Adverse events prediction in nonischemic dilated cardiomyopathy by two machine learning models integrating clinical and cardiovascular magnetic resonance data

Funding Acknowledgements Type of funding sources: None. Background Risk stratification of patients with nonischemic dilated cardiomyopathy (NIDCM) is challenging. Machine learning (ML) models represent novel technologic tools that promise to improve predictive ability of cardiovascular adverse outcomes. Purpose To implement two ML models to predict the risk of adverse outcome (composite endpoint of all-cause death, heart transplantation, implantation of ventricular assistant device), in patients with NIDCM using clinical and radiological data. Methods We retrospective collected data from patients affected by NIDCM referred to our cardiologic unit between 2002 and 2018. Two ML models, an extreme gradient boosting machine (XGBoost) and an artificial neural network (ANN), were developed using cardiovascular magnetic resonance (CMR) and clinical parameters. Both models were trained on 67% of patients and evaluated based on the other 33% as internal validation. To investigate the contribution of different features in our models, we used Shapley additive explanations (SHAP) analysis. Results The final cohort resulted of 161 patients (mean age 48.9±13.6 years, 71% males), with a mean follow up duration of 79.1 months. During the follow-up 29 adverse event were observed (16 deaths, 12 heart transplantation, 1 implantation of ventricular assistant device). ANN model resulted in an accuracy of 76% with an area under the receiver operating characteristic (ROC) curve of 0.66; SHAP analysis identified male sex, percentage of late gadolinium enhancement (LGE) and LGE distribution in the lateral wall of myocardium as the main markers of adverse outcomes. XGBoost model resulted in an accuracy of 78% with a ROC curve of 0.76; SHAP analysis identified NT-proBNP level, age and serum creatinine mainly related with adverse outcomes. About CMR parameters, left ventricular diameter and ejection fraction better related with adverse outcomes. Conclusions Our models achieved a good performance in prediction of adverse events, suggesting that ML could improve our ability of risk stratification of NIDCM. XGBoost better performed than ANN, and SHAP analysis suggested that both CMR parameters and clinical parameters contribute to the definition of risk. ANN performance and SHAP analysisXGBoost performance and SHAP analysis