O002 Machine learning applied to oximetry to detect paediatric sleep apnoea

Abstract Introduction Polysomnography(PSG) is the gold-standard test for diagnosing paediatric obstructive sleep apnoea, but resource-intensiveness limits availability. Although oximetry testing is widely available, manual scoring methods such as McGill scoring have poor sensitivity, and decreased positive predictive value(PPV) for those with complex medical comorbidities. We hypothesised that computerised oximetry analysis could overcome these limitations. We developed and tested a novel support vector classifier(SVC) machine learning algorithm, with a component for discarding of likely movement/wake artefact, that classifies the overnight oximetry as being either low risk (predicted apnoea hypopnoea index[AHI] <5/hour) or high risk(≥5/hour). Methods Oxygen saturation(SpO2) and pulse rate(PR) data were extracted from 3476 PSGs performed at Queensland Children’s Hospital (QCH). 90% were selected for model training, with 10% reserved for final testing. A rules-based filter discarded periods of SpO2 < 60% or PR > 225bpm or < 40bpm. Thirteen SpO2/PR features were used for SVC model training; model performance was then evaluated on held-out data from QCH and other sources. Results Using oximetry data from the reserved 344 QCH PSGs, the algorithm showed a PPV of 93%, negative predictive value of 90%, specificity of 99%, and sensitivity of 65%. Using oximetry data from the Childhood Adenotonsillectomy Trial, Paediatric Adenotonsillectomy Trial for Snoring and British Columbia Children’s Hospital PSG datasets (n = 1215, n = 721 and n = 2895 respectively) model performance showed specificities of 99%, 98% and 84% respectively, with sensitivities of 35-62%. Conclusion A SVC algorithm can be used to classify oximetry into likely AHI ≥5 with a high degree of certainty, for children with and without complex comorbidities.