Construction of a predictive model for the efficacy of enhanced external counterpulsation therapy in patients with heart failure with preserved ejection fraction based on automated machine learning and echocardiography

Abstract

Objective This study aimed to integrate baseline clinical characteristics and multi-dimensional echocardiographic parameters of patients with heart failure with preserved ejection fraction (HFpEF), construct and validate an individualized predictive model for the efficacy of enhanced external counterpulsation (EECP) therapy in HFpEF patients based on an automated machine learning (AutoML) framework, and develop a corresponding clinical decision support system. Methods A retrospective cohort study design was adopted, consecutively enrolling 550 HFpEF patients who received complete EECP treatment at the cardiovascular center of two tertiary grade-A hospital between January 2018 and December 2023. An AutoML predictive framework driven by an improved Newton's downhill optimizer (INDO) was constructed, automatically performing joint optimization of feature subset selection, model algorithm selection, and hyperparameter configuration under a rigorous double-layered nested cross-validation strategy. Five benchmark models, including logistic regression, support vector machine, adaptive boosting, extreme gradient boosting, and light gradient boosting machine, were simultaneously established for performance comparison. Multi-dimensional evaluation was conducted using the area under the receiver operating characteristic curve (ROC-AUC), area under the precision-recall curve (PR-AUC), accuracy, sensitivity, specificity, F1 score, Brier score, and decision curve analysis. The robustness of features was verified through LASSO regression, and the SHAP framework was employed for model interpretability analysis. Results Among the 550 HFpEF patients, significant differences were observed between the treatment-effective group and the treatment-ineffective group in terms of age, body mass index, left ventricular ejection fraction, left atrial volume index, and E/e' ratio. The AutoML model optimized based on INDO achieved a ROC-AUC of 0.9254, a PR-AUC of 0.8425, an F1 score of 0.8285, and a Brier score of 0.1206 on the independent test set, with all metrics outperforming the five benchmark models. Decision curve analysis demonstrated that the model yielded a positive net clinical benefit within a threshold probability range of 16% to 96%. SHAP interpretability analysis revealed the 6-minute walk distance (6MWD), left ventricular ejection fraction (LVEF), age (Age), left atrial volume index (LAVI), E/e' ratio, and body mass index (BMI) as key predictors influencing the efficacy of EECP therapy. Conclusion The predictive model constructed by integrating echocardiographic parameters and clinical features based on the AutoML framework driven by the INDO optimization algorithm can relatively accurately predict the individualized efficacy of EECP therapy in HFpEF patients, providing a interpretable, and clinically actionable intelligent assessment tool for indication screening and individualized decision-making regarding EECP therapy in HFpEF patients.