Cardiovascular risk assessment enhanced by automated machine learning in a multi-phase study

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Informatics

Background:

• Cardiovascular diseases (CVDs) are a major global health concern, with existing risk predictors having limitations.
• Lipoprotein (a) [Lp(a)] is a known CVD risk factor, but its predictive power can be enhanced.
• Automated machine learning (AutoML) presents a promising approach to develop sophisticated CVD risk prediction models.

Purpose of the Study:

• To develop and validate AutoML models for predicting lipoprotein (a) levels, specific CVDs, and CVD-related mortality.
• To identify key determinants of CVD risk using machine learning.
• To assess the performance and generalizability of AutoML models in clinical datasets.

Main Methods:

• Utilized clinical datasets from the LURIC (n=3316) and UMC/M (n=423) studies.
• Developed and applied AutoML models in three distinct phases for prediction tasks.
• Employed SHAP analysis to interpret model predictions and identify significant clinical variables.

Main Results:

• Phase 1 models demonstrated good accuracy in identifying CVD determinants (AUC 0.6249–0.9101).
• Phase 2 validation in the UMC/M dataset showed robust performance (AUC 0.7224–0.8417), with statin therapy, age, and NTproBNP highlighted as key predictors.
• Phase 3 models for cardiovascular mortality prediction achieved high AUC values (0.74–0.85), though data drift necessitates model adjustment.

Conclusions:

• AutoML models show significant potential for improving CVD risk prediction and identifying key clinical determinants.
• Validated AutoML models offer robust performance in diverse clinical cohorts.
• Ongoing monitoring and adjustment are crucial for maintaining model performance due to potential data drift.