Deep learning detection of cardiac amyloidosis from ECG images: validation across diverse clinical and prospective cohorts

Introduction Transthyretin amyloid cardiomyopathy (ATTR-CM) is increasingly treatable with disease-modifying therapy but remains underdiagnosed due to limited scalable screening strategies and reliance on resource-intensive confirmatory testing. Purpose To develop a format-agnostic artificial intelligence tool capable of detecting ATTR-CM from routinely acquired 12-lead ECG images and to validate its performance across multinational and prospective screening cohorts. Methods We designed an EfficientNet-B3 architecture-based model using ECG images to detect ATTR-CM in a large and diverse hospital system in the United States, and validated it across 8 multinational cohorts. External validation included two hospital systems in the United States, one in Greece, and the National Amyloidosis Center (NAC) in the UK. Prospective performance was assessed using scanned ECGs from SCAN-MP, a prospective screening cohort in older black and Hispanic adults with heart failure, and in two separate CACTUS cohorts in which individuals with prior carpal tunnel syndrome (CTS) were screened for ATTR-CM. The AI-ECG model was benchmarked against an established echocardiography-based risk score. Results Across eight cohorts encompassing 4,276 ATTR-CM cases, the model demonstrated strong discrimination. In a temporally distinct internal test set (n = 20,222), AUROC was 0.87 (95% CI 0.82–0.91) (Figure 1). Performance remained robust in ATTR-CM mimics with severe left ventricular hypertrophy or aortic stenosis on echocardiography, as well as in consecutive patients referred for amyloid nuclear scintigraphy. Across external validation cohorts, the model’s performance was good, including in 3,902 (2,595 amyloid cases) ECG images from patients referred to the NAC, where the AUROC was 0.89 (95% CI 0.87–0.90). In the prospective screening populations (n = 1,017), AUROC was 0.76 (95% CI 0.68–0.83) in older Black individuals with heart failure and 0.79 (95% CI 0.65–0.93) and 0.91 (95% CI 0.82–0.97) in separate cohorts including individuals with CTS surgery. Compared with an echocardiography-based risk score, the AI model achieved similar discrimination while requiring fewer confirmatory tests per true-positive case in broader clinical cohorts (Figure 2). Conclusion AI-ECG enables scalable detection of transthyretin amyloid cardiomyopathy using routinely acquired electrocardiogram images and demonstrated clinically meaningful performance when deployed in prospective screening cohorts. These findings support AI-ECG as a practical, frontline triage strategy to facilitate earlier identification of individuals for diagnostic testing.Figure 1FIgure 2