Joseph Hegenbart
Accurate prediction of thirty-day hospital readmissions in large clinical cohorts depends on data quality, the integrity of pre-processing, and the choice of learning strategy. The aim of this study was to develop and empirically validate an artificial-intelligence-based method for generating clinically coherent medical data in order to improve the precision, stability, and real-world applicability of models forecasting adverse outcomes in patients with chronic diseases. The analysis used the full Diabetes 130-Hospitals dataset (n = 101,766), applying multi-stage pre-processing, diffusion-based generation of synthetic records, three learning regimes based on real, synthetic, and combined data. The results show that pre-processing markedly strengthened dataset integrity: up to 18% of missing laboratory values were fully reconstructed, numerical ranges were stabilised, and the minority class representing 11.2% of cases was rebalanced to 45-50% while preserving correlation structures. Synthetic data demonstrated close correspondence to real distributions, with similarity values of 0.97 and correlation alignment of 0.95. In all comparisons, the combined learning regime was the strongest: gradient-boosting models achieved the highest discrimination, neural-network models offered the most accurate probability calibration, and synthetic-only learning consistently produced the weakest outcomes. Models trained on combined data maintained 95% of baseline performance under noise and simulated missingness, compared with 88% for real-only learning and 84% for synthetic-only learning. The findings suggest that synthetic augmentation, when integrated with real clinical data, can improve predictive discrimination, calibration, and robustness. However, the modest F1-scores indicate that further threshold optimisation and prospective external validation are required before clinical deployment.