N-of-1 modeling of wearable and survey data identified five distinct burnout signatures among ICU nurses, though differences in burnout scores between clusters were not significant (p>0.2).
Observational (n=68)
Does N-of-1 machine learning modeling using wearable data identify distinct burnout signatures in ICU nurses?
N-of-1 machine learning modeling using wearable data can identify individualized predictors of well-being and distinct burnout signatures among ICU nurses.
p-value: p=>0.2
Introduction: Burnout among healthcare workers is multifactorial and highly individualized. Traditional group-level analyses may obscure the personal dynamics that drive well-being. This study used a personalized machine learning approach (N-of-1 modeling), along with SHAP features (importance scores) to identify key influences on well-being for each ICU nurse, using wearable sensor data and daily surveys. We hypothesized that well-being predictors differ across individuals and aimed to identify distinct patterns by grouping them and assess their association with specific burnout domains defined by conventional burnout scores. Methods: We collected 35 days of physiological (sleep, heart rate, steps), contextual (workday/off-day, number of patients), and self-reported well-being data (score: 0–500) from ICU nurses wearing Fitbit devices. For each participant, we trained an N-of-1 machine learning model to predict next-day well-being from these physiological data and contextual data. SHAP values were used to quantify the contribution of each feature per individual. We then clustered participants based on their SHAP profiles and examined associations with Maslach Burnout Inventory scores: emotional exhaustion (EE), depersonalization (DP), and personal achievement (PA). Results: Data from 68 participants(average age: 32.5 years, 81% female) were included in the analysis. The average prediction error across participants was mean absolute error=39.8 and mean absolute percentage error=0.19, indicating moderate predictive performance. Clustering based on SHAP features revealed five groups (cluster sizes: 7–24): each cluster was related to sleep duration, workload, workday, indifferent, and physiological stress. While differences were not statistically significant (p>0.2), the sleep-sensitive group showed the highest DP score, the workday-sensitive group showed the highest EE score, and the workload-sensitive group showed the lowest DP score. Conclusions: This study demonstrates the power of N-of-1 modeling in capturing person-specific predictors of well-being and uncovers individualized burnout signatures. These insights will enable tailored intervention design, essential in high-stress settings like critical care, where burnout drivers and coping patterns are diverse.
Ito et al. (Sun,) conducted a observational in Burnout in ICU nurses (n=68). N-of-1 machine learning modeling with SHAP features was evaluated on Next-day well-being prediction and association of SHAP profile clusters with Maslach Burnout Inventory scores (p=>0.2). N-of-1 modeling of wearable and survey data identified five distinct burnout signatures among ICU nurses, though differences in burnout scores between clusters were not significant (p>0.2).