JMIR Publications

ABSTRACT

Background:

Medical residency is associated with long working hours, demanding schedules, and high stress levels, which can lead to burnout among resident physicians. Although wearable and machine learning-based interventions can be useful in predicting potential burnout, existing models fail to clinically explain their predictions, thereby undermining the trustworthiness of the research findings and rendering the intervention apparently useless to residents. This paper develops, EMBRACE, Explainable Multitask Burnout pRediction using AdaptivE deep learning, that employs a novel framework for predicting burnout that is clinically explainable. At first, we develop, a wearable sensor based improved workplace activity and stress detection algorithm, using deep multi-task learning. Next, we present a novel Adaptive Multi-Task Learning (MTL) framework built on top of our activity and stress detection algorithm, to automatically detect burnout. Additionally, this model also completes the resident burnout survey automatically such a way that it can clinically estimate the same burnout level i.e., clinically explainable and trustworthy estimation. We evaluated the efficacy and explainability of EMBRACE using a real-time data collected from 28 resident physicians (2-7 days each) with appropriate IRB approval (IRB# 2021-017).

Objective:

This study aims to address the high incidence of burnout among resident physicians by developing EMBRACE, an Explainable Multitask Burnout pRediction using Adaptive deep learning framework. The objective is to predict burnout in a clinically explainable manner using wearable sensor data and machine learning techniques.

Methods:

EMBRACE employs a novel multitask learning approach that integrates wearable sensor data for improved detection of workplace activity and stress. The framework includes a deep learning model for simultaneous activity recognition and stress level classification, followed by an adaptive multitask learning algorithm for burnout prediction and explanation. The approach was validated using real-time data collected from 28 resident physicians over periods ranging from 2 to 7 days, with subsequent analysis based on the Mini-Z Burnout Survey.

Results:

The EMBRACE framework demonstrated significant accuracy in detecting workplace activities and stress levels, with overall accuracy rates of 91% for activity recognition and 94% for burnout classification. The multitask learning approach effectively predicted burnout measures and provided clinically relevant explanations by automatically completing a burnout survey that aligns with physicians' burnout levels. Correlation analysis further elucidated the relationship between work activities, stress levels, and burnout, highlighting the impact of work interruptions and time pressures.

Conclusions:

EMBRACE represents a significant advancement in the prediction and explanation of burnout among resident physicians. By leveraging wearable sensor data and adaptive deep learning, it offers a clinically explainable tool for early detection and intervention. Future work will focus on expanding the dataset, comparing the framework with existing models, and exploring user satisfaction with the explainability aspect of the predictions. This research underscores the potential of machine learning and wearable technology in addressing critical mental health challenges in the medical profession.