EHR-Based Machine Learning Models Outperform STOP-BANG for OSA Screening: Nathanael Hwang

Obstructive sleep apnea (OSA) affects an estimated 1 billion people worldwide, yet roughly 80% of cases remain undiagnosed, leaving patients at risk for cardiometabolic and other complications, Nathanael Hwang, a research intern at Beth Israel Deaconess Medical Center, explained at SLEEP 2026.

Current screening approaches, such as the widely used STOP-BANG questionnaire, require clinician time and patient engagement and do not fully leverage routinely collected clinical data. To address these gaps, investigators developed and validated machine learning models based on electronic health record (EHR) data to identify patients at risk for OSA in a scalable, low-effort manner.

Using a cohort of approximately 265,000 patients who underwent diagnostic sleep testing, the team trained 2 models, 1 to predict any OSA and another focused on moderate-to-severe OSA to better stratify disease severity. The primary models incorporated 106 EHR features, including comorbidities, laboratory values, vital signs, and demographic information, and demonstrated strong predictive performance.

Recognizing that highly complex models can be difficult to implement, the researchers also developed a “minimal model” that relies on just 4 routinely available variables: age, sex, body mass index, and race/ethnicity. Hwang noted that the minimal model performed nearly as well as the full-feature models, with both models significantly outperforming STOP-BANG with respect to area under the receiver operating characteristic curve (AUC) and net clinical benefit.

The models were externally validated, with performance maintained in independent datasets, supporting their generalizability. Kaiser Permanente is now piloting integration of these tools into primary care and perioperative workflows. Embedded within the EHR, the models continuously and passively synthesize existing patient data to generate OSA risk scores and alerts, enabling clinicians to identify high-risk individuals even when they present for non–sleep-related complaints.

Hwang highlighted that future work will assess model performance and clinical impact in real-world practice, beyond cohorts already referred for diagnostic sleep testing. If successful, this EHR-based machine learning approach could replace traditional questionnaires as the primary OSA screener, reduce underdiagnosis, and streamline referrals to sleep medicine, all while adding virtually no burden to clinicians or patients.