1078 Machine Learning to Predict PAP Adherence and Compliance in Tele-Health Management

Abstract

Introduction: CPAP is the optimal treatment for obstructive sleep apnea, but is limited by low adherence. Fairview's sleep program actively tracks PAP usage and outcomes and employs tele-health coaching to improve adherence. This approach has achieved 6-month adherence rates of 71%. However, this protocol is applied uniformly and is labor-intensive, reaching some patients belatedly and contacting others unnecessarily. Machine learning can facilitate efficient contact strategies through early identification of therapy trajectories. Methods: We constructed prediction models for compliance (≥4 hours/day ≥70% of days) and adherence (average ≥4 hours/day) at days 151-180. Compliance and adherence during days 1-30 (standard) were compared to two prediction intervals chosen based on current tele-health contact protocols: days 1-13 (D13) and 1-30 (D30). Patient data used for prediction included demographics, health information, questionnaires, and daily PAP metrics. A feature selection algorithm was used to find the best features to improve prediction accuracy. Support Vector Machines and Random Forest learning methods were used. Results: Demographics for the group (N=3588) included mean (SD) age of 53.3 (12.9) years, BMI 36.5 (8.0) kg/ m2, and baseline AHI of 38.3 (30.4) events/hour; 68% were male. For the compliance outcome, standard 30-day compliance versus D13 and D30 had PPVs of 0.716, 0.732, and 0.765, respectively while NPVs were 0.718, 0.694, and 0.707, respectively. For the adherence outcome, standard 30-day adherence versus D13 and D30 had PPVs of 0.781, 0.782, and 0.806, respectively while NPVs were 0.675, 0.697, and 0.686, respectively. Accuracy improved from standard to D13 to D30 for both outcomes. Baseline characteristics and daily CPAP metrics contributed positively to performance. Conclusion: We demonstrate an efficient method using machine learning to predict long-term PAP adherence based on early PAP usage, with comparable prediction over two weeks earlier. Initial machine learning efforts improved performance over standard predictors. Further refinement will improve performance using different weighting/ cost functions, advanced learning methods, and non-dichotomous models. This type of model is easily localized and implementable in most EHRs. Developing predictors of long-term adherence will allow for tailoring of follow-up and timely care.