Development and validation of a deep neural network for predicting coronary heart disease in hypertensive patients using 24-hour ambulatory blood pressure monitoring: a retrospective study.
Coronary heart disease (CHD) remains a leading cause of morbidity and mortality worldwide. Early identification of high-risk hypertensive patients is crucial for preventing cardiovascular events. While traditional risk scores rely on static clinical measurements, 24-h ambulatory blood pressure monitoring (ABPM)-derived time in target range (TTR) captures dynamic blood pressure control patterns that may improve risk stratification. Machine learning methods, particularly deep neural networks, offer an enhanced capability to model complex non-linear relationships in high-dimensional clinical data, compared with conventional statistical approaches.
This single-center retrospective cohort study included 1,026 patients admitted between January 2023 and December 2024, with 718 patients allocated to model development and 308 to internal validation. A deep neural network model with three hidden layers was developed and compared against eight conventional machine learning algorithms (logistic regression, naïve Bayes, k-nearest neighbors, random forest, support vector machine, XGBoost, LightGBM, and CatBoost). Thirty-two variables spanning demographics, clinical data, laboratory results, echocardiographic measures, and blood pressure indices were evaluated. Continuous variables were discretized into quartile-based categories to enhance clinical interpretability. Feature selection employed a two-step process combining the Boruta algorithm and least absolute shrinkage and selection operator (LASSO) regression, with variance inflation factor analysis confirming the absence of collinearity. Model selection prioritized balanced performance across discrimination (AUC), calibration (Brier score), and clinical utility (decision curve analysis) in the independent validation cohort. Interpretability was evaluated using SHAP (SHapley Additive exPlanations) values.
The deep neural network model achieved optimal balanced performance with an AUC of 0.822 (95% CI: 0.793-0.850) in the training cohort and 0.796 (95% CI: 0.749-0.846) in the validation cohort, accompanied by the lowest Brier score (0.172), indicating superior calibration. Nine predictors were retained: diabetes mellitus, mean systolic blood pressure, time in target range of systolic blood pressure, left atrial diameter, left ventricular end-systolic diameter, left ventricular ejection fraction, use of antihypertensive medications, calcium channel blockers, and β-blockers. SHAP analysis identified TTR and blood pressure control parameters as the primary drivers of model predictions.
The developed deep neural network model enables early identification of high-risk CHD patients with hypertension through interpretable, routinely available clinical variables. Prospective multicenter external validation is warranted to confirm its generalizability across diverse populations and clinical settings.
This single-center retrospective cohort study included 1,026 patients admitted between January 2023 and December 2024, with 718 patients allocated to model development and 308 to internal validation. A deep neural network model with three hidden layers was developed and compared against eight conventional machine learning algorithms (logistic regression, naïve Bayes, k-nearest neighbors, random forest, support vector machine, XGBoost, LightGBM, and CatBoost). Thirty-two variables spanning demographics, clinical data, laboratory results, echocardiographic measures, and blood pressure indices were evaluated. Continuous variables were discretized into quartile-based categories to enhance clinical interpretability. Feature selection employed a two-step process combining the Boruta algorithm and least absolute shrinkage and selection operator (LASSO) regression, with variance inflation factor analysis confirming the absence of collinearity. Model selection prioritized balanced performance across discrimination (AUC), calibration (Brier score), and clinical utility (decision curve analysis) in the independent validation cohort. Interpretability was evaluated using SHAP (SHapley Additive exPlanations) values.
The deep neural network model achieved optimal balanced performance with an AUC of 0.822 (95% CI: 0.793-0.850) in the training cohort and 0.796 (95% CI: 0.749-0.846) in the validation cohort, accompanied by the lowest Brier score (0.172), indicating superior calibration. Nine predictors were retained: diabetes mellitus, mean systolic blood pressure, time in target range of systolic blood pressure, left atrial diameter, left ventricular end-systolic diameter, left ventricular ejection fraction, use of antihypertensive medications, calcium channel blockers, and β-blockers. SHAP analysis identified TTR and blood pressure control parameters as the primary drivers of model predictions.
The developed deep neural network model enables early identification of high-risk CHD patients with hypertension through interpretable, routinely available clinical variables. Prospective multicenter external validation is warranted to confirm its generalizability across diverse populations and clinical settings.