Synthetic data-based risk modelling for 30-day mortality after infective endocarditis surgery: a leakage-free methodological evaluation

Code, de-identified data and results for "Synthetic data-based risk modelling for 30-day mortality after infective endocarditis surgery: a leakage-free methodological evaluation" (European Heart Journal - Digital Health). One participating centre (n = 148) is excluded from every development step. The remaining 5,107 patients are split into a training partition of 4,085, the only data the generator is fitted on, and an internal hold-out of 1,022 that neither the generator nor the predictive model sees. The primary model uses 35 pre-operative predictors; seven variables recorded after operation, one intra-operative variable and the two EuroSCORE-derived terms are excluded, so that the comparison with the logistic EuroSCORE is independent. The archive contains a train-on-real benchmark and a real-plus-synthetic model as formal comparators, paired DeLong tests for every comparison in both cohorts, a sensitivity analysis over synthetic training-set size and label prior, raw unenforced synthetic draws alongside the enforced draws, a detection test on two column sets with three classifiers, a privacy evaluation by membership and attribute inference, a recalibration audit, a restricted cubic spline sensitivity analysis, alternative model classes, and a variable-level provenance and missingness table for all 157 source columns. Contents: data, splits, synth, scripts, results, figures, checksums and a reproduction script. Every number in the manuscript and the appendix is produced by a script in this archive.