Major adverse cardiovascular events prediction in patients with suspected myocarditis using multimodal data and artificial intelligence

Abstract Background Predicting patient outcomes in myocarditis remains challenging, as current methods and subjective assessments often lack reliable prognostic indicators. This study proposes a AI-driven approach that integrates multimodal data to predict Major Adverse Cardiovascular Events (MACE) in myocarditis patients, aiming to enhance risk stratification and clinical decision-making. Method We included 1,385 consecutive patients with clinically suspected myocarditis from two registries, who were referred to cardiac MRI (CMR). Multimodal data at the diagnosis time point were collected, including cine CMR images and Late Gadolinium Enhancement (LGE), ECGs, clinical information, and lab tests. We developed survival models using three approaches (1) Handcrafted features: Machine learning (ML) models were trained on a tabular dataset with features from clinical, laboratory, electrocardiography, echocardiography, and CMR data. (2) Radiomics features: We automatically segmented the myocardium in LGE images and extracted radiomic features to train ML models. (3) End-to-end deep learning: A convolutional neural network (CNN) was trained using the DeepHit method to predict MACE risk directly from LGE images. The ML models considered were Cox proportional hazards model with Lasso and Ridge regularization (CoxNet), Random Survival Forest (RSF), and Gradient Boosting Survival model (GBS). For radiomics, we applied the ComBat harmonization algorithm to correct scanner-related variability. All models were trained on data from the first center (65.6%) and evaluated on the remaining patients from second center as external test set using Harrell’s c-index and cumulative-dynamic AUC Results Major adverse cardiac events (MACE) - a composite endpoint of death, sustained ventricular arrhythmia, heart failure hospitalization, and recurrent myocarditis- occurred in 178 patients (12.9%) over a median follow-up of 35 months (IQR 16, 35). Models trained on handcrafted features performed best, with a c-index of 0.71 and a 1-year AUC of 0.75 for both the CoxNet and GBS models. Performance was marginally inferior with radiomic features only, where the RSF model achieved the highest discrimination (c-index: 0.69, 1-year AUC: 0.73). The end-to-end deep learning model trained on LGE images showed similar performance, with a c-index of 0.67 and a 1-year AUC of 0.68. Conclusion Machine learning models trained on multimodal handcrafted features provided the best MACE prediction in myocarditis patients, with slightly lower performance using LGE radiomic features. While the deep learning model showed lower accuracy, it offers a fully automated process, making it a promising and scalable approach for future risk prediction.