Predicting mental health-related sick leave using stress check data: development and validation of a workplace risk model.

Abstract

Objectives: To develop and validate a predictive model for mental health-related sick leave using data from Japan's Stress Check program linked with personnel records (demographics and sick-leave history), and to evaluate its predictive performance compared with the conventional high-stress classification.

Methods: We conducted a retrospective cohort study of employees (2020-2024) in a 14-company corporate group spanning diverse occupations. A mixed-effects logistic regression model was developed using data from 2020-2023, with company and year as random intercepts. Variable robustness was confirmed through least absolute shrinkage and selection operator (LASSO) logistic regression. Model calibration and discrimination were evaluated using the Brier score and area under the curve (AUC), respectively. The model was temporally validated on 2024 data.

Results: Among 87 138 person-years in the development phase, 695 employees took mental health-related sick leave (0.8%). Seven predictors-younger age, anxiety, depression, sleep disturbance, low job control, poor job suitability, and low coworker support-were significantly associated with sick leave. The recalibrated model showed good discrimination (AUC = 0.708) and calibration (Brier score = 0.0079) in development, and higher performance in validation (AUC = 0.819; Brier score = 0.0026). The conventional high-stress classification performed poorly.

Conclusions: The proposed model demonstrated robust predictive validity and outperformed the conventional high-stress classification. By providing quantitative and continuous risk stratification within the Stress Check framework, it offers a practical approach to support risk-based prioritization and decision-making in occupational health practice.