Robert Joseph Thomas

STUDY OBJECTIVES: Pathological respiratory chemoreflex activation (high loop gain) in sleep apnea may activate the carotid body and induce excessive sympathetic nerve output, and independently add to the risk of adverse outcomes. This study aims to determine whether a high loop gain of the respiratory control system during sleep, estimated by respiratory self-similarity (SS) and the algorithm-derived central respiratory event index (CREI), is associated with new-onset cardiovascular disease (CVD) and CVD-related mortality. METHODS: A secondary analysis of The Sleep Heart Health Study (SHHS) and Multi-Ethnic Study of Atherosclerosis (MESA). High SS% and CREI were defined as the highest quartile. Lasso regression and restricted cubic spline assessed the relationship between SS%, CREI and CVD, CVD-mortality and all-cause mortality. A cumulative incidence function and a Fine-Gray subdistribution hazard model were used to evaluate this outcome accounting for competing risks. RESULTS: In adjusted models with age as a categorical variable divided by 65 years old, the highest quartile of SS% (HR 1.693, p: 0.017 in MESA and HR 1.240, p : 0.009 in SHHS) was associated with the incidence of CVD and was more prominent in males. A positive relationship between SS% (MESA: HR 1.429, p: 0.039; SHHS HR1.176, p:0.018 ) for the continuous metrics and CVD mortality in both cohorts was more apparent in men than in women. The categorical high-exposure association with CVD mortality was not consistent across cohorts.The associations were significantly attenuated when adjusted by age as a continuous metric. CONCLUSIONS: An endotype of sleep apnea reflecting high loop gain is associated with CVD and CVD mortality, especially in males.

Obstructive sleep apnea (OSA) is a heterogeneous disorder whose associations across the entire phenome remain incompletely characterized by hypothesis-driven outcome-specific analyses. We constructed a multidomain atlas for phenome-wide association study of OSA in 222,109 adults undergoing diagnostic sleep testing in Kaiser Permanente Southern California. Using apnea-hypopnea index, we examined cross-sectional associations across diagnoses, laboratory tests, medications, procedures, healthcare utilization, behaviors, and symptoms; associations with incident diseases; and data-driven cross-domain phenotypes. After full adjustment, 169 prevalent phecodes and 16 incident phecodes remained Bonferroni-significant. Positive associations converged most consistently on cardiometabolic diseases, including hypertension, impaired glucose metabolism, hyperlipidemia, obesity, and heart failure, with concordant laboratory, medication, and cardiac-measurement signals. Multidimensional physiologic, symptom, and comorbidity phenotypes showed heterogeneous disease incidence after adjustment, with patterns not aligned with apnea-hypopnea index severity alone. These findings define the broad clinical landscape of OSA and provide a comprehensive atlas for replication and hypothesis generation.

Study Objectives: Polysomnography (PSG) is the gold standard for diagnosing sleep disorders, which are recognized as cardiovascular risk factors. Single-lead electrocardiograms (ECGs) are recorded during PSG but rarely analyzed, presenting an opportunity to leverage overnight ECGs for cardiovascular risk prediction. Materials and Methods: We applied a deep residual neural network with attention to single-lead ECGs from PSG, combined with sleep stage data, to predict 10-year risk of atrial fibrillation (AF), stroke, myocardial infarction (MI), heart failure (HF), and death. The network, trained for arrhythmia detection, was finetuned on 15 809 Massachusetts General Hospital patients. External validation used cohorts from Emory University Hospital (EUH) (n = 9810) and Beth Israel Deaconess Medical Center (BIDMC) (n = 12 576), with outcomes derived from electronic health records (ICD-9/10 codes). Results: Cox proportional hazards models showed that neural-network output (NN-output) was strongly associated with long-term cardiovascular risk (p < .0001), even after adjusting for risk factors (age, sex, body mass index, smoking, hypertension, diabetes) and sleep characteristics (apnea-hypopnea-index, arousal-index, periodic limb movement index, time spend in N1-N3, REM and sleep efficiency). A one-standard deviation increase in NN-output was associated with elevated hazard ratios (95% CI): AF 2.03 (1.82-2.27) (EUH) and 2.72 (2.31-3.21) (BIDMC); stroke 1.19 (1.09-1.31) and 1.40 (1.12-1.74); MI 1.38 (1.22-1.55) and 1.10 (0.86-1.40); HF 1.69 (1.55-1.85) and 2.33 (2.05-2.65); all-cause mortality 1.82 (1.60-2.06) and 1.65 (1.41-1.93). Conclusions: Deep learning applied to nocturnal ECGs with sleep stage information captures cardiac abnormalities, offering a tool for cardiovascular risk stratification. Integrating such prediction into PSG could enhance detection and support personalized clinical decision-making.

