A new 5-year risk score for hypertrophic cardiomyopathy: integrating echocardiography, clinical, and ECG data

Hypertrophic cardiomyopathy (HCM) is associated with increased risk of heart failure and sudden cardiac death. Risk prediction remains challenging, and existing risk predictors underutilize echocardiographic and ECG data routinely available in clinical practice.

This study aimed to develop an integrated risk score, by incorporating echocardiographic, clinical and electrocardiographic (ECG) variables.

We retrospectively analyzed data from 1012 HCM patients enrolled in a monocentric cohort of the SHARE registry. The target outcome was the occurrence of a 5-year composite cardiovascular event including heart failure (cardiac transplant, device implantation, LVEF<35%, or new onset NYHA III/IV) and ventricular arrhythmic events (SCD, cardiac arrest, appropriate ICD therapy). In addition to conventional echocardiographic parameters, we included clinical variables (NYHA class, systolic and diastolic blood pressure) and ECG-derived features (heart rate, ECG rhythm, PR, QRS, QT intervals). Four supervised machine learning models, namely, logistic regression, support vector machine, random forest, and gradient boosting classifiers, were trained and evaluated using 5-fold nested cross-validation.

Models trained using only echocardiographic features achieved a balanced accuracy of 0.749±0.025 and F1 score of 0.654±0.031 for random forest (sensitivity 0.741±0.029, specificity 0.757±0.031), and a balanced accuracy of 0.730±0.022 and F1 score of 0.632±0.028 for logistic regression (sensitivity 0.713±0.035, specificity 0.748±0.034). The addition of clinical and ECG variables improved the F1 score by approximately 5% for both top-performing models: random forest reached a balanced accuracy of 0.790±0.015 and F1 score of 0.704±0.021 (sensitivity 0.797±0.041, specificity 0.783±0.042), while logistic regression achieved a balanced accuracy of 0.776±0.012 and F1 score of 0.684±0.013 (sensitivity 0.794±0.050, specificity 0.757±0.039). Comparison with the ESC risk score was performed using receiver operating characteristic (ROC) curves, shown in Figure 1 alongside the curves for all four models.

Combining echocardiographic, clinical, and ECG features enhances risk prediction in HCM. The logistic regression model, in particular, offers a practical and interpretable tool for clinical settings while maintaining good predictive performance. Moreover, the automatically and hypothesis-free selected features provide valuable insights into risk stratification in this specific population, contributing to a better understanding of the underlying determinants of adverse outcomes.

Mean ROC curves (5-CV) comparison.