Enhancing Explainable AI Stability with Realistic Synthetic Data for Cardiovascular Risk Prediction

Data-driven such as machine learning models in rare disease research faces a significant challenge: scarce data cause performance degradation and unstable explanations that undermine clinical trust. This study investigated whether synthetic data augmentation can maintain prediction performance and improve explainability stability for cardiovascular disease risk prediction in Chronic Myeloid Leukemia patients. Using 235 real-world patients' data, we employed DP-CGANS to generate synthetic records and evaluated their impact across decreasing sample sizes, comparing real-data-only versus real + synthetic data scenarios. Synthetic augmentation showed stable performance while real-data-only models degraded substantially. We observed synthetic data can significantly improve SHAP stability. These findings demonstrate that synthetic data augmentation has potential to address performance degradation and explanation instability in building reliable ML models for diseases with scarce patient data.