O043 Performance of a machine-learning based automated oximetry algorithm for detection of sleep apnoea in children

Abstract Background In-laboratory polysomnography (PSG) remains the gold-standard method for the detection of significant paediatric sleep-disordered breathing, although it is resource-intensive. We therefore trained, using data from 2380 diagnostic PSGs, a support vector regression (SVR) algorithm to predict apnoea hypopnea index (AHI) using oximetry data alone, for potential use as a pre-PSG screening/triage tool. This study aimed to test the algorithm’s performance in detecting potentially significant sleep apnoea, defined as AHI ≥5. Methods We extracted oximetry data (oxygen saturation and heart rate, down-sampled to 1Hz) from 300 PSGs performed in 2018 in a single Australian paediatric centre. De-identified demographic and clinical information was extracted from medical records. Oximetry data was then evaluated by the SVR algorithm, which utilised 19 heart rate and oxygen saturation-derived features to predict AHI. The SVR-predicted AHI was compared to gold-standard AHI determined by original PSG, with AHIs undergoing binary classification into AHI <5 or ≥5. Results Of 297 patients (median age 7.4 years), 26% had a significant neuromuscular, neurologic or genetic disorder diagnosed, 15% had undergone previous adenoidectomy+/-tonsillectomy and 14% were obese. On PSG 102 (34%) had AHI ≥5. Sensitivity and specificity of the SVR algorithm for detecting this was 0.75 and 0.76 respectively; positive predictive value was 0.62 and negative predictive value was 0.85. Conclusion This demonstration of a machine learning algorithm classifying oximetry traces from a real-world cohort of a diverse group of children undergoing evaluation for possible sleep apnoea demonstrates the automated tool may be useful in helping with PSG waitlist prioritisation.