 using the butter and filtfilt functions from the scipy.signal module, which apply zero-phase forward and reverse filtering to avoid phase distortion \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e35\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eFollowing filtering, we evaluated several peak detection strategies to identify heartbeats from the PPG waveform. Among these, the ppg-findpeaks function from the NeuroKit2 library \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e36\u003c/span\u003e\u003c/sup\u003e provided reliable results in terms of peak timing consistency and robustness to signal noise. Figure\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e shows the effects of preprocessing: the top panel displays the raw PPG signal with notable baseline fluctuations (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003ea), the middle panel shows the filtered waveform with clearly resolved peaks (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003eb), and the bottom panel plots the computed PPG signal quality over time (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003ec). This quality metric, ranging from 0 to 1, reflects the reliability of the signal for physiological analysis.\u003c/p\u003e\n\u003cp\u003eOnce peaks were detected, HR and HRV features were extracted from each 60-second window. HR metrics included minimum, maximum, and mean HR. HRV features encompassed time-domain measures (e.g., RMSSD, SDNN), frequency-domain indices (e.g., low-frequency/high-frequency ratio), and nonlinear metrics such as entropy, coefficient of signal irregularity, coefficient of variation of intervals, and fractal complexity (e.g., Higuchi fractal dimension).\u003c/p\u003e\n\u003cp\u003eSignals from the 3-D accelerometer and 3-D gyroscope were high-pass filtered with a cutoff frequency of 0.2 Hz to reduce low-frequency drift and artifacts. Each axis (X, Y, Z) was segmented into non-overlapping 60-second windows and processed to extract statistical features (mean, standard deviation, variance, skewness, kurtosis, minimum, maximum, and range), signal energy features (RMS and AUC), and spectral characteristics (dominant frequency and spectral entropy). Object and ambient temperature signals were not filtered but were similarly segmented into 60-second windows and processed to extract basic descriptive statistics, including mean, median, standard deviation, minimum, maximum, and range.\u003c/p\u003e\n\u003cp\u003eAll features across modalities were combined into a unified feature matrix indexed by timestamp and subject ID. Arousal labels were resampled into 60-second non-overlapping windows to match the feature segmentation. A window was labeled as an arousal event if it contained any arousal occurrence within its duration, ensuring sensitivity to even brief arousal activity. This binary labeling approach allowed the model to learn from both isolated and clustered arousal events, supporting robust temporal prediction. The dataset was imbalanced, with arousal windows (Class 1) comprising 6.6% of the data and non-arousal windows (Class 0) accounting for 93.4%, reflecting the rarity of cortical arousals during sleep.\u003c/p\u003e\n\u003cp\u003eWhile this approach simplifies the classification task, it introduces a limitation: multiple arousals occurring within the same 60-second window are treated as a single event. This may underestimate the actual number of arousals in windows with dense activity. We initially experimented with shorter windows (e.g., 30 seconds) to capture finer temporal dynamics. However, this led to increased false positives, likely because pre- and post-arousal changes over the signals extended beyond the arousal itself. Thus, the 60-second window length was selected as an optimal trade-off between capturing relevant signal changes and maintaining specificity. Additionally, arousals that spanned multiple windows, a potential source of edge effects, were observed in approximately 10% of cases. Given that most arousals lasted 8 to 12 seconds, this level of boundary overlap was considered acceptable within the 60-second segmentation framework.