0464 Deep Learning to Predict PAP Adherence in Obstructive Sleep Apnea

Autor: Samuel Rusk, Yoav Nygate, Chris Fernandez, Jiaxiao M Shi, Jessica Arguelles, Matthew T Klimper, Nathaniel F Watson, Robert Stretch, Michelle Zeidler, Anupamjeet Sekhon, Kendra Becker, Joseph Kim, Dennis Hwang
Rok vydání: 2023
Předmět:
Zdroj: SLEEP. 46:A206-A206
ISSN: 1550-9109
0161-8105
Popis: Introduction Machine Learning (ML) algorithms to predict Positive Airway Pressure (PAP) adherence may support personalized clinical management. Models were developed to predict adherence at various time-points after PAP initiation and in moving time windows. Methods Deep neural network (DNN) models were trained utilizing daily PAP data (Kaiser Permanente, Southern California). The DNN was evaluated with 10-fold cross-validation on N=21,397 patients. Algorithms developed included (a) Models 1 and 2 which utilized early usage to predict adherence at 90-days and 1-year respectively, and (b) Model 3 which utilized 14 and 30-day moving windows to predict subsequent usage. Regression analyses compared ML and Naïve (i.e., future use equals previous use) predictions versus Actual adherence. Results Model 1 predicted “% days without usage” for first 90-days based on first 7, 14, 21, 30-days of input and at 1-year (90-day window) based on first 30, 60, 90, 180-days of input. ML was superior to Naïve in predicting adherence [R 2 for ML versus Naïve compared to Actuals for different input days— 0.495-vs-0.193; 0.660-vs-0.465; 0.748-vs-0.607; 0.828-vs-0.735 at 90-days and 0.362-vs-0.104; 0.463-vs- 0.247; 0.513-vs-0.339; 0.680-vs-0.547 at 1-year; all p< 0.05]. Model 2 predicted “hours/night” of use—ML did not outperform the Naïve prediction with similar R 2 ; however, when ML predicted < 3 hours/night, nearly all patients had “no significant usage” at 1-year (comparatively, the naïve model had no differentiating threshold to predict this outcome.) Model 3 utilized different windows of PAP usage to predict subsequent usage. ML predictive accuracy was similar using 14 or 30-days of input [R 2 for ML vs. Actuals in predicting 7, 14, and 30-day “% days used ≥4 hours” were 0.687, 0.701, 0.699 using 14- days input and 0.582, 0.702, 0.77 using 30-days input; all p< 0.05.] Conclusion ML algorithms based on PAP usage can predict future adherence, potentially supporting personalized treatment decisions and pre-emptive interventions when upcoming non-adherence is predicted. Support (if any) AASM Foundation SRA205-SR-19
Databáze: OpenAIRE
Externí odkaz: