Identifying Burnout in Pediatric Clinicians Using Wearable Biometric Data: Development and Feasibility Study
Journal of Medical Internet Research ·
Background: Addressing the burnout epidemic among health care providers remains one of the foremost challenges of modern health care systems. Past researchers have turned to passively gathered biometric data from wearable devices to identify key variables that predict burnout. However, this has resulted in limited success. Objective: This pilot and feasibility study aimed to test the feasibility and accuracy of developing an Extreme Gradient Boosting machine learning model to classify health care providers as being at higher or lower risk of burnout based on their biometric data. Methods: We conducted a 3-month pilot and feasibility study where health care providers self-reported burnout every 6 weeks and wore a Garmin Venu 3 smartwatch to collect continuous biometric data. Participants were recruited from 3 pediatric urgent care sites and an inpatient hospitalist division of a pediatric health care system. We had a final sample of 41 participants who met our inclusion criteria requiring completion of burnout measurements at 2 of the 3 time points and having worn the watch for at least 70% of the study period. Participants were asked to wear the Garmin watch continuously over the course of the study. Burnout self-report surveys were delivered via email. Our main outcome was burnout as measured using the Maslach Burnout Inventory–Human Services Survey. The measure consists of 3 subscales that together constitute burnout: emotional exhaustion, depersonalization, and reduced personal efficacy. Results: First, we found evidence of feasibility for wearable watches to be used in situ for health care providers to gather data passively for the purposes of classifying burnout. A total of 82% (51/62) of participants met our criteria of wearing the watch 70% of the time, with 66% (41/62) having corresponding burnout scores to be included in analyses. Second, our machine learning model achieved strong overall performance in identifying risk of burnout, with strong sensitivity and specificity, demonstrating that biometric data collected using wearable devices can accurately predict burnout. Conclusions: Our evidence of feasibility offers promising results for a passive method of data collection to be used for classifying burnout risk in other health care settings. Furthermore, the strong performance of our machine learning model provides a foundational step toward improving an organization’s ability to respond and mitigate burnout through thoughtful intervention strategies.