Leveraging machine learning with real-world data for hypothesis generation by identifying exposomic predictors in...

BackgroundMachine learning offers new avenues for complementing traditional epidemiological approaches by analyzing routinely collected, population-based administrative health data.ObjectiveThis study aimed to identify potential exposomic predictors (hypothesis generation) for Parkinson's disease (PD) across the entire French agricultural workforce.MethodsWe applied XGBoost adapted for Cox proportional hazards modeling to assess approximately 180 exposomic factors derived from nationwide administrative health data within the TRACTOR project. Shapley Additive Explanation (SHAP) values were used to assess the importance of each predictor. To provide both model-based and statistical perspectives, SHAP analysis was complemented with classical Cox regression, allowing for transparent assessment of each predictor's contribution to the model and its statistical association with survival. Sensitivity analyses incorporating different exposure lags were conducted. The study included 424,725 farm managers (6,265 PD cases) and 544,788 farmworkers (2,848 PD cases) aged 50+, analyzed separately due to differences in available variables and coding structures.ResultsSeveral occupational factors, including duration of involvement in crop farming and viticulture, emerged as key promoting predictors, surpassing age in predictive importance. Beyond conventional predictors such as type 2 diabetes, less conventional predictors were identified, including work diversification, seasonal employment, hypercholesterolemia, epilepsy, antidepressant use, anxiolytic use, and antibiotic use.ConclusionsThese results contribute to a growing body of evidence supporting the integration of occupational health considerations into PD research and highlight the importance of exploring and identifying potential farming-related risk factors in PD development.