Machine learning-based risk prediction for major adverse cardiovascular events in a Brazilian hospital: Development,...

Area of Science:

• Cardiovascular Disease Research
• Machine Learning in Healthcare
• Predictive Analytics

Background:

• Machine learning models for cardiovascular disease (CVD) risk often lack external validation and interpretability.
• This study developed and validated predictive models for major adverse cardiovascular events (MACE).
• Interpretability analyses were performed to enhance model reliability and personalize interventions.

Purpose of the Study:

• To develop and validate machine learning models for predicting 5-year risk of MACE.
• To assess the generalization ability of models across different populations.
• To analyze model interpretability using LIME and Shapley values for improved clinical application.

Main Methods:

• Trained and validated eight machine learning algorithms on retrospective data from Brazil (RPMS) and the USA (BIDMC).
• Utilized balanced datasets of MACE and non-MACE cases for internal and external validation.
• Evaluated predictive performance using accuracy and ROC AUC; applied LIME and Shapley for interpretability.

Main Results:

• Random Forest demonstrated superior predictive performance with AUCs of 0.871 (internal) and 0.786 (external).
• Accuracy for Random Forest was 0.794 (internal) and 0.710 (external).
• Shapley values provided more consistent feature explanations than LIME for model interpretability.

Conclusions:

• Random Forest exhibited the best generalization for MACE risk prediction.
• Shapley values offered more informative local interpretability compared to LIME.
• Machine learning models with strong generalization and interpretability are recommended for personalized CVD risk assessment and prevention.