Rationale: Obstructive sleep apnea (OSA) is highly prevalent yet largely under-diagnosed. Current screening strategies rely on effortful questionnaires with modest accuracy, and prior machine learning models often require resource-intensive inputs and lack external validation. Objectives: To develop and validate machine learning models to predict OSA (apnea-hypopnea index [AHI4%]≥5) and moderate-severe OSA (AHI4%≥15) using only routinely collected electronic health record (EHR) data, test cross-system transferability and compare performance with clinical screeners and a Bayesian risk-updating framework. Methods: Three tree-ensemble families (XGBoost, random forests, and histogram-based gradient boosting) were trained on an adult cohort with sleep-study results at Kaiser Permanente (KPSC), using full (106-feature) and minimal (age, sex, body-mass index, race/ethnicity) predictor sets and tuned by nested cross-validation. XGBoost models were selected, internally and temporally validated, and externally validated at University of Kansas Medical Center (KUMC). Measurements and Main Results: On our internal hold-out set (n=53,048), the full XGBoost model reached a ROC-AUC of 0.791 for any OSA and 0.768 for moderate-severe OSA. The minimal model retained most performance (0.764 and 0.736). Temporal performance was stable. Externally, frozen models showed lower AUCs (0.718 and 0.690 for full models; 0.725 and 0.706 for minimal models) but similar to KUMC-native models and improved further with transfer learning. The models significantly outperformed STOP-BANG (AUC difference 0.06-0.09 for predicting any OSA; p<0.001). Bayesian updating with select patient-reported items further improved discrimination. Conclusions: Machine learning models based on EHR data can predict OSA probability that outperform existing screeners, transfers across systems, and offers potential for EHR implementation. At a Glance Commentary: Scientific Knowledge on the Subject: Obstructive sleep apnea is common and heavily underdiagnosed, yet upstream patient identification still depends on symptom and anthropometry questionnaires such as STOP-BANG, which are effort-dependent and achieve modest performance. Prior machine learning models for OSA are heterogeneous, often developed on small or demographically narrow cohorts, frequently rely on physiologic signals that reintroduce a testing step, and are rarely validated across independent health systems. What This Study Adds to the Field: We developed machine learning models from routinely collected electronic health record data, comprising a full (106-feature) and a minimal (4-feature) model, to predict any OSA and moderate-severe OSA. The study is distinguished by multiple validation efforts, including temporal validation and external validation at a second health system. We further explored adaptation strategies for external adoption using transfer learning techniques. Finally, the models outperformed existing screeners such as STOP-BANG, and we describe a Bayesian framework that optionally incorporates select patient-reported symptoms to refine individual risk estimates. Overall, these models are feasible to implement within EHR systems as a low-effort, passive screener that can support encounter-based and population health outreach screening strategies.

Introduction: Maintaining long-term positive airway pressure (CPAP) adherence is a major challenge when treating obstructive sleep apnea (OSA). We evaluated factors associated with one-year CPAP adherence and developed models predicting Month-12 adherence and CPAP-use trajectories. Methods: CPAP-naive adults with OSA (apnea-hypopnea index [AHI]≥5) at Kaiser Permanente Southern California with 364-day observation period were analyzed (development/internal-validation cohort, n=14,906; temporal holdout, n=1,676). Logistic regression evaluated baseline associations with Month-12 adherence (mean≥4 hours/night during days 331-360). XGBoost models utilized baseline features (clinical, sleep-study, questionnaire data), early CPAP-only data (days 1-7), or combined to predict long-term CPAP use. A latent-class model identified trajectories over days 8-364. Predictive models were evaluated in both random and temporal holdouts. Results: About one-third (34.6%) of the primary cohort (49.7±13.8 years, 63.2% male) were adherent during Month-12. Higher AHI and longer self-reported sleep duration favored adherence while greater comorbidity burden, depression, and substance-use disorder were associated with lower adherence. In held-out testing, adherence prediction using baseline features yielded an AUROC of 0.678. Using CPAP-only data, AUROCs were higher from 0.787 with seven days of usage data to 0.816 and 0.854 with 14 and 30 days. Models combining CPAP and baseline features increased Day-7 AUROC to 0.810. We identified four one-year trajectories: Early-Decliners (42.9%), Gradual-Decliners (17.4%), Stable-Moderate (19.0%), Stable-High (20.7%). Models predicting patient trajectory revealed macro-average AUROCs of 0.619, 0.756, and 0.770 for baseline, CPAP-only, and combined models. Conclusion: Baseline pAbstractTextredictors may identify strategies for adherence support, while early CPAP data substantially improved predictive performance. These models may guide personalized follow-up but should not restrict CPAP access. External and implementation validation are needed.