\u003c/p\u003e\n\u003cp\u003eMachine Learning Framework and Feature Selection\u003c/p\u003e\n\u003cp\u003eWe evaluated and compared the performance of three classifiers:\u003c/p\u003e\n\u003cp\u003eLogistic Regression\u003c/p\u003e\n\u003cp\u003eAs a baseline, we trained a Logistic Regression model with L2 regularization (Ridge penalty), which helps prevent overfitting and handles multicollinearity. The model was trained with subject-level z-scored features, class balancing, and LOOCV. Hyperparameters, including the regularization strength, were tuned using RandomizedSearchCV with 50 randomized iterations. While it offers greater interpretability, it lacks the capacity to model nonlinear interactions present in physiological time-series data.\u003c/p\u003e\n\u003cp\u003eGradient-Boosted Decision Tree Model (XGBoost)\u003c/p\u003e\n\u003cp\u003eWe also implemented XGBoost, a high-performance gradient-boosted decision tree model that incorporates both first- and second-order gradients. We tuned hyperparameters including learning rate, tree depth, subsampling rate, and L1/L2 penalties using RandomizedSearchCV with 50 randomized iterations. All training followed the same LOOCV protocol as the previous model.\u003c/p\u003e\n\u003cp\u003eBagged Tree Ensemble Model (Random Forest)\u003c/p\u003e\n\u003cp\u003eWe used a Random Forest classifier, known for its robustness to noise, ability to model nonlinear relationships and embedded feature importance analysis. Hyperparameters were optimized using RandomizedSearchCV with 50 randomized iterations. Tuned parameters included the number of trees, maximum depth, minimum samples per split and leaf node, and feature subsampling ratio. All training followed the same LOOCV protocol as the other models. The best-performing hyperparameters for each model, selected based on cross-validation performance across folds, are summarized in Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e\n\u003cp\u003eTo account for inter-individual variability in physiological signals, all features were standardized per subject using z-score normalization. Columns with excessive missingness were removed, and the remaining missing values were imputed using subject-level k-nearest neighbors \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e37\u003c/span\u003e\u003c/sup\u003e. This method estimates missing values by averaging the feature values from the most similar observations in the dataset. Dimensionality reduction and feature selection were performed using Recursive Feature Elimination\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e38\u003c/span\u003e\u003c/sup\u003e within the training folds to retain only the most informative features for classification.\u0026nbsp;\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n\u003ctable id=\"Tab2\" border=\"1\"\u003e\u003ccaption\u003e\n\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\n\u003cdiv class=\"CaptionContent\"\u003e\n\u003cp\u003eBest-performing Hyperparameters for Each Classifier\u003c/p\u003e\n\u003c/div\u003e\n\u003c/caption\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eClassifier\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eBest Hyperparameters\u003c/p\u003e\n\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eLogistic Regression\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eC\u0026thinsp;=\u0026thinsp;0.1336, penalty = 'l2', solver = 'liblinear', fit_intercept\u0026thinsp;=\u0026thinsp;True, max_iter\u0026thinsp;=\u0026thinsp;1000, tol\u0026thinsp;=\u0026thinsp;0.0039\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eXGBoost\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003ecolsample_bytree\u0026thinsp;=\u0026thinsp;0.7547, gamma\u0026thinsp;=\u0026thinsp;4.6836, learning_rate\u0026thinsp;=\u0026thinsp;0.0513, max_depth\u0026thinsp;=\u0026thinsp;6, n_estimators\u0026thinsp;=\u0026thinsp;91, reg_alpha\u0026thinsp;=\u0026thinsp;0.2579, reg_lambda\u0026thinsp;=\u0026thinsp;2.9800, scale_pos_weight\u0026thinsp;=\u0026thinsp;2, subsample\u0026thinsp;=\u0026thinsp;0.9929\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eRandom Forest\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eccp_alpha\u0026thinsp;=\u0026thinsp;0.0036, criterion = 'entropy', max_depth\u0026thinsp;=\u0026thinsp;15, max_features\u0026thinsp;=\u0026thinsp;0.3986, min_samples_leaf\u0026thinsp;=\u0026thinsp;3, min_samples_split\u0026thinsp;=\u0026thinsp;6, min_weight_fraction_leaf\u0026thinsp;=\u0026thinsp;0.0037, n_estimators\u0026thinsp;=\u0026thinsp;120\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eA LOOCV scheme was used, where each subject was held out in turn as the test fold while the remaining subjects were used for training. This approach ensured strict subject-level separation and prevented data leakage, supporting robust evaluation of model generalizability.