Modeling interactions of health-based and socioeconomic work disability risks with Bayesian networks

Abstract Background Work disability pensions pose a significant burden on society, both in terms of economic costs and the availability of a sustainable workforce, especially as populations age. Individual risk factors for pensions, such as poor health, unemployment and low socioeconomic status are well-known. This study aims to uncover the complex interactions between socioeconomic and health-related risk factors using machine learning and displaying them in a graphical model in a way that traditional multivariate statistical models are not able to. Methods Using repeated survey data from the Northern Finland Birth Cohort 1966 (NFBC1966) we predicted the work ability score, with low scoring known as a strong predictor of disability pensions, of the cohort members at age 46 using Bayesian network structure learning algorithms. Our predictors included socioeconomic and health measures at birth and follow-ups at ages 14, 31 and 46. Results Using 22 predicting variables we modeled a structure with 114 interactions. The fit of the model to the data was BIC = -15.01. According to cross-validation, the predictive accuracy of the model is .80, SD = .009. Conclusions The high interconnectedness between the variables implies that modeling risks with traditional statistical methods and controlling for all interactions may not be sufficient nor viable to understand the complex interactions between socioeconomic background, employment history and health that contribute to work disability risk. Probabilistic graphical modeling with Bayesian networks allows all these interactions to be included in predictive modeling. By visualizing and pinpointing key variables with significant influence on work ability, this study enables more precise intervention planning and resource targeting, maximizing public health impact for work disability risk management and related societal cost savings. Key messages • Work disability risk is shaped by complex interactions between health and socioeconomic factors, which traditional statistical models may either oversimplify or ignore. • Interpretations of Bayesian networks can offer a powerful tool for policymakers to design data-driven, targeted interventions that reduce disability pension occurrence.