Using a Machine Learning–Based EHR Model to Enhance Screening for Obstructive Sleep Apnea: Nathanael Hwang

WATCH TIME: 3 minutes | Captions are auto-generated and may contain errors.

"We are envisioning basically a world where a patient can walk into the clinic and, even if it's not sleep related, the doctor will already be alerted that this person may need to be referred to the sleep clinic because they're at risk for OSA."

Obstructive sleep apnea (OSA) affects an estimated 936 million adults worldwide, yet approximately 80% of cases remain undiagnosed, leaving many patients without access to treatment for a disorder linked to cardiovascular and metabolic disease.¹ Current screening relies largely on the STOP-BANG questionnaire, which requires clinicians to actively query patients about symptoms and risk factors. A machine learning model built from electronic health record (EHR) data presented at the 2026 SLEEP Annual Meeting, held June 14-17 in Baltimore, may offer a passive, more scalable alternative for identifying at-risk patients.²

Investigators from Kaiser Permanente Southern California and Beth Israel Deaconess Medical Center developed the model using a cohort of 285,292 adults who underwent diagnostic sleep testing between 2016 and 2025, with candidate features spanning demographics, comorbidities, vitals, and laboratory values.² Two classification models were built, one predicting any OSA and another predicting moderate-to-severe disease. All told, results from a random forest model achieved the strongest discrimination for detecting any OSA (ROC-AUC, 0.84; sensitivity, 0.96) and for moderate-to-severe OSA (ROC-AUC, 0.82), outperforming STOP-BANG on both metrics. A reduced "minimal model" using only age, sex, BMI, and race/ethnicity retained comparable performance, and external validation supported the models' generalizability.

During the meeting, NeurologyLive^® spoke with lead author of the study Nathanael Hwang, a research intern at Beth Israel Deaconess Medical Center, who works on applying machine learning to sleep medicine. Throughout the interview, Hwang explained the rationale for developing an EHR-based screening model to close the OSA diagnostic gap, describing how it passively identifies at-risk patients within clinical workflows rather than relying on clinician-administered questionnaires. Finally, he addressed the model's current limitations and next steps as it moves toward real-world implementation as a clinical pilot at Kaiser Permanente.

Click here for more coverage of SLEEP 2026.

REFERENCES
1. Benjafield AV, Ayas NT, Eastwood PR, et al. Estimation of the global prevalence and burden of obstructive sleep apnoea: a literature-based analysis. Lancet Respir Med. 2019;7(8):687-698. doi:10.1016/S2213-2600(19)30198-5
2. Hwang N, Westover MB, Mazzotti D, et al. Applying machine learning to predict obstructive sleep apnea using electronic health records. Presented at: 2026 SLEEP Annual Meeting; June 14-17; Baltimore, Maryland. Abstract 0509.