\u003c/p\u003e\n\u003cp\u003eTo address the natural class imbalance between arousal and non-arousal events, a two-step resampling strategy was applied within each training fold. First, Tomek Links \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e39\u003c/span\u003e\u003c/sup\u003e were removed to clean the decision boundary, followed by Random Undersampling \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e to balance the class distribution during model fitting. Importantly, the held-out test subject was never undersampled, preserving the original data distribution for evaluation. Thresholds for classification were selected based on the precision-recall curve computed on the raw (non-resampled) version of the training data, ensuring that decision thresholds reflected realistic class ratios. The selected threshold was then applied to the test fold.\u003c/p\u003e\n\u003cp\u003eTogether, these classifiers enabled direct performance comparisons. The outputs were evaluated using window-based overlap metrics and correlation analyses, described in the next section.\u003c/p\u003e\n\u003cp\u003eModel Comparison and Evaluation\u003c/p\u003e\n\u003cp\u003eModel performance was assessed using both classification-based metrics and agreement-based statistical analyses, with careful consideration given to subject-level separation through LOOCV. For each model, the area under the ROC-AUC was computed to quantify overall discriminative ability. In addition, precision, recall, and F1-score, defined in equations (1) through (3), were calculated separately for arousal (Class 1) and non-arousal (Class 0) classes on a per-window basis. These equations quantify the performance of the model in different aspects:\u003c/p\u003e\n\u003cdiv class=\"Heading\"\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:Precision\\:=\\frac{True\\:Positives}{True\\:Positives+False\\:Positives}\\)\u003c/span\u003e\u003c/span\u003e \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\left(1\\right)\\)\u003c/span\u003e\u003c/span\u003e\u003c/div\u003e\n\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\n\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:Recall\\:=\\frac{True\\:Positives}{True\\:Positives\\:+\\:False\\:Negatives}\\)\u003c/span\u003e\u003c/span\u003e \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\left(2\\right)\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{F}_{1}\\:=2\\frac{\\left(Precision\\:x\\:Recall\\right)}{(Precision\\:+Recall)}\\)\u003c/span\u003e\u003c/span\u003e \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\left(3\\right)\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003eTo ensure equal contribution from each subject and prevent performance estimates from being skewed by subjects with longer recordings or more events, all metrics (precision, recall, F1-score) were first computed individually for each left-out subject in the LOOCV framework. The final reported values (Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e) represent the mean of per-subject metrics, formalized as:\u003c/p\u003e\n\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\n\u003cdiv id=\"FileID_Equa\" class=\"mathdisplay\"\u003e$$\\:\\stackrel{-}{M}\\:=\\frac{1}{S}{\\sum\\:}_{s=1}^{S}{M}^{\\left(s\\right)}$$\u003c/div\u003e\n\u003cdiv class=\"mathdisplay\"\u003e(4)\u003c/div\u003e\n\u003c/div\u003e\n\u003cp\u003eWhere:\u003c/p\u003e\n\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:\\stackrel{-}{M}\\)\u003c/span\u003e \u003c/span\u003e Subject-averaged metric (e.g., precision, recall, F1-score)\u003c/p\u003e\n\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:S\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003eTotal number of subjects\u003c/p\u003e\n\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\n\u003cdiv id=\"FileID_Equb\" class=\"mathdisplay\"\u003e$$\\:{M}^{\\left(s\\right)}:Metric\\:value\\:(e.g.,\\:{Precision}^{\\left(s\\right)}=\\:\\frac{{True\\:Positives}^{\\left(s\\right)}}{{True\\:Positives}^{\\left(s\\right)}+\\:{False\\:\\:Positives}^{\\left(s\\right)}}$$\u003c/div\u003e\n\u003cdiv class=\"mathdisplay\"\u003e\u0026nbsp;\u003c/div\u003e\n\u003c/div\u003e\n\u003cp\u003eIn addition to discrete classification metrics, we evaluated the agreement between predicted arousals and ground truth arousals across subjects. The predicted arousal rate for each subject, defined as the number of arousal events per hour of total sleep time, was compared with the true arousal rate using Spearman\u0026rsquo;s rank correlation coefficient (\u0026rho;) and Kendall\u0026rsquo;s tau (\u0026tau;) to assess monotonic relationships. Agreement between predicted and true arousal rate were further examined using Bland\u0026ndash;Altman analysis \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e41\u003c/span\u003e\u003c/sup\u003e, which visualizes the bias and limits of agreement between model estimates and expert-scored references.