The Sleep Disorders Channel is supported through educational grants from Alkermes and Takeda.
VJNeurology is an independent medical education platform. Channel supporters have no influence over the production of content.
SLEEP 2026 | A novel OSA prediction model using machine learning
Nathanael Hwang, Beth Israel Deaconess Medical Center, Boston, MA, discusses the development of a machine learning model designed to improve screening for obstructive sleep apnea (OSA) using routinely collected electronic health record data. He highlights the performance of both comprehensive and simplified models, their validation against existing tools such as STOP-BANG, and the potential for automated, scalable OSA detection in clinical practice. This interview took place at the 40th annual meeting of the Associated Professional Sleep Societies (APSS) in Baltimore, MD.
These works are owned by Magdalen Medical Publishing (MMP) and are protected by copyright laws and treaties around the world. All rights are reserved.
Transcript
So we obviously know that OSA is a very prevalent disorder, right? It affects 1 billion people worldwide, but how come only 80% of them remain undiagnosed? And so we wanted to fill this gap by using a machine learning model to basically screen for OSA. And currently, the paradigm for screening for OSA is they’re using STOP-BANG. However, clinicians frequently report that STOP-BANG is very inconvenient, takes a lot of effort, and they just don’t want to do it anymore...
So we obviously know that OSA is a very prevalent disorder, right? It affects 1 billion people worldwide, but how come only 80% of them remain undiagnosed? And so we wanted to fill this gap by using a machine learning model to basically screen for OSA. And currently, the paradigm for screening for OSA is they’re using STOP-BANG. However, clinicians frequently report that STOP-BANG is very inconvenient, takes a lot of effort, and they just don’t want to do it anymore. And so what if we can make a model that just passively runs in the background and alerts the clinician when someone’s at risk for OSA? And so that’s why we use electronic health records because it is so broadly available and everyone has electronic health records. So we can just make a model that can run in the background and alert the physician. So our model, we developed two of them, can predict both OSA and OSA severity. So any OSA or moderate to severe OSA. And our model uses labs, comorbidities, demographics and vitals all in a 106 feature model; however, we wanted to simplify things so we also created a minimal four feature model that only includes age, sex, BMI, and race ethnicity and we found that the minimal model actually performs almost as well as the full model. And so, but we consider this actually a very encouraging finding because the minimal model is significantly easier to implement because it only takes four features. And so we’re thinking that the minimal model should be our primary model to implement in all sorts of healthcare systems. But for healthcare systems who prefer accuracy and a better performance, they would prefer to use the full model for their studies. We validated this model internally and externally at a site called KUMC. And we also compared it to STOP-BANG, where it outperformed STOP-BANG statistically significantly in both AUC and net benefit. This model, we believe, has the potential to basically replace STOP-BANG as the primary OSA screener just because of how low effort and how scalable it is.
This transcript is AI-generated. While we strive for accuracy, please verify this copy with the video.
Read more...
