Abstract
INTRODUCTION: Obstructive sleep apnoea (OSA) is a heterogeneous disorder characterised by repeated upper-airway collapse during sleep. Given the major role of obesity in OSA pathogenesis, weight loss is commonly a recommended therapy, though its effectiveness can be highly variable. As weight loss therapies often require considerable time, effort, or cost, better tools are needed to identify likely responders and enable timely, targeted treatment decisions. METHODS: 149 adults (age[mean ± SD] = 49 ± 10) from five weight loss studies were included, all of whom achieved >5% reduction in body-mass index (BMI) following intervention. Baseline polysomnographic (PSG) metrics from clinical sleep studies were used as predictors. Two machine learning models, XGBoost and a deep neural network (DNN), were implemented to predict 1) a > 50% reduction in apnoea–hypopnoea index (AHI), and 2) a categorical improvement in OSA severity (e.g., from moderate to mild). Models were trained on ~80% of the sample and tested on the remaining 20%. RESULTS: For predicting a > 50% reduction in AHI, both models showed moderate accuracy (XGBoost: 69%; DNN: 66%), with higher negative (79%/73%) than positive (60%/57%) predictive values. For predicting categorical improvement, overall accuracy was lower (59%/52%), though positive predictive values remained stronger (70%/60% vs. 53%/47%). DISCUSSION: Models using routine PSG data were modestly effective at predicting OSA response to weight loss, particularly in identifying non-responders. While this suggests potential utility for triaging patients, performance was more limited than models for other OSA treatments. Consequently, there may be need for additional features such as behavioural, metabolic, or phenotypic markers to improve future prediction.