O067 Personalisation of comfort settings on CPAP devices using machine learning

Abstract Introduction Machine learning (ML) technology provides a novel mechanism to identify patterns in large data sets, draw inferences and develop algorithms to simplify complex problems. Sleep apnea is a uniquely positioned health condition in the ML space due to the large amounts of data collected by diagnostic and CPAP devices. We developed, validated, and tested a ML model which recommends personalised combinations of CPAP comfort settings for new users. Methods Data sets from 718,421 anonymous OSA patients were used to develop the ML models. Our ML method adopted the Doubly-Robust Learner to estimate the effect of comfort settings combinations on CPAP compliance. A multilayer perceptron model has been developed at each stage of the DR-learner method to provide the propensity model and the outcome models. A clinical trial was then conducted to measure the impact of the model on patient outcomes. Results Development and validation of the ML models, utilising observational data, demonstrated a promising capability to enhance CPAP compliance through suggesting personalised comfort settings. A double-blinded, RCT was conducted on 58 patients with newly diagnosed OSA. 30 days after setup, more patients in the ML model group were still using CPAP compared with the control group (96% vs 80%, p=ns). Discussion We developed and validated a unique machine learning algorithm aimed at providing personalised comfort CPAP settings in an area where limited guidance currently exists. This pilot study has shown that the ML model may provide benefit to new CPAP patients.