Study Objectives: Sleep research has been limited by the lack of large, diverse polysomnography (PSG) datasets. Existing resources are often single-center, age-specific, or epidemiological in nature, restricting generalizability and large-scale clinical research. The objective of this study was to create a large, multi-center clinical PSG dataset to support scalable and broadly representative research in sleep medicine and clinical neuroscience. Materials and Methods: The Human Sleep Project (HSP) dataset comprises 119 234 overnight sleep recordings from more than 90 000 unique patients collected across five major US institutions: Stanford University, Massachusetts General Brigham, Emory University, Beth Israel Deaconess Medical Center, and Boston Children’s Hospital. Each study includes at least a standard attended PSG montage, along with clinical annotations and standardized automated sleep annotations. Data were harmonized using the Brain Imaging Data Structure framework and linked to de-identified demographic and electronic health record data. Results: The HSP spans the full human lifespan, from infancy to individuals over 90 years old, and includes diagnoses across 22 ICD-10 clinical categories. Analysis of the aggregated data demonstrates expected age-related trends in sleep architecture, including reduced total sleep time and N3 sleep and increased sleep fragmentation with advancing age. Standardization procedures successfully addressed technical and scoring variability across sites, creating a consistent and high-quality resource for large-scale analyses. Conclusions: The HSP provides harmonized physiological signals, manual and automated annotations, and quality metadata for 119 234 sleep studies, including 115 129 PSGs and 4105 home sleep apnea tests, enabling reproducible sleep research at an unprecedented scale.

At-home collected salivary Dim Light Melatonin (DLMO) assessments have demonstrated agreement with laboratory-determined DLMO estimates, but key translational gaps remain. We compiled 32 at-home salivary DLMO assessments obtained over four years from a healthy adult male (35– 40 years old) with stable sleep-wake patterns to characterize the magnitude of behavioral, environmental, and pharmacologic influences on DLMO and melatonin profile morphology at the individual level. The series included 10 assessments conducted under standardized dim light conditions to establish a baseline pattern, 21 assessments following controlled contextual manipulations (including exogenous melatonin, ambient and bright-light exposures, blue-blocking glasses, diazepam, behaviorally delayed bedtime scheduling, melatonin-rich foods, magnesium, caffeine, alcohol, late-evening anaerobic exercise), and one seasonal photoperiod comparison. Across a series of 10 baseline assessments, DLMO timing demonstrated high intra-individual stability (range: 7:13 PM–8:43 PM; SD = 28 min), and peak melatonin levels were highly consistent (M = 13.26 pg/mL, SE = 0.46 pg/mL). Within this individual, 5-day magnesium supplementation (3.6 mg/kg per day), caffeine (300 mg), alcohol (6oz of 80-proof), melatonin-rich foods, or intense anaerobic exercise during the DLMO assessment window produced negligible deviation in DLMO timing or changes in peak concentrations. In contrast, continuous bright light (∼1500 lux) during the established pre-onset interval suppressed melatonin production and obscured central circadian phase estimation, while bright light exposure after melatonin onset also produced steep declines in melatonin levels. Diazepam (0.12 mg/kg per day, over a 5-day period) delayed melatonin timing and attenuated melatonin levels. In contrast, Escitalopram (0.24 mg/kg per day, over a 60-day period) elevated baseline melatonin concentrations without obvious alteration of the underlying onset of melatonin secretion at the dosage evaluated. A behaviorally delayed sleep-wake time (i.e., 5 hours for 10 days) resulted in a corresponding delay in melatonin onset consistent with entrainment to the altered bedtime. Exogenous melatonin (0.06 mg/kg) produced supraphysiologic concentrations, which precluded interpretation of the endogenous circadian phase. These findings demonstrate that contextual influence on at-home DLMO assessment may differ substantially in effect magnitude at the individual level of analysis. Discussion focuses on the distinction between higher-impact, first-order threats and low-negligible-impact “second-order” threats to at-home DLMO measurement validity.