Interpretable deep learning with pharmacogenomics predicts myocardial infarction in angiotensin receptor blocker treated hypertension.

OBJECTIVE

A significant residual risk of myocardial infarction (MI) persists in hypertensive patients treated with angiotensin receptor blockers (ARBs). Conventional risk models often fail to capture the complex, nonlinear interactions among clinical phenotypes, comorbidities, and pharmacogenetic factors.

METHODS

This retrospective cohort study included 1229 hospitalized hypertensive patients treated with ARBs. We analyzed 26 clinical variables and two key genetic polymorphisms: AGTR1 rs5186 and CYP2C9 rs1057910. A hybrid downsampling technique was employed to address severe class imbalance. Ten machine learning models were developed, with a focus on the interpretable deep learning model, TabNet.

RESULTS

The TabNet model demonstrated superior predictive performance, achieving an area under the receiver operating characteristic curve (AUC) of 0.887 and a recall of 0.795 in the validation cohort. Model interpretation using Shapley Additive Explanations (SHAP) identified pre-existing coronary heart disease, hypertension stage, chronic heart failure, sex, hyperlipidemia, and the AGTR1 rs5186 polymorphism as the most significant predictors of MI. The CYP2C9 rs1057910 variant showed a risk-modifying interaction effect in patients with pre-existing coronary heart disease.

CONCLUSION

An interpretable deep learning model integrating clinical and pharmacogenetic data effectively predicts MI risk in ARB-treated hypertensive patients. The AGTR1 rs5186 polymorphism is an independent predictor of MI, highlighting its potential as a prognostic biomarker. This clinico-pharmacogenetic approach offers a powerful tool for personalized risk stratification and may guide more targeted preventive interventions.