ABSTRACT
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 OSAInsomnia. 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.