Robust forecasting of sedentary bouts in chronic pelvic pain disorders for on-device learning and real-time deployment.

Abstract

Reducing sedentary behavior through personalized digital interventions holds particular promise for individuals with chronic pelvic pain disorders (CPPDs), who face unique barriers to physical activity. We present a self-contained, missing data-resilient framework for real-time forecasting of a physical activity score (PAS) using wearable Fitbit data from 134 females with CPPDs. Comparing online and offline learning approaches, we demonstrate that models leveraging recent activity and daily recurring patterns perform the best. When applied to 15-min sedentary bouts (SBs), the PAS forecasts support timely alerts, yielding approximately one true alert per day vs 0.6 false alerts at a conservative operating point. By integrating real-time imputation, the system supports forecasting of SBs for potential use in just-in-time adaptive interventions. Our work demonstrates a privacy-preserving, scalable pathway for integrating precision forecasting into just-in-time adaptive interventions, laying the groundwork for more equitable and effective digital health solutions for women with CPPDs.