Data-Driven Phenotyping of Obstructive Sleep Apnea and Sex-Specific Cardiovascular Consequences in a Multi-Modal EHR Cohort

About This Grant

ABSTRACT Obstructive Sleep Apnea (OSA) is a widespread yet underdiagnosed sleep disorder that significantly contributes to the development and progression of cardiovascular disease (CVD), including heart failure (HF), atrial fibrillation (AFib), and coronary artery disease (CAD). Standard clinical assessment relies heavily on the ApneaHypopnea Index (AHI), a single-dimensional metric that inadequately captures the physiological diversity of OSA and fails to explain substantial inter-individual differences in cardiovascular outcomes. This oversimplification leads to poor risk stratification and missed opportunities for timely intervention, particularly in groups where diagnostic sensitivity is lower and cardiovascular manifestations may be distinct. Emerging evidence highlights that physiologic endotypes, such as hypoxic burden, ventilatory control instability (loop gain), and arousal burden, better reflect OSA pathophysiology and associated cardiovascular risk. However, these dimensions have not been studied at scale in large patient populations with longitudinal clinical outcomes. We propose to address this gap by integrating high-dimensional physiological and clinical data from the Mass General Brigham health system, which includes over 144,000 patients with structured electronic health records (EHRs) and ~17,000 individuals with detailed polysomnography metrics. In Aim 1, we will derive physiologically informed OSA phenotypes using unsupervised machine learning techniques (e.g., manifold embedding, LASSO-regularized representation learning) that incorporate hypoxia indices, arousal dynamics, loop gain estimates, and clinical features such as comorbidities, medications, and lab trajectories. These phenotypes will be evaluated for differential cardiovascular risk (e.g., HF, AFib, CAD) and validated across age, sex, BMI, and racial/ethnic strata to identify subgroup-specific risk patterns invisible to conventional AHI-based assessment. In Aim 2, we will develop predictive models to identify patients likely to have undiagnosed OSA using structured EHR features. We will apply pseudo-negative learning strategies to avoid misclassification and build robust, scalable classifiers that flag high-risk individuals lacking a formal diagnosis. We will then examine the cardiovascular consequences of delayed diagnosis by comparing comorbidity trajectories and disease burden between diagnosed and model-identified undiagnosed cases, stratified by demographic variables. By combining rich physiologic signal processing with scalable EHR modeling, we will redefine OSA classification and improve risk prediction for CVD. Our work moves beyond AHI to reveal mechanistic heterogeneity and unmask high-risk patients overlooked by current diagnostics. This precision health approach will support early identification, targeted surveillance, and personalized treatment strategies across the population.