O031 Independent validation of a physiology-based model to predict oral appliance treatment outcomes

BACKGROUND: Oral appliance therapy (OAT) is a well-tolerated treatment for obstructive sleep apnoea (OSA). However, efficacy varies. On average, OAT reduces OSA severity by ~50% leaving a substantial proportion of OAT patients either undertreated or untreated. Thus, a major clinical priority is to develop clinically feasible prediction tools to accurately predict which patients will respond to OAT and vice versa. Accordingly, we developed a model to estimate OSA endotypes and have adapted it to predict OAT outcomes. Here we used our recently developed physiology-based OAT prediction model to predict OAT responses in an independent validation sample. METHODS: 91 people with OSA who were prescribed OAT by their treating physician had an in-laboratory diagnostic study followed by an OAT efficacy polysomnography after ≥4 weeks acclimatisation. 7 polysomnography variables from the diagnostic study plus age and body mass index were included in our machine learning based model designed to predict OAT response. Prediction performance characterises according to standard apnoea/hypopnoea index (AHI) definitions were assessed. RESULTS: Mean accuracy of the model to predict OAT responders vs. non-responders in this independent validation sample was high across different OSA treatment outcome definitions ranging from 70% (>50% reduction in AHI), 79% (AHI<10 events/h) to 95% (AHI<5 events/h). Corresponding specificity was also high at 87%, 97% and 99%, respectively. DISCUSSION: These new findings provide additional support for the use routinely collected sleep study and clinical data with machine learning based approaches underpinned by OSA endotype concepts to help predict OAT treatment outcomes for people with OSA.