Finding Associative and Causal Effects of Temporal Changes in Health Features for Prevalent and Incident Cancer in Males: A Machine Learning Approach.

Cancer remains a major public health challenge driven by complex interactions among sociodemographic, behavioral, clinical, and environmental factors. This study investigated how temporal changes in health-related features are associated to prevalent and causally related to incident cancer cases, using longitudinal data from 6,409 male participants in the Atlantic Partnership for Tomorrow's Health (PATH) cohort. For prevalent cancer cases, associations were assessed using univariate and bivariate analyses. For incident cancer cases, statistically significant predictors identified through these analyses were examined within a potential-outcomes-based causal inference framework. The X-Learner algorithm, implemented with Gradient Boosting (GB), Light Gradient-Boosting Machine (LightGBM), and Random Forest (RF) models, estimated Average Treatment Effects (ATEs) and corresponding 95% Confidence Intervals (CIs). Results showed that changes in 12 binary, two continuous, and two categorical features were associated with prevalent cancer cases. Among the three binary, two continuous, and one categorical feature changes that were statistically significant for incident cancer cases, changes in age, arthritis, and hypertension demonstrated modest but consistent causal effects. These findings suggest associative patterns and causal effects, conditional on observational data and modeling assumptions, supporting effective cancer prevention, control, and care planning.