Development of a predictive model for in-hospital new-onset atrial fibrillation in older adults with hypertension and acute myocardial infarction, enhanced by SHAP interpretability: a retrospective cohort study

TL;DR

This study creates a predictive model for new-onset atrial fibrillation in older adults with hypertension and heart attacks, using machine learning and an online tool for real-time risk assessment.

Contribution

The study introduces an interpretable predictive model and an online tool for real-time NOAF risk assessment in a specific high-risk population.

Findings

The model identified eight key predictors with strong discrimination (AUC 0.895 in training, 0.883 in validation).

An interactive web-based tool was developed for real-time NOAF risk prediction and clinical use.

SHAP values improved model interpretability and clinical relevance for high-risk patient management.

Abstract

Acute myocardial infarction (AMI) remains a leading cause of mortality, particularly among older adults with hypertension, who are at a heightened risk for complications such as new-onset atrial fibrillation (NOAF). Despite existing research, predictive models for NOAF in this population are limited in both scope and clinical utility, often lacking interpretability, which hinders their use in clinical practice.

This study aims to develop and validate a predictive model for NOAF in older adults with hypertension who have experienced AMI, incorporating machine learning techniques and SHapley Additive exPlanations (SHAP) value to enhance the model’s interpretability and clinical utility.

A retrospective cohort study was conducted on 2,140 older hypertensive adults hospitalized with AMI at the First Hospital of Qinhuangdao. Key features were selected using Boruta, LASSO regression, and logistic regression. A predictive nomogram was constructed via multivariate logistic regression, and SHAP value was utilized to explain the model’s predictions. The model’s performance was assessed using ROC, AUC, calibration curves, and clinical utility was evaluated via Decision Curve Analysis and Clinical Impact Curves.

The model identified eight key predictors: age, left atrial diameter, ejection fraction, white blood cell count, triglycerides, low-density lipoprotein, NT-proBNP, and potassium. The nomogram demonstrated excellent discrimination with an AUC of 0.895 in the training set and 0.883 in the validation set. An interactive web-based tool was developed (https://longmao.shinyapps.io/NOAF/) to provide real-time NOAF risk predictions. SHAP values clarified feature contributions, enhancing the model’s interpretability and clinical relevance.

This study presents an interpretable predictive model for NOAF in older adults with hypertension and AMI, introduces an online tool for real-time exploratory risk assessment. The model demonstrates high discriminative performance and potential clinical relevance for early detection and personalized management of high-risk patients.

Figures