Abstract P465: Linking Electronic Health Record and Wearable Device Data Can Provide Insight Into Expected Changes in Captured Heart Rate Due to Medication Usage

Introduction: From remote patient monitoring to arrythmia detection, heart rate (HR) data collected by wearable devices is widely utilized across healthcare practice and research. Although accuracy varies across devices, data are generally assumed to be “internally consistent” (i.e., within-person trends are reliable). As a result, individuals’ change from baseline is commonly used as a biomarker for risk models. However, defining a baseline from device data alone is problematic. External influences such as therapeutic use of HR-altering medications (e.g., beta-blockers) may alter patterns derived from HR data, and computational models not accounting for such may become less effective. Hypothesis: Initiation of a HR-altering medication as denoted in an individual’s EHR will correspond to a change in HR data collected by their wearable device. Methods: This study utilized data from 1938 patients at Cedars-Sinai Medical Center who linked Apple Watches to their EHR 4/2015 – 11/2018. It focused on 32 patients with ≥10 days of HR data before and after their 1 st prescription of a HR-altering medication, with no prior orders of these medication over the prior year. Mean daytime (9a-9p) hourly HR were compared before and after medication start (Fig 1) using a linear mixed model, adjusted for hour of the day, measurements per hour, time since medication was started, and time from the start of observation (to account for pre-medication trends) with a patient-level random effect. Results: HR-altering medication produced significant changes in device-measured HR. Post-medication HR averaged 3.2 BPM lower [95% CI, 2.8-3.6, p<.001] than pre-medication periods. Secondary analyses identified decreased hourly HR range (p<.001), and increased wear time (p=.005) in pre/post medication periods that may also impact model efficacy. Conclusion: Linking EHR and wearable data offers a promising route in identifying events that may bias computational tools and indicate a need for model-retraining or use of adjusted relationships.