Federated Analysis of Cross-European Health Data: Improving CVD prediction with Machine Learning

Abstract Background Federated analysis of nationwide health registers using machine learning could enhance cardiovascular disease (CVD) prediction. As a part of the HealthData@EU Pilot, the current study developed and validated CVD prediction models using data from 22 million individuals across four European countries. Methods We used nationwide register data from 2010 to 2018 including individuals aged 18-85 from Denmark (n = 4.18M), Finland (n = 4.15M), France (11.74M) and Norway (n = 3.64M). Incident CVD was defined as the first occurrence (fatal or nonfatal) of ischemic heart disease, myocardial infarction or ischemic stroke. To predict the 5-year CVD risk, we trained a baseline model with sex and age, and a full gradient boosted tree model including diagnoses and prescribed medications. Model performances were compared using ROC AUC and PR AUC. Feature importance was based on Gini entropy gain. All analyses were performed within each country's secure processing environment, sharing only aggregated data across countries. Results The background characteristics showed differences in the French cohort (female: DK 50.6%, FIN 50.9%, FR 53.9%, NO 49.9%; mean age: DK 47.4y, FIN 48.5y, FR 47.0y, NO 46.7y; death rate: DK 8.9%, FIN 11.4%, FR 7.3%, NO 11.6%). CVD incidence per 1000 person-years: 8.5 DK, 10.2 FIN, 5.2 FR, 9.2 NO. For the baseline model, the ROC AUC varied between 0.80-0.85. For the full model, the ROC AUC improved by + 3% to + 5%. The most influential predictors shared across countries included traditional CVD predictors, and non-traditional ones such as systemic corticosteroids (ATC:H02A). Conclusions A register-based CVD prediction model showed high performance though the improvement over sex and age-based model remained moderate.