Machine learning algorithms for predicting arrhythmic events in hypertrophic cardiomyopathy: limited enhancement beyond late gadolinium enhancement

Introduction and objectives

The optimal tool for predicting arrhythmic risk in Hypertrophic cardiomyopathy (HCM) remains a topic of ongoing debate. Artificial intelligence techniques, particularly machine learning (ML) predictive modelling, hold promise for improving risk stratification. The purpose of this study was to develop and assess the performance of a ML model integrating common clinical features to predict arrhythmic events in patients with HCM.

Methods

We conducted a post-hoc analysis of an international multicenter registry of 530 HCM patients (median age of 49 years (IQR 35-61), 57% male) who underwent cardiac magnetic resonance (CMR) for diagnostic confirmation and risk stratification. The dataset comprised clinical, echocardiographic, and CMR variables, including quantification of late gadolinium enhancement (LGE) using the 6 SD method. The study endpoint was a composite of sudden cardiac death (SCD), aborted SCD, and sustained ventricular tachycardia (VT). A total of 28 events (15 SCDs, 6 aborted SCD and 7 sustained VTs) were accrued over a median follow-up of 4.1 (IQR 1.8-7.3) years.

Using these data, several ML models [including Logistic Regression, Decision Trees, Gradient Boosting Machines, Support Vector Machines and Random Forest (RF)] were developed to predict the study endpoint. The predictive performance of the best model was then compared to the ESC HCM risk score and to the amount of LGE.

Results

After testing several models, the RF was the most effective method. Key predictive features included LGE percentage, left ventricular ejection fraction, left atrial diameter, and left ventricular indexed mass – Figure 1A. After 5-fold cross-validation, the RF model showed good performance for predicting arrhythmic events, achieving a time-weighted AUC of 0.78 (95% CI: 0.76–0.82, p<0.001).

This performance substantially outperformed the ESC HCM risk score, which achieved a time-weighted AUC of 0.64 (95% CI 0.62-0.67; p < 0.001), p<0.001 for comparison. However, when compared to LGE alone, which attained a time-weighted AUC of 0.76 (95% CI: 0.73–0.84, p<0.001), the RF model provided only a modest, statistically non-significant improvement in predicting the study endpoint (p = 0.817 for comparison).

Conclusion

A machine learning model using readily available clinical variables significantly outperformed the ESC HCM risk score in predicting arrhythmic events in HCM. However, its incremental value over LGE alone was modest, underscoring the strong predictive value of this imaging biomarker. Future research exploring AI-driven image analysis and other innovative approaches may yield better results.