An algorithm reading a photographed ECG can put cardiac amyloidosis on the list

Design

Diagnostic test study: AI-ECG model developed at one health system, temporally validated, then externally validated in five multinational and three screening cohorts

Population

Development 28,174 ECGs from 11,291 patients (293 with amyloid); internal validation 44,123 patients, mean age 68.5, 49.7% female

Primary outcome

Discrimination of transthyretin amyloid cardiomyopathy confirmed by radionuclide imaging or biopsy

Effect

Internal AUROC 0.84 (95% CI 0.79-0.89), sensitivity 0.72, specificity 0.86; external 0.78-0.89; sequential AI-echo raised PPV 0.24 to 0.66, cutting sensitivity 0.84 to 0.68

Transthyretin amyloid cardiomyopathy is treatable and routinely missed, because the diagnosis requires someone to think of it and then order imaging that is not everywhere available. This study built an artificial-intelligence model that reads routine ECG images - photographs or scans of a paper trace, not just raw digital signals - and asks whether that patient warrants confirmatory evaluation.

The model was developed on 28,174 ECGs from 11,291 patients, of whom 293 had confirmed amyloid cardiomyopathy, then validated temporally, across five multinational cohorts and in three screening cohorts. Internal validation in 44,123 patients gave an area under the curve of 0.84 (95% CI 0.79-0.89), with sensitivity 0.72 and specificity 0.86 at the prespecified threshold. External performance held between 0.78 and 0.89. Crucially, it survived a stress test against the conditions that mimic amyloid on an ECG - left ventricular hypertrophy and severe aortic stenosis without amyloid - at 0.81.

A second finding is about how to use it. Running the ECG model and then an AI-enabled echocardiogram raised positive predictive value from 0.24 to 0.66, but cut sensitivity from 0.84 to 0.68. That is the classic screening trade and it should be made deliberately: a sequential pathway sends fewer people for expensive nuclear imaging and misses more cases.

The reason this belongs on a general desk is the deployment claim. A model that works on an image of a paper ECG can run where there is no digital ECG archive and no cardiac MRI - which describes most district hospitals in India. It is retrospective work and the authors say calibration and intended use need prospective study, so this is not a test to start ordering. It is a reason to keep amyloid on the differential in an older patient with heart failure, hypertrophy on the ECG that does not fit, and a history of carpal tunnel surgery.

• Think of transthyretin amyloidosis in heart failure with unexplained ventricular hypertrophy
• Ask about bilateral carpal tunnel surgery and lumbar canal stenosis - both precede the cardiac diagnosis by years
• A sequential AI pathway trades sensitivity for precision; decide which error you can live with
• Do not treat an AI-ECG score as a diagnosis - confirmation is bone scintigraphy or biopsy
• This is retrospective validation; prospective calibration is still outstanding

The statistics, in plain English

An area under the curve of 0.84 means that given one patient with amyloid and one without, the model ranks them correctly 84% of the time. Sensitivity 0.72 with specificity 0.86 means it misses about a quarter of cases and falsely flags about one in seven of the rest - and because the disease is uncommon, most positive results will still be false, which is why the positive predictive value starts at 0.24. Adding a second test raises that to 0.66 by discarding cases: the honest way to describe sequential screening is that it makes each referral more likely to be right and each missed case more likely to stay missed.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.