An interpretable machine learning model for predicting mortality risk in patients with hypertension and heart failure with preserved ejection fraction

Objective To develop an interpretable machine learning model to predict all-cause mortality risk in patients with hypertension and heart failure with preserved ejection fraction (HFpEF).

Methods A prospective cohort of 847 patients diagnosed with hypertension and HFpEF from 3 tertiary hospitals in Shanxi Province between April 2014 and March 2019 was followed until April 1, 2022. All-cause mortality was used as the outcome event. The cohort was randomly divided into training (70%) and testing (30%) sets. The training set was used to construct prediction models, and the testing set was used for performance evaluation. Predictors were selected using the least absolute shrinkage and selection operator (LASSO)-Cox regression. Six machine learning models, including extreme gradient boosting (XGBoost), logistic regression, random forest (RF), decision tree (DT), support vector machine (SVM), and multilayer perceptron (MLP), were developed to predict the 3-year all-cause mortality risk. Model performance was evaluated using receiver operating characteristic (ROC) curves, calibration curves, and clinical decision curves. The Shapley additive explanations (SHAP) framework was applied to the simplified optimal model for interpretability analysis, and restricted cubic spline models were used to explore the nonlinear relationships between key predictors and all-cause mortality.

Results After a median follow-up of 4.25 (P₂₅, P₇₅ 2.86, 6.17) years, 224 patients (26.4%) experienced all-cause mortality. Using the LASSO-Cox regression, 17 predictors were identified from patients’ clinical characteristics, including vital signs, laboratory tests, and imaging results. The RF model achieved the best performance, with an area under the ROC curve (AUC) of 0.823 (95% CI: 0.693–0.950), accuracy of 84.0%, sensitivity of 82.3%, specificity of 83.0%, and an F1 score of 0.810. Calibration and clinical decision curves confirmed the RF model's good calibration and clinical applicability. SHAP feature importance analysis revealed that age, estimated glomerular filtration rate (eGFR), systolic blood pressure, and body mass index (BMI) were the top 4 factors influencing all-cause mortality in patients with hypertension and HFpEF. Further restricted cubic spline analysis indicated that age ＞72 years, eGFR ＜72.9 mL/(min·1.73 m²), systolic blood pressure ＞136 mmHg, and BMI ＞26.6 kg/m² were associated with increased all-cause mortality risk. To enhance the clinical applicability of risk warning thresholds, clinically practical thresholds were selected for Cox regression analysis. The results showed that systolic blood pressure ＞135 mmHg (HR=1.362, 95% CI: 1.020–1.819) and eGFR ＜70 mL/(min·1.73 m²) (HR=1.519, 95% CI: 1.135–2.034) were both significantly associated with an increased risk of all-cause mortality.

Conclusions The RF-based prediction model can effectively estimate the 3-year all-cause mortality risk in patients with hypertension and HFpEF after discharge. SHAP-based interpretability analysis can provide clear insights for clinical decision-making.