Quick Heart Test May Hold Hidden Clues to Diabetes and Kidney Disease, New AI Finds

A five-minute heart test that millions of people get every year at a doctor’s office could soon do far more than check for heart problems. According to new research from Scripps Research, that same recording, run through a newly built artificial intelligence system, can flag the risk of death within 30 days of an emergency room visit or surgery, along with the longer-term odds of developing chronic kidney disease or type 2 diabetes. The tool needed far less labeled training data to get there than the comparison models the researchers tested.

Called ECG-CLIP, the system was designed to read electrocardiograms, the squiggly-line tests that track the heart’s electrical activity. But instead of only flagging heart attacks or irregular rhythms, it was trained on the actual written notes cardiologists attach to those tests, not just the raw signal. That combination let it pick up patterns that go well beyond typical heart disease detection, including risks tied to organs that have nothing to do with the heart’s wiring. The work appears in the journal The Lancet Digital Health.

Heart disease remains one of the leading causes of death worldwide, and a large share of it strikes older adults. Catching problems early matters, yet doctors reading these tests by eye can miss subtle warning signs, especially when they’re short on time or comparing scans taken years apart. Researchers have been trying to build AI models that catch what a busy clinician might overlook, and this one claims to do that while needing a fraction of the labeled examples earlier tools required.

How Scientists Trained This AI Heart Test Tool

Researchers pulled together more than 1.7 million heart tests from over 500,000 patients treated at Scripps Health between 2008 and 2019, each one paired with the final written interpretation a cardiologist had signed off on. That pairing, the recording alongside a doctor’s plain-language description of what it showed, became the training material.

Training happened in two stages. First, the model learned to fill in missing pieces of a heart recording after part of it was hidden, similar to teaching someone to finish a puzzle so they understand the shape of the pieces. Second, it learned to match those recordings with the matching doctor’s notes, building an internal sense of which patterns in the electrical signal line up with specific medical language.

Once trained, the model was fine-tuned and tested on data it had never seen before: an outside collection called MIMIC-IV, containing more than 800,000 heart tests from over 161,000 patients. This step matters because a tool that works well only in the hospital system that built it isn’t especially useful elsewhere. Testing on a separate patient population lets researchers check whether a model works broadly rather than just memorizing quirks from one dataset.

To see how it stacked up, the researchers compared ECG-CLIP against several other systems, including standard machine learning models built for single tasks, a general-purpose AI system not designed for medical signals, and three other recently published heart test AI models. Accuracy was measured on a scale where 1.0 is a perfect score and 0.5 is a coin flip, letting researchers directly compare how well each tool told sick patients from healthy ones.

What This AI Heart Test Revealed

ECG-CLIP beat the other models at detecting heart attacks, a rare thickening disease of the heart muscle called hypertrophic cardiomyopathy, and a rare protein-buildup condition called cardiac amyloidosis, even though amyloidosis and cardiomyopathy cases made up less than half a percent of the training data. It also outperformed rivals at predicting which patients with a currently normal heart rhythm would go on to develop atrial fibrillation, the most common heart rhythm disorder.

Where the model really stood out was how little labeled data it needed. When researchers tested it using just ten confirmed cases of a disease, an extremely small sample, ECG-CLIP still beat its closest competitor on every task tested: heart attack detection, amyloidosis detection, cardiomyopathy detection, and atrial fibrillation prediction. It matched what other top-performing tools could do with the full dataset while using anywhere from about 83 percent to more than 92 percent less training data, depending on the task.

Beyond heart disease, the model predicted death within 30 days after an emergency department visit or surgery, plus the three-year onset of chronic kidney disease and Type 2 diabetes, all from a single ECG. It also outperformed comparison models on every one of these predictions. As the researchers put it in the paper, “ECG-CLIP enables accurate prediction of cardiovascular disease and several adverse health outcomes across diverse clinical contexts.”

One more finding could matter for wearable health tech. Portable devices and smartwatches typically capture only one or two of the twelve electrical “leads” a full hospital ECG machine records. When researchers restricted the model to individual leads, ECG-CLIP still outperformed competing models at detecting certain heart attacks, even when it had to rely on leads that normally carry less useful information. The researchers say that resilience could eventually prove useful for portable or wearable devices, though the model has not yet been tested on data from a smartwatch or similar device.

Why This AI Heart Test Breakthrough Matters

None of this means an ECG is about to replace blood tests or genetic screening for diabetes or kidney disease. Building this kind of tool into everyday care would require much more testing, including trials across multiple hospitals, different patient populations, and different recording equipment, all of which the researchers themselves flagged as necessary next steps.

Still, the core finding is worth sitting with: a decades-old, widely available test that costs little and requires no needles appears to carry hidden signals about a person’s future health well beyond the heart, and a machine can now learn to read those signals from far fewer examples than anyone thought possible. That mix of broader predictive reach and lower data requirements is what could eventually make tools like this practical in hospitals, clinics, and wearable devices that currently have no way to catch these risks at all.