\u003c/p\u003e\n\u003cp\u003eFeature Importance Analysis\u003c/p\u003e\n\u003cp\u003eAfter model training and evaluation, we analyzed feature importances using the Random Forest model trained on the entire dataset to capture generalizable patterns across all subjects. Random Forest determines feature importance by evaluating the total decrease in node impurity, such as Gini impurity, each feature contributes across all decision trees in the ensemble. Features that result in larger impurity reductions when used for splitting are considered more important \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e42\u003c/span\u003e\u003c/sup\u003e. This approach allows the model to naturally account for nonlinear relationships and feature interactions. To enhance interpretability and reduce noise from low-importance variables, we selected the top ranked features for post hoc analysis. This number was chosen empirically: including more than 30 features resulted in only marginal improvements in classification performance while increasing model complexity and risk of overfitting. The selected features represented a balanced trade-off between performance and interpretability and were used in downstream visualizations and interpretation.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"Abbreviations","content":" \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Taba\" border=\"1\"\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cdiv class=\"SimplePara\"\u003eAASM\u003c/div\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cdiv class=\"SimplePara\"\u003eAmerican Academy of Sleep Medicine\u003c/div\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cdiv class=\"SimplePara\"\u003eADHD\u003c/div\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cdiv class=\"SimplePara\"\u003eAttention-Deficit/Hyperactivity Disorder\u003c/div\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cdiv class=\"SimplePara\"\u003eAUC\u003c/div\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cdiv class=\"SimplePara\"\u003eArea Under the Curve\u003c/div\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cdiv class=\"SimplePara\"\u003eCSI\u003c/div\u003e \u003cdiv class=\"SimplePara\"\u003eEEG\u003c/div\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cdiv class=\"SimplePara\"\u003eCardiac Sympathetic Index (HRV-derived)\u003c/div\u003e \u003cdiv class=\"SimplePara\"\u003eElectroencephalographic\u003c/div\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cdiv class=\"SimplePara\"\u003eFuzzyEn\u003c/div\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cdiv class=\"SimplePara\"\u003eFuzzy Entropy (HRV-derived)\u003c/div\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cdiv class=\"SimplePara\"\u003eHRV\u003c/div\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cdiv class=\"SimplePara\"\u003eHeart Rate Variability\u003c/div\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cdiv class=\"SimplePara\"\u003eHFD\u003c/div\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cdiv class=\"SimplePara\"\u003eHiguchi fractal dimension (HRV-derived)\u003c/div\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cdiv class=\"SimplePara\"\u003eLOOCV\u003c/div\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cdiv class=\"SimplePara\"\u003eLeave-One-Subject-Out Cross-Validation\u003c/div\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cdiv class=\"SimplePara\"\u003ePPG\u003c/div\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cdiv class=\"SimplePara\"\u003ePhotoplethysmography\u003c/div\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cdiv class=\"SimplePara\"\u003eREM\u003c/div\u003e \u003cdiv class=\"SimplePara\"\u003eNREM\u003c/div\u003e \u003cdiv class=\"SimplePara\"\u003eRMS\u003c/div\u003e \u003cdiv class=\"SimplePara\"\u003eRMS AUC\u003c/div\u003e \u003cdiv class=\"SimplePara\"\u003eRMSSD\u003c/div\u003e \u003cdiv