Utilizing machine learning based on multi-modal data to predict PAP adherence in patients with OSA

Dennis Hwang1 • Sam Rusk2 • Chris Fernandez, MS2 • Yoav Nygate, MS2 • Tom Vanasse, PhD3 • Jan Wodnicki3 • Anupamjeet Sekhon1 • Jiaxiao Shi1 • Rui Yan2

Introduction

Predicting future Positive Airway Pressure (PAP) adherence may support treatment decisions and timing of interventions in patients with obstructive sleep apnea (OSA). We developed machine learning (ML) prediction algorithms based on daily PAP metrics and diagnostic sleep study data to predict adherence at 3-months and 1-year.

Methods

ML models (eXtreme Gradient Boost) were trained on a dataset of 37,076 patients (Kaiser Permanente, Southern California) that comprised of daily PAP data (“usage” [minutes/night] and “leak”) and diagnostic sleep study data (Polysomnography and Home Sleep Apnea Tests). Models were trained to predict PAP adherence (≥70% days, ≥4 hours) at 3-months (i.e., 61-90 days) and 1-year (i.e., 330-360 days) after PAP initiation. Developed algorithms were based on: (a) Usage; (b) Leak; (c) Diagnostic sleep study data (i.e., AHI, oximetry metrics, sleep-wake metrics); (d) combination of those data types. For the models utilizing daily PAP data, different input days were applied (i.e., 7, 30, 90, etc.) to the training.

Results

Models based on daily usage alone demonstrated excellent predictive performance with relatively short input days (ROC-AUC 0.86 and 0.96 in predicting 3-month adherence utilizing 7 and 30 input days; 0.74, 0.82, 0.88 in pre- dicting 1-year adherence utilizing 7, 30, and 90 input days). Models based on daily leak demonstrated more modest performance (0.69 and 0.83 predicting 3-month adherence with 7 and 30 input days; 0.60, 0.70, 0.80 predicting 1-year adherence with 7, 30, 90 input days.) Models based only on diagnostic sleep study data (without PAP data) demonstrated only small but similar 3-month and 1-year predictive accuracy (0.57 and 0.56, respectively). The addition of leak and sleep study data to usage did not significantly improve performance.

Conclusion

ML algorithms based on daily usage data early after therapy initiation can accurately predict future adherence to at least 1 year after therapy, enabling decision support after starting PAP. Algorithms using sleep study data could enable treatment decisions prior to prescribing PAP, but performance was modest. Whether the addition of electronic health record or direct ML processing of raw sleep study data could improve predictive accuracy requires further exploration.