A new artificial intelligence model named DiaCardia has been developed by researchers at the Institute of Science Tokyo, Japan, capable of detecting prediabetes using only electrocardiogram (ECG) data—without the need for blood tests. This breakthrough could pave the way for widespread, non-invasive prediabetes screening through consumer wearable devices. Prediabetes, a condition where blood glucose levels are elevated but not yet in the diabetic range, is a critical window for preventing the progression to type 2 diabetes through lifestyle changes. However, early detection remains difficult due to the lack of symptoms, low participation in health screenings, and the cost and invasiveness of blood tests. ECG data, traditionally used to assess heart function, has shown promise in identifying early signs of metabolic disorders. Since prediabetes is linked to increased cardiovascular risk, changes in heart electrical activity may reflect underlying glucose dysregulation. Led by Junior Associate Professor Chikara Komiya, graduate student Dr. Ryo Kaneda, and Professor Tetsuya Yamada, the research team created DiaCardia, an AI model based on the LightGBM machine learning algorithm. The model was trained and tested on 16,766 health check-up records from a single clinic in Tokyo, including 12-lead ECGs, fasting plasma glucose (FPG), and hemoglobin A1c (HbA1c) levels. Prediabetes or diabetes was defined by meeting at least one of three criteria: FPG ≥110 mg/dL, HbA1c ≥6.0%, or current diabetes treatment. The model analyzed 269 waveform features extracted from ECG signals. In internal testing, DiaCardia achieved an area under the receiver operating characteristic curve (AUROC) of 0.851—indicating strong accuracy—in identifying prediabetes using ECG data alone. The model also performed well in external validation using data from another institution without retraining, demonstrating high generalizability. SHAP (Shapley additive explanations) analysis revealed that higher R-wave amplitudes in certain leads and reduced heart rate variability were key predictors. These findings align with known physiological effects of insulin resistance and autonomic neuropathy, supporting the model’s biological plausibility. Notably, DiaCardia maintained strong performance even when using only single-lead ECG data from lead I—just 28 features—achieving results nearly as accurate as those from full 12-lead ECGs. This is a major step toward integrating the technology into wrist-worn wearables like smartwatches. The researchers also confirmed the model’s predictive power after adjusting for six major confounding factors, showing that its performance stems from ECG features linked specifically to impaired glucose regulation. “This is the first robust, interpretable, and generalizable AI model capable of identifying prediabetes using ECG data alone,” said Komiya. “DiaCardia has the potential to make screening scalable, accessible, and available anytime, anywhere—without a blood test.” By enabling early, non-invasive detection, DiaCardia could significantly improve diabetes prevention efforts globally.