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

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Cardiovascular Disease Research

Background:

• Machine learning models for rare diseases struggle with scarce data, leading to performance degradation and unstable explanations.
• Clinical trust in AI models is undermined by unreliable predictions and explanations in low-data scenarios.

Purpose of the Study:

• To investigate the efficacy of synthetic data augmentation in maintaining prediction performance.
• To assess the improvement in explainability stability using synthetic data.
• To address challenges in cardiovascular disease risk prediction for Chronic Myeloid Leukemia patients.

Main Methods:

• Utilized 235 real-world patient datasets for cardiovascular disease risk prediction.
• Employed Differential Privacy-Conditional Generative Adversarial Networks (DP-CGANS) to generate synthetic patient data.
• Evaluated model performance and explanation stability (SHAP) across decreasing sample sizes, comparing real-data-only versus real + synthetic data.

Main Results:

• Synthetic data augmentation maintained stable prediction performance.
• Real-data-only models exhibited substantial performance degradation with reduced sample sizes.
• Synthetic data significantly improved the stability of SHAP (SHapley Additive exPlanations) values.

Conclusions:

• Synthetic data augmentation is a promising approach to overcome performance degradation in rare disease ML models.
• Augmented data enhances the stability of model explanations, fostering greater clinical trust.
• This method offers a viable solution for building reliable ML models for diseases with limited patient data.