Nonresponse adjustment using clinical and perioperative patient characteristics is critical for understanding post-discharge opioid consumption

BackgroundPost-discharge opioid consumption is an important source of data in guiding appropriate opioid prescribing guidelines, but its collection is tedious and requires significant resources. Furthermore, the reliability of post-discharge opioid consumption surveys is unclear. Our group developed an automated short messaging service (SMS)-to-web survey for collecting this data from patients. In this study, we assessed its effectiveness in estimating opioid consumption by performing causal adjustment and comparison to a phone-based survey as reference. MethodsPatients who underwent surgical procedures at our institution from 2019-2020 were sent an SMS message with a link to a secure web survey to quantify opioids consumed after discharge. Several patient factors extracted from the EHR were tested for association with survey response. Following targeted learning (TL) nonresponse adjustment using these EHR-based factors, opioid consumption survey results were compared to a prior telephone-based survey at our institution as a reference. Results6,553 patients were included. Opioid consumption was measured in 2,883 (44%), including 1,342 (20.5%) through survey response. Characteristics associated with inability to measure opioid consumption included age, length of stay, race, tobacco use, and missing preoperative assessment. Among the top 10 procedures by volume, EHR-based TL nonresponse bias adjustment corrected the median opioid consumption reported by an average of 57%, and corrected the 75th percentile of reported consumption by an average of 11%. This brought median estimates for 6/10 procedures closer to telephone survey-based consumption estimates, and 75th percentile estimates for 3/10 procedures closer to telephone survey-based consumption estimates. ConclusionWe found that applying electronic health record (EHR)-based machine learning nonresponse bias adjustment is essential for debiased opioid consumption estimates from patient surveys. After adjustment, post-discharge surveys can generate reliable opioid consumption estimates. Clinical factors from the EHR combined with TL adjustment appropriately capture differences between responders and nonresponders and should be used prior to generalizing or applying opioid consumption estimates to patient care.