Application of machine learning for personalized advice about workers’ lifestyle changes to improve sustainable employability: Development and evaluation of a decision support tool

Abstract

Background Occupational Health Check-ups (OHCs) have evolved from identifying work-related health risks to supporting sustainable employability through individualized lifestyle advice, but translating multidimensional demographic, occupational, and health characteristics into effective, personalized recommendations remains complex and lacks validated decision-support methods. Objective Aims of this study are to: 1) develop a Machine Learning Decision Support Tool (ML-DST) based on workers’ demographic, occupational, and health characteristics to support personalized advice about lifestyle changes to improve sustainable employability, and 2) evaluate performance of the ML-DST. Methods Real-world OHC data from a population-based cohort (2010–2020) were used to develop ML-DST models providing lifestyle advice (physical activity, smoking, alcohol, nutrition, relaxation, or no change) to improve workability and vitality. Paired baseline and 1–4-year follow-up data were used to train and evaluate 12 models. Per model, three rule-based algorithms were applied. Performance was considered useful when both True Positive Rate (TPR) and macroF1 score were ≥0.7. The best performing algorithm is proposed per model for inclusion in the final ML-DST. Results Of 20,440 individual OHCs, 3412 paired sets were included. Best TPRs ranged from 0.51 to 0.71. One model achieved TPR and macroF1 scores ≥0.7, supporting nutritional behavior change to improve vitality; however, its profiles were not conceptually meaningful. The other 11 models did not achieve useful quantitative performance. Conclusion Training algorithms for an ML-DST to support advice about lifestyle changes to improve long-term sustainable employability using data from regular Occupational Health Check-ups did not lead to useful model performance for improving workability and vitality.

The authors' abstract, as published at the source. Work, 2026 · DOI ↗