Machine learning and atherosclerotic cardiovascular disease risk prediction in a multi-ethnic population

Area of Science:

• Cardiovascular Medicine
• Medical Informatics
• Machine Learning in Healthcare

Background:

• The Pooled Cohort Equations (PCE) for atherosclerotic cardiovascular disease (ASCVD) risk prediction have limitations in diverse populations, including Asian and Hispanic individuals, and for patients with out-of-range or missing variables.
• The performance of machine learning (ML) models in improving ASCVD risk prediction across broader, real-world, multi-ethnic populations remains largely unknown.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting 5-year ASCVD risk in a multi-ethnic cohort using electronic health record (EHR) data.
• To compare the performance of ML models against the traditional Pooled Cohort Equations (PCE), particularly in populations excluded or underserved by PCE.

Main Methods:

• Developed and trained various ML models (logistic regression, random forest, gradient boosting machine [GBM], extreme gradient boosting) on a large EHR database from Northern California.
• Included 262,923 patients aged 18+, stratified into PCE-eligible and PCE-ineligible groups, with analysis of Asian and Hispanic subgroups.
• Assessed 5-year ASCVD risk prediction performance using the area under the receiver-operating characteristic curve (AUC), with and without additional EHR variables.

Main Results:

• The Gradient Boosting Machine (GBM) model demonstrated superior performance (AUC 0.835) in the full multi-ethnic cohort, including PCE-ineligible patients, compared to the PCE (AUC 0.775) in the eligible cohort.
• ML models achieved comparable or improved ASCVD risk prediction accuracy across diverse patient groups, including those with missing or out-of-range variables for PCE.
• GBM performance remained robust even after incorporating additional EHR data for patients aged 40-79.

Conclusions:

• Electronic health record-trained machine learning models offer a promising approach to enhance ASCVD risk prediction accuracy and equity in diverse, real-world populations.
• ML models can effectively bridge existing gaps in cardiovascular risk assessment, particularly for patients not well-represented by current risk prediction tools like the PCE.