This review provides a current perspective of the classes of tools at hand, along with a sense of their relative strengths and areas for further improvement.

Introduction Electroencephalogram (EEG) provides clinically relevant information for personalized patient health evaluation and comprehensive assessment of sleep. EEG-based indices have been associated with neurodegenerative conditions, psychiatric disorders, and metabolic and cardiovascular disease, and hold promise as a biomarker for brain health. Methods A deep neural network (DNN) model was trained to predict the age of patients using raw EEG signals recorded during clinical polysomnography (PSG). The DNN was trained on N=126,241 PSGs, validated on N=6,638, and tested on a holdout set of N=1,172. The holdout dataset included several categories of patient demographic and diagnostic parameters, allowing us to examine the association between brain age and a variety of medical conditions. Brain age was assessed by subtracting the individual’s chronological brain age from their EEG-predicted brain age (Brain Age Index; BAI), and then taking the absolute value of this variable (Absolute Brain Age Index; ABAI). We then constructed two regression models to test the relationship between BAI/ABAI and the following list of patient parameters: sex, BMI, depression, alcohol/drug problems, memory/concentration problems, epilepsy/seizures, diabetes, stroke, severe excessive daytime sleepiness (e.g., Epworth Sleepiness Scale ≥ 16; EDS), apnea-hypopnea index (AHI), arousal index (ArI), and sleep efficiency (SE). Results The DNN brain age model produced a mean absolute error of 4.604 and a Pearson’s r value of 0.933 which surpass the performance of prior research. In our regression analyses, we found a statistically significant relationship between the ABAI and: epilepsy and seizure disorders, stroke, elevated AHI, elevated ArI, and low SE (all p<0.05). This demonstrates these health conditions are associated with deviations of one’s predicted brain age from their chronological brain age. We also found patients with diabetes, depression, severe EDS, hypertension, and/or memory and concentration problems showed, on average, an elevated BAI compared to the healthy population sample (all p<0.05). Conclusion We show DNNs can accurately predict the brain age of healthy patients based on their raw, PSG derived, EEG recordings. Furthermore, we reveal indices, such as BAI and ABAI, display unique characteristics within different diseased populations, highlighting their potential value as novel diagnostic biomarker and potential “vital sign” of brain health. Support (if any):

Introduction Clinical management of CPAP adherence remains an ongoing challenge. Behavioral and technical interventions such as patient outreach, coaching, troubleshooting, and resupply may be deployed to positively impact adherence. Previous authors have described adherence phenotypes that retrospectively categorize patients by discrete usage patterns. We design an AI model that predictively categorizes patients into previously studied adherence phenotypes and analyzes the statistical significance and effect size of several types of interventions on subsequent CPAP adherence. Methods We collected a cross-sectional cohort of subjects (N = 13,917) with 455 days of daily CPAP usage data acquired. Patient outreach notes and resupply data were temporally synchronized with daily CPAP usage. Each 30-days of usage was categorized into one of four adherence phenotypes as defined by Aloia et al. (2008) including Good Users, Variable Users, Occasional Attempters, and Non-Users. Cross-validation was used to train and evaluate a Recurrent Neural Network model for predicting future adherence phenotypes based on the dynamics of prior usage patterns. Two-sided 95% bootstrap confidence intervals and Cohen’s d statistic were used to analyze the significance and effect size of changes in usage behavior 30-days before and after administration of several resupply interventions. Results The AI model predicted the next 30-day adherence phenotype with an average of 90% sensitivity, 96% specificity, 95% accuracy, and 0.83 Cohen’s Kappa. The AI model predicted the number of days of CPAP non-use, use under 4-hours, and use over 4-hours for the next 30-days with OLS Regression R-squared values of 0.94, 0.88, and 0.95 compared to ground truth. Ten resupply interventions were associated with statistically significant increases in adherence, and ranked by adherence effect size using Cohen’s d. The most impactful were new cushions or masks, with a mean post-intervention CPAP adherence increase of 7-14% observed in Variable User, Occasional Attempter, and Non-User groups. Conclusion The AI model applied past CPAP usage data to predict future adherence phenotypes and usage with high sensitivity and specificity. We identified resupply interventions that were associated with significant increases in adherence for struggling patients. This work demonstrates a novel application for AI to aid clinicians in maintaining CPAP adherence. Support

