1104 Personalisation of Breathing Comfort for CPAP Users Through a Machine-Learning Algorithm
Sarah Pongrass,Nathan Liu,Chinh Nguyen,Vinh Tran,Hussein Emami,Jeff Armistead,Alison Wimms
DOI: https://doi.org/10.1093/sleep/zsae067.01104
IF: 6.313
2024-04-20
SLEEP
Abstract:Abstract Introduction Many factors impact CPAP adherence, some common issues include trouble falling asleep, difficulty breathing against pressure, and being disturbed by high pressures. Various settings to address breathing comfort have been developed, such as ramp time, reducing pressure during exhalation, and APAP response rate, though evidence of their impact is minimal. We sought to develop a machine-learning model to recommend personalised combinations of comfort settings at setup with the goal of improving CPAP adherence in patients newly starting CPAP therapy. Methods Data sets from 718,421 anonymous OSA patients (myAir, ResMed) were used to create the machine-learning model (MLM). The model was developed using a causal inference approach to control for bias from observational data and utilised features such as age, sex, diagnostic AHI, mask type and start pressure to predict optimal comfort settings for various patient types. These comfort settings can then be applied on an individual level. A proof-of-concept, double-blinded, randomised controlled study was conducted with people newly diagnosed with OSA (AHI≥5) and indicated for CPAP therapy attending a sleep clinic. The primary outcome was treatment continuance after 30 days comparing those set up on MLM vs default comfort settings (control). Results 58 patients (26% female, mean age 49.9±14.9, 46.6% non-white, mean diagnostic AHI 43.6±30.4) were included in the study analysis. At 30 days the MLM group had greater treatment continuance vs control (96% vs 80%) though not statistically significant. All other endpoints were also not statistically significant. 54% of MLM group had zero setting changes vs 37% control. In this subgroup, MLM showed increased average daily usage (+46.9min), number of days used (+2.9days), and number reaching CMS compliance (+11%), however these endpoints did not reach statistical significance. Conclusion This program developed a machine-learning model to predict optimal comfort setting combinations for individual CPAP patients to improve CPAP adherence. The proof-of-concept study did not reveal any harm to patients assigned personalised settings and showed some trends towards positive outcomes. Personalising comfort settings may benefit new CPAP patients, however a larger study is required to fully validate the machine-learning model. Support (if any) Funded by ResMed
neurosciences,clinical neurology