Identifying risk factors of long sickness absences: a registry-based study using explainable AI methods.

Area of Science:

• Occupational Health
• Data Science
• Biostatistics

Background:

• Predicting long-term sickness absence is crucial for workforce management and intervention planning.
• Traditional predictors like prior sickness absence and healthcare utilization have limitations in accuracy.

Purpose of the Study:

• To identify and explore novel variable groups and individual predictors of long sickness absences.
• To enhance prediction accuracy using machine learning (ML) and explainable artificial intelligence (XAI) with a submodel approach.

Main Methods:

• Retrospective analysis of prospectively collected registry data and health examination questionnaires from 11,533 employees.
• Utilized ML and XAI techniques, including a submodel approach, to analyze electronic medical record data.
• Focused on predicting at least one long sickness absence period (≥30 days) over a 2-year follow-up.

Main Results:

• An ensemble model combining all submodels achieved an area under the receiver operating characteristic curve (AUROC) of 0.79.
• Submodels for sickness absence and service use showed the highest AUROC values (0.68-0.74).
• Key predictors identified included reported pain, number of symptoms/diseases, body mass index, short sleep duration, and work/mental health factors.

Conclusions:

• Variables beyond service use and prior sickness absence significantly improve long sickness absence prediction.
• Incorporating these novel predictors offers valuable insights for developing targeted interventions.
• The findings support proactive planning to mitigate work disability risk.