Abstract 4342247: Machine learning of echocardiographic left ventricular myocardial radiomics for identification of transthyretin amyloid cardiomyopathy

Background: Transthyretin amyloid cardiomyopathy (ATTR-CM) is an important cause of heart failure (HF) characterized by increased left ventricular (LV) wall thickness that can be challenging to distinguish from other LV hypertrophic processes. Myocardial radiomic texture analysis using echocardiographic images may identify subtle differences in tissue structure, invisible to the human eye, that result from amyloid deposition. Our aim was to train a machine learning model utilizing echo-derived radiomics features to identify amyloid cardiomyopathy. Methods: Retrospective multi-center study of echocardiographic images from 749 patients: 200 with ATTR-CM and 549 with non-amyloid HF. Conventional radiomic features such as histogram, two-dimensional, gray-level co-occurrence matrix, and gray-level run-length matrix, in addition to the novel application of chi-square, gray-level gradient matrix, and Laws’ texture features, were extracted from parasternal short-axis views at 3 cardiac levels (base, mid, apex). Data preprocessing via median imputation and z-score standardization was performed. A machine learning pipeline was developed and implemented using a Random Forest (RF) classifier with Least Absolute Shrinkage and Selection Operator (LASSO) in Python. A balanced training set was randomly subsampled (150 ATTR-CM and 150 non-amyloid HF) with hyperparameter tuning. Model performance was evaluated on an independent testing set of real-world prevalence comprising the remaining 50 ATTR-CM and 399 non-amyloid HF samples (conferring 11.1% ATTR-CM prevalence). Results: Using the realistic low-prevalence testing dataset, the machine learning model achieved a sensitivity of 86%, specificity of 92%, positive predictive value (PPV) of 57.3%, and negative predictive value (NPV) of 98.1%. The F1 score was 0.92, respectively. Overall accuracy was 0.91. The area under the receiver operating characteristic curve (ROC-AUC) was 0.938. Analysis of features incorporated into the model demonstrated that all 3 LV levels contributed and that novelly applied features sensitive to various local texture patterns (lines, edges,&spots) highly contributed to discrimination performance. Conclusion: We have developed a novel tool using echocardiographic radiomics that differentiates ATTR-CM from non-amyloid HF with high precision. The high NPV has clinical utility to exclude ATTR-CM, while further refinement using available demographics should increase performance.