class=\"SimplePara\"\u003eROC\u003c/div\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cdiv class=\"SimplePara\"\u003eRapid Eye Movement\u003c/div\u003e \u003cdiv class=\"SimplePara\"\u003eNon-Rapid Eye Movement\u003c/div\u003e \u003cdiv class=\"SimplePara\"\u003eRoot Mean Square\u003c/div\u003e \u003cdiv class=\"SimplePara\"\u003eRoot Mean Square Area Under the Curve\u003c/div\u003e \u003cdiv class=\"SimplePara\"\u003eRoot Mean Square of Successive Differences\u003c/div\u003e \u003cdiv class=\"SimplePara\"\u003eReceiver Operating Characteristic\u003c/div\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cdiv class=\"SimplePara\"\u003eSDNN\u003c/div\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cdiv class=\"SimplePara\"\u003eStandard Deviation of NN Intervals\u003c/div\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003cbr/\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003ch2\u003eCompeting Interests\u003c/h2\u003e\u003cp\u003eJB, CF, and NB are shareholders of Tanzen Medical Inc. All other authors have no competing interests.\u003c/p\u003e\u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eThis study was supported by the National Institutes of Health under award number 1R43MH133495-01A1 (NIH SBIR Phase I). The funding agency was not involved in the study design, data collection, data analysis, decision to publish, or preparation of the manuscript.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eNB and CF conceptualized the research. MK and SK performed the research. NB supervised the research. MK, SC, and KB prepared the figures and wrote the manuscript. YA and QD contributed to data processing and supported manuscript revision. GK and JB contributed to the discussion section and manuscript revision. All authors reviewed and approved the final manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e\u003cp\u003eThis research was supported by the National Institutes of Health under grant #1R43MH133495-01A1 (NIH SBIR Phase I). The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the National Institutes of Health or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eAll data generated or analyzed during the current study are not publicly available due to institutional restrictions but are available from the senior author (NB) upon reasonable request and with approval from the University of Maryland, Baltimore County.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eMorgan, B. J. et al. Neurocirculatory consequences of abrupt change in sleep state in humans. \u003cem\u003eJ. Appl. Physiol.\u003c/em\u003e \u003cb\u003e80\u003c/b\u003e, 1627\u0026ndash;1636 (1996).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXue, Y. et al. Durative sleep fragmentation with or without hypertension suppress rapid eye movement sleep and generate cerebrovascular dysfunction. \u003cem\u003eNeurobiol. Dis.\u003c/em\u003e \u003cb\u003e184\u003c/b\u003e, 106222 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChouchou, F. et al. Sympathetic overactivity due to sleep fragmentation is associated with elevated diurnal systolic blood pressure in healthy elderly subjects: the PROOF-SYNAPSE study. \u003cem\u003eEur. Heart J.\u003c/em\u003e \u003cb\u003e34\u003c/b\u003e, 2122\u0026ndash;2131 (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCappuccio, F. P., Cooper, D., D\u0026rsquo;Elia, L., Strazzullo, P. \u0026amp; Miller, M. A. Sleep duration predicts cardiovascular outcomes: a systematic review and meta-analysis of prospective studies. \u003cem\u003eEur. Heart J.\u003c/em\u003e \u003cb\u003e32\u003c/b\u003e, 1484\u0026ndash;1492 (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDuan, D., Kim, L. J., Jun, J. C. \u0026amp; Polotsky, V. Y. Connecting insufficient sleep and insomnia with metabolic dysfunction. \u003cem\u003eAnn. N Y Acad. Sci.\u003c/em\u003e \u003cb\u003e1519\u003c/b\u003e, 94\u0026ndash;117 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWajszilber, D., Santiseban, J. A. \u0026amp; Gruber, R. Sleep disorders in patients with ADHD: impact and management challenges. \u003cem\u003eNat. Sci. Sleep.\u003c/em\u003e \u003cb\u003e10\u003c/b\u003e, 453\u0026ndash;480 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLal, C., Strange, C. \u0026amp; Bachman, D. Neurocognitive impairment in obstructive sleep apnea. \u003cem\u003eChest\u003c/em\u003e \u003cb\u003e141\u003c/b\u003e, 1601\u0026ndash;1610 (2012).