AI Model Accurately Identifies Prediabetics Using Only ECG Data

Low participation in routine health checkups and the cost and invasiveness of laboratory testing further limit early detection. Yet this stage represents a critical window when lifestyle changes can prevent progression to type 2 diabetes. Researchers have now shown that subtle patterns in heart electrical activity can be used to identify people with prediabetes accurately, using only ECG data.

Researchers at the Institute of Science Tokyo (Tokyo, Japan), in collaboration with academic partners, have developed an artificial intelligence (AI) model called DiaCardia designed to detect prediabetes using either standard 12-lead ECGs or simplified single-lead ECG recordings. The goal was to create a noninvasive, scalable screening tool that could eventually work with consumer wearable devices.

DiaCardia is built on the LightGBM machine learning framework and analyzes waveform-level ECG features without relying on demographic inputs such as age or sex. The model extracts hundreds of quantitative features from ECG signals that reflect cardiac structure and autonomic regulation, both of which are known to be affected early in metabolic disease.

To train and evaluate the model, the researchers analyzed 16,766 health checkup records from a Tokyo clinic, each including fasting glucose, HbA1c values, and 12-lead ECG data. Prediabetes or diabetes was defined using standard clinical thresholds or ongoing diabetes treatment. In internal testing, DiaCardia achieved an AUROC of 0.851 using ECG data alone.

The findings, published in Cardiovascular Diabetology, show that the model also performed well in an external validation cohort from another institution without retraining, demonstrating strong generalizability. Explainable AI analysis showed that higher R-wave amplitudes and reduced heart rate variability were key predictors, aligning with known physiological effects of insulin resistance and autonomic dysfunction.

Notably, DiaCardia retained nearly the same predictive accuracy when applied to single-lead ECG data using far fewer features. This suggests the model could be deployed through wrist-worn wearables or home ECG devices, enabling large-scale, low-cost screening. Such an approach could significantly expand access to early prediabetes detection beyond clinical settings.

The researchers plan to further validate the model in broader populations and explore real-world integration with consumer ECG platforms. If confirmed, the technology could support continuous, noninvasive monitoring and earlier preventive interventions.

“This is the first robust, interpretable, and generalizable AI model capable of identifying individuals with prediabetes using either 12-lead or single-lead ECG data,” said Junior Associate Professor Chikara Komiya, who led the research team. “DiaCardia has the potential to make prediabetes screening scalable, accessible, and available anytime, anywhere, without a blood test. By promoting widespread screening of prediabetes, this work will ultimately contribute to the prevention of diabetes.”

Related Links:
Institute of Science Tokyo