Introduction Despite an appreciable rise in sleep wellness and sleep medicine A.I. research publications, public data corpuses, institutional support, and health A.I. research funding opportunities, the availability of controlled-retrospective, hybrid-retrospective-prospective, and prospective-RCT quality clinical validation study evidence is limited with respect to their potential clinical impact. Furthermore, only a few practical examples of A.I. technologies are validated, in use today clinically, and widely adopted, to assist in sleep diagnoses and treatment. In this study, we contribute to this growing body of clinical A.I. validation evidence and experimental design methodologies with an interoperable A.I. scoring engine in Adult and Pediatric populations. Methods Stratified random sampling with proportionate allocation was applied to a database of N>10,000 retrospective diagnostic clinical polysomnography (PSG), selected by evidence grading standards, with controls applied for OSA severity, diagnoses; sleep, psychiatric, neurologic, neurodevelopmental, cardiac, pulmonary, metabolic disorders, medications; benzodiazepines, antidepressants, stimulants, opiates, sleep aids, demographic groups of interest; sex, adult age, pediatric age, BMI, weight, height, and patient-reported sleepiness, to establish representative N=100 Adult and N=100 Pediatric samples. Double Blinded scoring was prospectively collected for each sample by 3 experienced RPSGT certified sleep technologists randomized from a pool of 9 scorers. Sensitivity (PA), Specificity (NA), Accuracy (OA), Kappa (K), and 95% Bootstrap CI’s are presented for sleep stages, OSA/CSA, hypopnea 3%/4%, arousals, limb movements, Cheyenne-Stokes respiration, periodic breathing, atrial fibrillation, and other events, and normative, mild, moderate, and severe OSA categories for global-AHI and REM-AHI. Results for Sleep Staging and OSA Severity Diagnostic Accuracy are summarized. Results A.I. scoring performance meet but in most cases exceeded initial clinical validation study (N=72 Adults, 2017) PA, NA, OA, K point-estimates and confidence-interval results for the 26 event types and 8 AHI-categories evaluated. The Adult sample showed 87%/94% Sensitivity/Specificity across all stages (Wake/N1/N2/N3/REM) and 94%/96% Sensitivity/Specificity for AHI>=15. The Pediatric sample showed 87%/93% Sensitivity/Specificity staging, 89%/98% Sensitivity/Specificity AHI>=15. Observed Accuracy was >90% for Adults and Pediatrics all 26 events and 7 AHI-categories analyzed, except REM-AHI>=5 (85%/82% Adults/Pediatrics). Conclusion We provide clinical validation evidence that demonstrates interoperable A.I. scoring performance in representative Adult and Pediatric patient clinical PSG samples when compared to prospective, double-blind scoring panel. Support (if any):

Introduction Current approach to processing polysomnography is labor intensive and produces metrics that are poor at identifying obstructive sleep apnea (OSA) phenotypes necessary to enhance personalized care. We describe our approach to utilize Dynamic Phenotype Learning (DPL) as an innovative machine learning technique to identify OSA subtypes that can better predict clinical risk and success with therapies. Methods This study is a collaboration between Kaiser Permanente Southern California (KPSC), a large integrated health system, and EnsoData Research, which specializes in applied A.I. analysis of physiologic waveforms. KPSC sleep medicine compiled a database of N=5,368–234,250 subjects that include Types I, II, III, or IV sleep study data, daily PAP data, patient reported data, and comprehensive electronic health record information, with present research applications to study the relationship between OSA and PAP adherence with cardiovascular outcomes, health economic impacts, novel coronavirus (COVID-19) outcomes, and predictive PAP adherence and OSA severity clinical decision tools. DPL is a machine learning method for studying known and new biomarkers and care-pathway indices, including personalized screening, diagnostic, treatment, adherence, and outcomes predictors, that can be rooted in physiologic data. DPL processes waveform signal data without scoring, annotations, or expert synthesis, by applying a novel machine learning mechanism that blurs supervised and unsupervised deep learning paradigms, to find relationships between physiome dynamics expressed in waveforms and phenotypes and endotypes of interest. Results We demonstrate DPL method with an illustrative study on known indices, to explain its ability to (1) lift theoretical-empirical predictive accuracy ceilings and (b) reduce several sources of bias and variance. We show DPL exceeds the ROC-AUC and PRC-AUC of equivalent deep learning models in N=30,000 Report-Demographic (ODI, PLMSI, Weight), Scoring (REM, OSA), and Waveform (EEG, PPG) datasets respectively to predict AHI, TST, Brain Age, and OSA-Insomnia. We present our current collaboration advancing DPL to identify specific phenotypes that better predict: (a) cardiovascular risk; (b) neurocognitive outcomes; (c) response to PAP and alternative therapies. Conclusion DPL methods are being applied to large and comprehensive patient dataset to identify phenotypic indices and biomarkers with potential to take us beyond the AHI, and uncover relationships between OSA sub-types, treatments, and health outcomes. Support (if any):