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOwens, J. A. A clinical overview of sleep and attention-deficit/hyperactivity disorder in children and adolescents. \u003cem\u003eJ. Can. Acad. Child. Adolesc. Psychiatry J. Acad. Can. Psychiatr Enfant Adolesc.\u003c/em\u003e \u003cb\u003e18\u003c/b\u003e, 92\u0026ndash;102 (2009).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKushida, C. A. et al. Practice parameters for the indications for polysomnography and related procedures: an update for 2005. \u003cem\u003eSleep\u003c/em\u003e \u003cb\u003e28\u003c/b\u003e, 499\u0026ndash;521 (2005).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDement, W. \u0026amp; Kleitman, N. Cyclic variations in EEG during sleep and their relation to eye movements, body motility, and dreaming. \u003cem\u003eElectroencephalogr. Clin. Neurophysiol.\u003c/em\u003e \u003cb\u003e9\u003c/b\u003e, 673\u0026ndash;690 (1957).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGerstenslager, B. \u0026amp; Slowik, J. M. \u003cem\u003eSleep Study. in StatPearls\u003c/em\u003e (StatPearls Publishing, 2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee, Y. J., Lee, J. Y., Cho, J. H., Kang, Y. J. \u0026amp; Choi, J. H. Performance of consumer wrist-worn sleep tracking devices compared to polysomnography: a meta-analysis. \u003cem\u003eJ. Clin. Sleep. Med.\u003c/em\u003e \u003cb\u003e21\u003c/b\u003e, 573\u0026ndash;582 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee, T. et al. Accuracy of 11 Wearable, Nearable, and Airable Consumer Sleep Trackers: Prospective Multicenter Validation Study. \u003cem\u003eJMIR MHealth UHealth\u003c/em\u003e. \u003cb\u003e11\u003c/b\u003e, e50983 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBogan, R. K. Effects of restless legs syndrome (RLS) on sleep. \u003cem\u003eNeuropsychiatr Dis. Treat.\u003c/em\u003e \u003cb\u003e2\u003c/b\u003e, 513\u0026ndash;519 (2006).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBansal, K. et al. A pilot study to understand the relationship between cortical arousals and leg movements during sleep. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e12\u003c/b\u003e, 12685 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCortese, S. et al. Restless legs syndrome and attention-deficit/hyperactivity disorder: a review of the literature. \u003cem\u003eSleep\u003c/em\u003e \u003cb\u003e28\u003c/b\u003e, 1007\u0026ndash;1013 (2005).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFerri, R. et al. Heart rate and spectral EEG changes accompanying periodic and non-periodic leg movements during sleep. \u003cem\u003eClin. Neurophysiol.\u003c/em\u003e \u003cb\u003e118\u003c/b\u003e, 438\u0026ndash;448 (2007).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBobovych, S. et al. Low-power accurate sleep monitoring using a wearable multi-sensor ankle band. \u003cem\u003eSmart Health\u003c/em\u003e. \u003cb\u003e16\u003c/b\u003e, 100113 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJha, A. et al. Pilot study: can machine learning analyses of movement discriminate between leg movements in sleep (LMS) with vs. without cortical arousals? \u003cem\u003eSleep. Breath.\u003c/em\u003e \u003cb\u003e25\u003c/b\u003e, 373\u0026ndash;379 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBansal, K. et al. A pilot study to understand the relationship between cortical arousals and leg movements during sleep. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e12\u003c/b\u003e, 12685 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi, A., Chen, S., Quan, S. F., Powers, L. S. \u0026amp; Roveda, J. M. A deep learning-based algorithm for detection of cortical arousal during sleep. \u003cem\u003eSleep\u003c/em\u003e \u003cb\u003e43\u003c/b\u003e, zsaa120 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePitson, D. J. \u0026amp; Stradling, J. R. Autonomic markers of arousal during sleep in patients undergoing investigation for obstructive sleep apnoea, their relationship to EEG arousals, respiratory events and subjective sleepiness. \u003cem\u003eJ. Sleep. Res.\u003c/em\u003e \u003cb\u003e7\u003c/b\u003e, 53\u0026ndash;59 (1998).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSomers, V. K., Dyken, M. E., Mark, A. L. \u0026amp; Abboud, F. M. Sympathetic-nerve activity during sleep in normal subjects. \u003cem\u003eN Engl. J. Med.\u003c/em\u003e \u003cb\u003e328\u003c/b\u003e, 303\u0026ndash;307 (1993).