LAUSR

Improving positive airway pressure (PAP) adherence is crucial to sleep apnea therapy success. Although behavioral interventions may be deployed to increase PAP adherence, operationalization remains an ongoing clinical challenge. Treatment outcomes may be optimized by forecasting PAP use to identify patients at risk for non-adherence enabling early intervention. We build upon our previous work by leveraging a larger dataset, additional metadata, and new Deep Learning approaches to forecast future PAP adherence. We collected a cohort of N=21,397 subjects with daily PAP usage recorded during 2015-2021. We defined the input to models as the number of minutes the PAP machine was used during each day for the first 30-days. The ground truth was defined as the PAP adherence of the patients at the 3-month, 6- month, and 1-year endpoints. Adherence was calculated based on a 30-day window as ≥4-hours of usage for ≥70% of nights. We evaluated a Deep Neural Network (DNN) model with 10-fold cross- validation to forecast future adherence by leveraging previous daily usage information. Results were compared to a naive method which assumes adherence at each time point equals adherence during the first 30-days. The DNN models predicted adherence with a sensitivity of 90%, 81%, 77% and a specificity of 90%, 81%, 77%, for 3-month, 6-month, and 1-year endpoints, with ROC-AUC values of 0.97, 0.89, and 0.84 respectively. The models converged to ROC-AUC performance >0.90 for the first 90-days within the first 7 to 14-days of PAP use. DNN models demonstrated strong predictive performance for PAP adherence, as defined by the CMS adherence criteria, measured by sensitivity, specificity, and overall ROC-AUC results at clinically relevant 90-day, 6-month, and 1-year timepoints. AI approaches show promise as early predictors of the likelihood to meet key therapy utilization thresholds within the first 1-2 weeks of therapy, enabling early PAP intervention or transition to alternative therapies. AASM Foundation SRA205-SR-19