Amsterdam Cardiovascular Sciences

Missing a treatable disease

In ATTR‑CM, misfolded transthyretin proteins accumulate in the heart muscle, making the ventricle stiff and impairing its ability to pump. Disease‑modifying therapies are now available, but many patients are only diagnosed once irreversible cardiac damage has occurred. Confirmation of cardiac amyloidosis typically relies on nuclear radionuclide imaging or biopsy: tests that are costly and not universally accessible. By contrast, 12‑lead ECGs are cheap, ubiquitous, and minimally burdensome. They contain subtle signatures of cardiac amyloidosis, yet these are extremely difficult to recognize with the naked eye. The study asked whether AI could uncover this hidden information and turn a routine test into a practical triage tool for ATTR‑CM.  

AI reads ECG images

The researchers developed deep learning software that analyses ordinary ECG images, such as PDFs and scanned tracings, as well as raw digital signals. The model was trained on tens of thousands of ECGs from the Yale New Haven Health System, including almost 300 patients with confirmed ATTR‑CM based on cardiac amyloid radionuclide imaging or treatment with transthyretin stabilizers.  

In a large internal validation cohort of more than 44,000 patients, the AI‑ECG model achieved strong discriminative performance and maintained accuracy in patients with conditions that often mimic amyloidosis, such as severe aortic stenosis and marked left ventricular hypertrophy. “Clinicians already use these ECG images every day,” says first author Philip Croon, cardiology resident and PhD candidate at Amsterdam UMC and Amsterdam Cardiovascular Sciences: “This study shows that the same images can carry meaningful diagnostic information about cardiac amyloidosis, without the need to rebuild hospital IT systems.” 

International validation 

To test robustness, the team deployed the AI‑ECG pipeline in eight external cohorts from nine hospitals in the United States, United Kingdom, the Netherlands, Denmark and Greece. Despite differences in patient populations, ECG vendors and formats, the model consistently distinguished ATTR‑CM from controls, with performance metrics comparable to the internal results.  

Focus on higher-risk groups

Because ATTR‑CM is rare, broad screening of all ECGs would very likely lead to many false positives. The study therefore focused on higher‑risk groups: older Black and Hispanic adults with heart failure, people with previous carpal tunnel surgery and carriers of transthyretin gene variants. In these enriched populations, the AI‑ECG software achieved higher positive predictive values and proved more suitable as a targeted screening aid rather than a universal test. “For rare diseases, context is everything,” Croon explains. “Our data suggest that ECG‑based AI is most powerful in groups where clinical suspicion should already be higher, helping clinicians decide who needs specialized imaging and who can safely be reassured.” 

Towards smarter pathways

The authors emphasize that AI‑ECG is not a stand‑alone diagnostic test and cannot replace echocardiography or nuclear imaging. Instead, it may serve as an accessible first step to prioritize patients for further evaluation, particularly where advanced imaging capacity is limited. In exploratory analyses, combining AI‑ECG with an AI‑echocardiography model further increased the positive predictive value and reduced the number of nuclear scans per true diagnosis, at the cost of sensitivity.  

The algorithm has been recognized by the US FDA as a “breakthrough device”, a status that gives priority to promising technologies during the regulatory review process. A prospective study needed for full FDA approval is underway. “International collaboration and diverse datasets were crucial to demonstrate that this works across different health systems,” says Croon. “The next step is to test how integrating AI‑ECG into real‑world care pathways changes diagnostic yield, resource use and, ultimately, patient outcomes.”