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePatel, A. K., Reddy, V., Shumway, K. R., Araujo, J. F. \u0026amp; Physiology \u003cem\u003eSleep Stages. in StatPearls\u003c/em\u003e (StatPearls Publishing, 2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRicci, A. et al. Association of a novel EEG metric of sleep depth/intensity with attention-deficit/hyperactivity, learning, and internalizing disorders and their pharmacotherapy in adolescence. \u003cem\u003eSleep\u003c/em\u003e \u003cb\u003e45\u003c/b\u003e, zsab287 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOlsen, M. et al. Automatic, electrocardiographic-based detection of autonomic arousals and their association with cortical arousals, leg movements, and respiratory events in sleep. \u003cem\u003eSleep\u003c/em\u003e \u003cb\u003e41\u003c/b\u003e, zsy006 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eB\u0026eacute;res, S. \u0026amp; Hejjel, L. The minimal sampling frequency of the photoplethysmogram for accurate pulse rate variability parameters in healthy volunteers. \u003cem\u003eBiomed. Signal. Process. Control\u003c/em\u003e. \u003cb\u003e68\u003c/b\u003e, 102589 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKazemi, K., Laitala, J., Azimi, I., Liljeberg, P. \u0026amp; Rahmani, A. M. Robust PPG Peak Detection Using Dilated Convolutional Neural Networks. \u003cem\u003eSensors\u003c/em\u003e \u003cb\u003e22\u003c/b\u003e, 6054 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBonnet, M. H. et al. The scoring of arousal in sleep: reliability, validity, and alternatives. \u003cem\u003eJ. Clin. Sleep. Med. JCSM Off Publ Am. Acad. Sleep. Med.\u003c/em\u003e \u003cb\u003e3\u003c/b\u003e, 133\u0026ndash;145 (2007).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDavies, R. J., Belt, P. J., Roberts, S. J., Ali, N. J. \u0026amp; Stradling, J. R. Arterial blood pressure responses to graded transient arousal from sleep in normal humans. \u003cem\u003eJ. Appl. Physiol. Bethesda Md.\u003c/em\u003e \u003cb\u003e1985 74\u003c/b\u003e, 1123\u0026ndash;1130 (1993).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCameli, N. et al. Restless Sleep Disorder and the Role of Iron in Other Sleep-Related Movement Disorders and ADHD. \u003cem\u003eClin. Transl Neurosci.\u003c/em\u003e \u003cb\u003e7\u003c/b\u003e, 18 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMartins, R. et al. Sleep disturbance in children with attention-deficit hyperactivity disorder: A systematic review. \u003cem\u003eSleep. Sci. Sao Paulo Braz\u003c/em\u003e. \u003cb\u003e12\u003c/b\u003e, 295\u0026ndash;301 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLazzaro, G., Galassi, P., Bacaro, V., Vicari, S. \u0026amp; Menghini, D. Clinical characterization of children and adolescents with ADHD and sleep disturbances. \u003cem\u003eEur. Arch. Psychiatry Clin. Neurosci.\u003c/em\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s00406-024-01921-w\u003c/span\u003e\u003cspan address=\"10.1007/s00406-024-01921-w\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBerry, R. B. et al. AASM Scoring Manual Updates for 2017 (Version 2.4). \u003cem\u003eJ. Clin. Sleep. Med. JCSM Off Publ Am. Acad. Sleep. Med.\u003c/em\u003e \u003cb\u003e13\u003c/b\u003e, 665\u0026ndash;666 (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGommers, R. et al. scipy/scipy: SciPy 1.9.0. Zenodo (2022). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.5281/zenodo.6940349\u003c/span\u003e\u003cspan address=\"10.5281/zenodo.6940349\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMakowski, D. et al. NeuroKit2: A Python toolbox for neurophysiological signal processing. \u003cem\u003eBehav. Res. Methods\u003c/em\u003e. \u003cb\u003e53\u003c/b\u003e, 1689\u0026ndash;1696 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePedregosa, F. Scikit-learn: Machine learning in Python. (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGuyon, I., Weston, J., Barnhill, S. \u0026amp; Vapnik, V. Gene Selection for Cancer Classification using Support Vector Machines. \u003cem\u003eMach. Learn.\u003c/em\u003e \u003cb\u003e46\u003c/b\u003e, 389\u0026ndash;422 (2002).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTOMEK, I. TWO MODIFICATIONS OF CNN. TWO Modif. CNN (1976).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHe, H. \u0026amp; Garcia, E. A. Learning from Imbalanced Data. \u003cem\u003eIEEE Trans. Knowl. Data Eng.\u003c/em\u003e \u003cb\u003e21\u003c/b\u003e, 1263\u0026ndash;1284 (2009).