Photographing an ECG to Find Clues to Cardiac Amyloidosis: Yale AI Validated Across Eight Populations

As the heart gradually stiffens and ultimately progresses to heart failure, the underlying cause is sometimes protein deposits that have gone unrecognized for years. Because its symptoms resemble those of other cardiovascular diseases, cardiac amyloidosis often remains undiagnosed until complications develop. An artificial intelligence platform developed by a Yale University team seeks clues in routine ECG images to help physicians determine which patients should receive priority for further testing.

The study focuses on transthyretin amyloid cardiomyopathy (ATTR-CM). Transthyretin, a protein produced primarily by the liver, can misfold because of genetic factors or aging, forming amyloid deposits in the heart that stiffen the heart muscle and interfere with electrical conduction. The AI analyzes electrical signal patterns that these changes may leave behind; it does not directly visualize protein deposits.

An overview published by Yale School of Medicine on September 25 and a report on the same development published by Medical Xpress on September 30 state that physicians can access the platform through a smartphone and photograph an ECG for the model to analyze. This allows test results on paper or in image form to serve as an entry point for screening. The model is intended to direct suspected cases to echocardiography and confirmatory testing, rather than establish a definitive diagnosis from a single photograph.

A diagnostic study published online in *JAMA* on August 28 provides more specific validation data: model development used 28,174 ECGs from 11,291 patients, 293 of whom had ATTR-CM. In addition to validation using data from patients seen at a later period within the Yale system, the study included eight external cohorts in the United States and Europe, covering different care settings and screening populations, including older Black and Hispanic adults with heart failure and people who had undergone carpal tunnel surgery.

The model’s area under the receiver operating characteristic curve (AUROC) was 0.84 in internal validation and ranged from 0.76 to 0.91 across the eight external cohorts. This measure reflects the model’s ability to distinguish between people with and without the disease and cannot be interpreted directly as “accuracy.” At a prespecified threshold, sensitivity in internal validation was 72% and specificity was 86%, meaning that some patients would still be missed and some people without the disease would be flagged as suspected cases.

The results across cohorts therefore support its use as a first step in arranging further testing, but do not yet demonstrate that implementation can reduce heart failure or deaths. The study authors noted that real-world use and model calibration still require prospective evaluation. Yale also stated that the TRACE-AI Network Study is examining the implementation of multimodal AI tools at 13 medical centers in the United States, assessing how these approaches can identify ATTR-CM within larger-scale care workflows.

The regulatory status also requires a clear distinction. According to Yale and Medical Xpress, the tool has received Breakthrough Device designation from the U.S. Food and Drug Administration (FDA) and remains under review; this designation does not constitute marketing approval. Whether it can translate warning signs on an ECG into more timely diagnoses will depend on whether clinical workflows can accommodate flagged patients and whether subsequent testing identifies those who need treatment.