Beyond Atrial Fibrillation: Machine Learning Algorithm Predicts Stroke in Adult Patients With Congenital Heart...

Area of Science:

• Cardiology
• Medical Informatics
• Artificial Intelligence in Medicine

Background:

• Adult congenital heart disease (ACHD) patients face increased risks of stroke and systemic embolism (SSE).
• Accurate risk prediction is crucial for managing SSE in this vulnerable population.
• Existing clinical risk scores may not fully capture SSE risk in ACHD.

Purpose of the Study:

• To develop and validate artificial intelligence (AI) models for robust SSE risk prediction in ACHD patients.
• To compare the performance of machine learning (ML) algorithms against traditional clinical risk scores.
• To identify key predictors of SSE in ACHD using explainable AI methods.

Main Methods:

• Utilized deidentified insurance claims data from 49,276 ACHD patients (2009-2014).
• Trained two ML algorithms, regularized Cox regression (RegCox) and extreme gradient boosting (XGBoost), on a development cohort (70%).
• Validated model performance using AUC and employed SHAP to identify risk drivers.

Main Results:

• ML models significantly outperformed the CHA₂DS₂-VASC score (AUC 0.66) in SSE prediction.
• RegCox achieved AUCs of 0.82, 0.81, and 0.80 at 1-, 2-, and 5-year follow-ups, respectively.
• Atrial septal defect (ASD) was identified as a significant SSE predictor by ML algorithms.

Conclusions:

• AI-driven ML models offer superior SSE risk prediction in ACHD patients compared to current clinical scores.
• The identification of ASD as a key risk factor by ML models can refine risk stratification.
• These findings support the integration of advanced ML techniques for improved cardiovascular risk management in ACHD.