Acute myocardial infarction prognosis prediction with reliable and interpretable artificial intelligence system

Area of Science:

• Artificial Intelligence in Medicine
• Cardiovascular Disease Prognosis

Background:

• Accurate prediction of mortality after acute myocardial infarction (AMI) is essential for effective patient management.
• Current AI systems lack the reliability and interpretability required for clinical adoption in AMI prognosis.

Purpose of the Study:

• To develop a reliable and interpretable AI framework, named RIAS, for predicting short- and long-term mortality in AMI patients.
• To provide clinicians with actionable insights and decision support tools for managing AMI.

Main Methods:

• The RIAS framework was developed, emphasizing reliability and interpretability with automated model optimization.
• RIAS integrates global and local explanations, including SHAP values and "what if" counterfactual scenarios.
• The framework was applied to data from the Korean Acute Myocardial Infarction Registry, comparing FT-Transformer against XGBoost and MLP.

Main Results:

• FT-Transformer demonstrated superior sensitivity and comparable AUROC and F1 scores to XGBoost in AMI prognosis.
• RIAS identified statin-based medications, beta-blockers, and age as significant predictors of mortality.
• The framework provided reliable, interpretable predictions with local explanations and counterfactual examples for realistic clinical scenarios.

Conclusions:

• RIAS overcomes the
• black-box
• issue in AI by offering interpretable predictions and explanations.
• The system's counterfactual explanations enhance clinical utility by enabling patient-specific scenario simulation.