Diagnosis of transthyretin amyloid cardiomyopathy using automated myocardial segmentation and CT-based radiomic phenotyping

Background Transthyretin amyloid cardiomyopathy (ATTR-CM) is an underdiagnosed cause of heart failure. Disease diagnosis remains challenging, as it requires a high index of clinical suspicion, disease-specific imaging modalities, and, in selected cases, histological confirmation. Aim We aimed to develop and validate an automated end-to-end pipeline for the non-invasive detection of ATTR-CM by routinely performed Computed Tomography Angiography (CTA) scans. Methods A total of 999 CTA scans were analyzed, including patients with ATTR-CM (n=27), AS and negative bone scintigraphy (n=209), various cardiomyopathy subtypes (n=51), ischemic heart disease (n=211) or no known structural heart disease (n=501). Left ventricular myocardium was automatically segmented using a previously externally validated deep-learning model developed in-house (1). Radiomic features (n=843) were automatically extracted using PyRadiomics and subjected to preprocessing and feature selection performed exclusively on the training set to avoid information leakage. Models were tested using a stratified 50/50 train-test split, with 5-fold cross-validation applied within the training set. Performance was assessed on an independent test set using ROC and precision-recall (PR) analyses, with additional threshold-based evaluation. Results In the initial multiclass analysis, radiomic feature distributions demonstrated varying phenotypic patterns across disease categories, indicating disease-specific myocardial texture signatures (Figure panel A). For ATTR-CM, the GLMNET model demonstrated strong discrimination of ATTR-CM patients vs. all other subjects (ROC AUC 0.92, Figure panel B). Model-predicted probabilities demonstrated clear separation between ATTR-CM and non-ATTR cases in both training and independent test cohorts (Figure panels C and D). Given the low disease prevalence (~2.7%), PR analysis confirmed substantial enrichment above baseline (PR AUC 0.26). At the maximum F1 threshold, GLMNET achieved a positive predictive value (PPV) of 57%, identifying 4 true ATTR cases among 7 flagged patients, with a sensitivity of 33.3% (Figure panel E). In clinical scenarios where minimizing missed ATTR-CM diagnoses is critical, we evaluated the GLMNET classifier at the threshold corresponding to maximum balanced accuracy. At this operating point, the model correctly identified 11 of 12 ATTR patients (sensitivity 91.7%), while flagging 82 patients overall, resulting in a positive predictive value of 13.4% (Figure panel F). Notably, only one ATTR case was missed, underscoring the suitability of this threshold for screening applications, where sensitivity is prioritized over diagnostic precision. Conclusion A CT-based myocardial phenotyping approach enabled detection of ATTR-CM from routinely performed CTAs. This framework supports a layered, non-invasive screening strategy that may complement existing diagnostic pathways and facilitate earlier identification of ATTR-CM.