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMartin Bland, J., Altman, D. G., STATISTICAL METHODS FOR \u0026amp; ASSESSING AGREEMENT BETWEEN TWO METHODS OF CLINICAL MEASUREMENT. \u003cem\u003eLancet\u003c/em\u003e \u003cb\u003e327\u003c/b\u003e, 307\u0026ndash;310 (1986).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBreiman, L. Random Forests. \u003cem\u003eMach. Learn.\u003c/em\u003e \u003cb\u003e45\u003c/b\u003e, 5\u0026ndash;32 (2001).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"cortical arousals, RestEaze, ADHD, wearables, machine learning, sleep monitoring","lastPublishedDoi":"10.21203/rs.3.rs-6574148/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6574148/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eCortical arousals are brief brain activations that disrupt sleep continuity and contribute to cardiovascular, cognitive, and behavioral impairments. Although polysomnography is the gold standard for arousal detection, its cost and complexity limit use in long-term or home-based monitoring. This study presents a noninvasive machine learning based framework for detecting cortical arousals using the RestEaze\u0026trade; system, a leg-worn wearable that records multimodal physiological signals including accelerometry, gyroscope, photoplethysmography (PPG), and temperature. Across multiple methods tested, including logistic regression, XGBoost, and Random Forest classifiers, we found that features related to movement intensity were the most effective in identifying cortical arousals, while heart rate variability had a comparatively lower impact. The framework was evaluated in 14 children with attention-deficit/hyperactivity disorder (ADHD) who were being assessed for possible restless leg syndrome related sleep disruption. The Random Forest model achieved the best performance, with a ROC AUC of 0.94. For the arousal class specifically, it reached a precision of 0.57, recall of 0.78, and F1-score of 0.65. These findings support the feasibility of wearable-based machine learning for real-world arousal detection, demonstrated here in a pediatric ADHD cohort with sleep-related behavioral concerns.\u003c/p\u003e","manuscriptTitle":"Detection of Cortical Arousals in Sleep Using Multimodal Wearable Sensors and Machine Learning","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-16 12:19:36","doi":"10.21203/rs.3.rs-6574148/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-09-23T11:18:43+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-09-22T21:55:08+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"183001383362137814033678314611431969973","date":"2025-09-12T07:48:33+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"9135714506570972876626326112387412219","date":"2025-09-12T01:24:15+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"157092797042111271608701229029044519607","date":"2025-09-11T15:50:16+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-09-11T00:58:55+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"134292512781586096294294694071488464673","date":"2025-09-01T14:35:45+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"135440797781230078441143643200960799363","date":"2025-08-25T14:48:50+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-08-25T10:28:16+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-05-06T11:21:06+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-05-04T18:22:56+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2025-05-01T20:30:35+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"f51d2b9d-0da0-492b-ac1f-90211af42b76","owner":[],"postedDate":"May 16th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":48370735,"name":"Biological sciences/Neuroscience"},{"id":48370736,"name":"Biological sciences/Physiology"},{"id":48370737,"name":"Health sciences/Biomarkers"},{"id":48370738,"name":"Health sciences/Medical research"}],"tags":[],"updatedAt":"2025-12-22T16:07:14+00:00","versionOfRecord":{"articleIdentity":"rs-6574148","link":"https://doi.org/10.1038/s41598-025-27739-7","journal":{"identity":"scientific-reports","isVorOnly":false,"title":"Scientific Reports"},"publishedOn":"2025-12-18 15:58:35","publishedOnDateReadable":"December 18th, 2025"},"versionCreatedAt":"2025-05-16 12:19:36","video":"","vorDoi":"10.1038/s41598-025-27739-7","vorDoiUrl":"https://doi.org/10.1038/s41598-025-27739-7","workflowStages":[]},"version":"v1","identity":"rs-6574148","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6574148","identity":"rs-6574148","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}