Estimating 10-Year Cardiovascular Disease Risk in Primary Prevention Using UK Electronic Health Records and a Hybrid...

Area of Science:

• Computational biology and bioinformatics
• Machine learning in healthcare
• Cardiovascular disease research

Background:

• Cardiovascular disease (CVD) is a major cause of preventable death, necessitating early risk stratification for primary prevention.
• Traditional risk models have limitations in flexibility and handling complex data.
• Integrating structured and textual electronic health record (EHR) data can improve CVD risk prediction and equity.

Purpose of the Study:

• To develop a hybrid multitask deep learning model (MT-BERT) for predicting 10-year CVD risk.
• To integrate structured and textual EHR features for enhanced individualized risk stratification.
• To support equitable CVD risk assessment across diverse demographic groups.

Main Methods:

• Developed MT-BERT using EHR data from 469,496 patients (aged 40-85 years).
• Jointly encoded structured variables and textual notes using DistilBERT and MLP.
• Employed a custom FocalCoxLoss function and multihead attention for cross-modal interactions and time-to-event prediction.

Main Results:

• MT-BERT achieved Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.744 (males) and 0.782 (females) on the test set.
• External validation confirmed model performance (AUROC 0.736 males, 0.775 females).
• Model performance showed heterogeneity across ethnicity and deprivation subgroups, with lower accuracy in certain minority groups.

Conclusions:

• The hybrid MT-BERT model effectively predicts 10-year CVD risk by integrating diverse EHR data.
• The multitask approach enables individualized risk stratification and time-to-event estimation.
• Findings support advancing equity-aware, data-driven strategies for CVD prevention in diverse populations.