Multimodality artificial intelligence-based major adverse cardiovascular events prediction in transthyretin amyloid cardiomyopathy patients

Abstract Background Transthyretin amyloid cardiomyopathy (ATTR-CM) is a progressive condition that increases myocardial stiffness, compromising cardiac function and leading adverse cardiovascular events. Predicting adverse cardiovascular events is crucial for personalized treatment and management of patients. Purpose This study aims to develop artificial intelligence (AI)-based models for predicting major adverse cardiovascular events (MACE) in ATTR-CM by integrating multimodal tabular data and advanced machine learning techniques. Methods Multimodality data, including clinical, medication, demographic, laboratory, electrocardiography, and echocardiographic images, were collected at the time of diagnosis for each patient from four different center. We excluded features with more than 25% of missing, and features were imputed with MissForest algorithm. After normalizing continuous features subtracting the median and dividing by the interquartile range, we removed features with high correlation. Three machine learning models were fitted: Cox proportional hazards model with Lasso and Ridge regularization (CoxNet), Random Survival Forest (RSF) and Gradient Boosting Survival model (GBS). Hyperparameters tuning was performed on data coming from Center 1 (training set, 76% of patients) with a 5-fold cross-validation (CV) approach. For each model, the hyperparameter setting with higher CV Harrell’s c-index was chosen and used to fit the model on the training set. Models were then evaluated external validation from 3 different centers. We assessed model performance using metrics such as the Harrell’s c-index and cumulative-dynamic AUC. Results A total of 372 consecutive patients with ATTR-CM diagnosed based on bone scintigraphy Perugini grad 2-3 or biopsy, recruited from 4 centers in Switzerland, were included. In this cohort, 157 patients (41.1%) experienced a MACE event (i.e., death, sustained ventricular tachycardia, heart failure hospitalization, thromboembolic event, pacemaker or heart defibrillator implantation), with a median time to event of 612 days (IQR [311, 1014]). RSF and GBS showed the highest CV performance on the training set, both with an average c-index of 0.72. Similar performance was achieved by CoxNet model, with a c-index of 0.71. Performance was confirmed in the external test datasets with a c-index of 0.73 for RSF and GBS models, and 0.72 for CoxNet models. Regarding AUC, the 1-year AUC was 0.74, 0.69 and 0.73, respectively for RSF, GBS and CoxNet in external test sets. Conclusion In this study, we developed and externally evaluated an artificial intelligence-based model for predicting MACE in ATTR-CM patients using multimodal data, including imaging, signal, laboratory, and clinical information. The integration of multimodal data sources with machine learning algorithms enabled accurate MACE prediction in ATTR-CM patients, potentially supporting personalized treatment strategies.Kaplan-Meier curves stratified by risk