Interpretable artificial intelligence-based transthyretin amyloid cardiomyopathy detection in patients with severe aortic stenosis

Abstract Background The concurrent presence of transthyretin amyloid cardiomyopathy (ATTR-CM) in candidates for transcatheter aortic valve implantation (TAVI) with severe aortic stenosis (AS) is frequently observed. Objective This study aims to develop a multimodal artificial intelligence-based model for ATTR-CM detection in patients with severe AS. Methods A total of 263 prospectively enrolled (57% males, 83±4.6 years) patients with severe AS were enrolled. All underwent DPD-scintigraphy, and coexisting ATTR-CM was confirmed in 27 (10.3%). 4D-cardiac computed tomography (CT), electrocardiogram, echocardiography, angiography, clinical, and laboratory data were assessed. Specifically, in the 4D-CT, LV myocardial radiomics features as well as global- and segmental myocardial strain. All features underwent z-score normalization, and those with high correlation were eliminated. Different feature selection and machine learning (ML) techniques were used for classification in single- and multimodality scenarios. Eighty percent of the dataset was used for pre-processing, feature extraction, and tuning, with the remaining 20% for evaluating the final model. The model's effectiveness was assessed with accuracy, F1 score, and ROC-AUC across 100 random seeds to establish a reliable classifier. The SHAP (SHapley Additive exPlanations) model is used to interpret the output of ML models. Results Among the different algorithms used, Logistic Regression offers the most effective detection of ATTR-CM, surpassing other models with AUC scores of 0.85. CT strain (AUC: 0.85), Radiomics (AUC: 0.76), and clinical features (AUC: 0.70) alone and in combination yielded the highest performance (i.e., AUC: >70), while the remaining modalities had AUC scores ranging from 0.52 to 0.56. Conclusion The study shows that AI is effective for non-invasively detecting ATTR-CM in severe aortic stenosis patients, offering a potential alternative to costly DPD-scintigraphy or invasive biopsy.