Multimodal Data-Driven Explainable Prognostic Model for Major Adverse Cardiovascular Events Prediction in Patients With Unstable Angina and Heart Failure With Preserved Ejection Fraction: Multicenter, Cross-Regional Cohort Study.

BACKGROUND

Heart failure with preserved ejection fraction (HFpEF) and unstable angina (UA) often coexist in clinical practice, constituting a high-risk cardiovascular phenotype with a markedly increased incidence of major adverse cardiovascular events (MACEs). The identification of high-risk patients within this population is crucial for reducing complications, improving outcomes, and guiding clinical decision-making.

OBJECTIVE

This study aimed to develop and externally validate predictive models based on machine learning algorithms to estimate the risk of MACEs in patients with coexisting UA and HFpEF, and to construct an online risk calculator to support individualized prevention strategies.

METHODS

This multicenter cohort study included 4459 patients with both HFpEF and UA admitted to 7 hospitals across eastern, central, and western China between January 1, 2015, and December 31, 2021. Patients were divided into the derivation cohort (n=2923) and external validation cohort (n=1536) based on geographic regions. Clinical, laboratory, and imaging data were extracted from electronic medical records. Key predictors were identified using a hybrid feature selection method combining least absolute shrinkage and selection operator and Boruta algorithms. A total of 33 survival models were developed, including a variety of machine learning algorithms and survival analysis models. The model with the best concordance index (C-index) performance was deployed as a web-based risk calculator. Additionally, we assessed other performance indicators of the best-performing model, including the area under the receiver operating characteristic curve, accuracy, sensitivity, specificity, recall, F1-score, Brier scores, calibration curves, and decision curve analysis.

RESULTS

Using a combination of the least absolute shrinkage and selection operator regression and the Boruta algorithm, 7 key predictors were identified: diabetes mellitus, blood platelet count, triglyceride, systemic inflammatory response index, triglyceride-glucose-BMI, N-terminal pro-brain natriuretic peptide, and atherogenic index of plasma. The surv.xgboost.cox model was used to predict MACEs in patients with UA and HFpEF due to its superior C-index. The model demonstrated the following performance metrics in the external validation cohort: a C-index of 0.788; cumulative/dynamic area under the curve of 0.81; and area under the curve values at 20, 30, and 40 months of 0.809 (95% CI 0.745-0.873), 0.784 (95% CI 0.745-0.824), and 0.807 (95% CI 0.776-0.838), respectively. The model exhibited satisfactory calibration and clinical utility in predicting 40-month MACEs. Model interpretability was enhanced using Shapley Additive Explanations for survival analysis to provide global and individual explanations. Furthermore, we converted the surv.xgboost.cox-based model into a publicly available tool for predicting 40-month MACEs, providing estimated probabilities based on the predictive indicators entered.

CONCLUSIONS

We developed a surv.xgboost.cox-based predictive model for MACEs in patients with the dual phenotype of HFpEF and UA. We implemented this model as a web-based calculator to facilitate clinical application.