JMIR Publications

ABSTRACT

Background:

Occupational burnout among healthcare and knowledge workers is associated with absenteeism, staff turnover, reduced well-being, and patient-safety risks. Existing burnout interventions are often reactive and depend heavily on self-reported symptoms. Wearable biosensors may provide continuous physiological and behavioral signals that allow earlier burnout-risk detection and more timely intervention delivery.

Objective:

This study aimed to evaluate whether a wearable biosensor–based machine learning system could predict 48-hour burnout risk and whether algorithmically triggered just-in-time adaptive interventions could reduce burnout symptoms over 16 weeks compared with wearable-only monitoring and passive control.

Methods:

A pre-registered, 16-week, three-arm parallel randomized controlled trial was conducted among 218 full-time healthcare and knowledge workers in the United Kingdom. Participants were assigned to a just-in-time intervention arm, a wearable-only arm, or a passive control arm. Wearable devices collected heart rate variability, electrodermal activity, skin temperature, actigraphy, and sleep data. Sixty-four physiological, sleep, activity, and ecological momentary assessment features were extracted. A stacked ensemble model combining XGBoost, bidirectional LSTM, and Random Forest classifiers predicted 48-hour burnout-risk onset. The primary outcome was the Maslach Burnout Inventory–General Survey score at 16 weeks. Linear mixed models assessed intervention effects, and structural equation modeling tested mediation through prediction accuracy and intervention engagement.

Results:

The ensemble model achieved moderate predictive performance, with AUROC = 0.78, sensitivity = 0.74, and specificity = 0.80, outperforming an HRV-only baseline. SHAP analysis identified RMSSD, sleep efficiency, and LF/HF ratio as leading predictors. At 16 weeks, the just-in-time intervention arm showed a significantly greater reduction in burnout than passive control, with an adjusted mean difference of -0.48 and a medium effect size. However, the just-in-time intervention did not significantly outperform the wearable-only arm on the primary outcome. Sequential mediation analysis indicated that prediction accuracy and intervention engagement jointly mediated the effect on burnout outcomes.

Conclusions:

A wearable biosensor–driven machine learning system can predict short-term burnout risk with moderate accuracy and may support clinically meaningful burnout reduction when linked to just-in-time adaptive interventions. However, the absence of clear superiority over wearable-only monitoring, the moderate false-positive burden, and limited follow-up duration suggest that larger, longer, and more diverse confirmatory trials are needed before large-scale implementation. Clinical Trial: ISRCTN14832991