Researchers in Japan have developed a new artificial intelligence model that can detect prediabetes based on electrocardiograms (ECGs), without the need for blood tests. The model, DiaCardia, analyses both traditional 12-lead ECGs and single-channel ECGs. This offers the prospect of low-threshold, large-scale screening, for example via wearable consumer devices such as smartwatches.
Prediabetes is an important precursor to type 2 diabetes. At this stage, blood sugar levels are already elevated, but not yet high enough for an official diagnosis. It is precisely during this period that timely intervention, for example through lifestyle changes, can significantly delay or even prevent the development of diabetes. In practice, however, prediabetes often goes unnoticed. This is partly because people experience few symptoms, do not participate in regular health checks, and because blood tests involve costs and logistical barriers.
ECG as an alternative screening route
An ECG is routinely used to measure the electrical activity of the heart. Prediabetes increases the risk of cardiovascular disease, which supports the hypothesis that subtle changes in ECG signals may indicate impaired glucose regulation. Previous studies combined ECG data with additional patient characteristics, such as age and gender, to detect diabetes. However, until now, there has been no reliable model that can identify prediabetes based solely on ECG signals.
The new research was conducted by a team led by Chikara Komiya, together with Ryo Kaneda and Tetsuya Yamada from the Institute of Science Tokyo. They developed DiaCardia based on the LightGBM machine learning algorithm, after comparing it with several other models.
For training and validation, the researchers used 16,766 health check data from a single clinic in Tokyo. This dataset included fasting plasma glucose values, HbA1c measurements and standard 12-lead ECGs. Prediabetes or diabetes was diagnosed when at least one criterion was met, such as elevated glucose or HbA1c levels or ongoing diabetes treatment. In total, the model analyses 269 characteristics from the ECG waveforms.
Good results
In internal tests, DiaCardia achieved an area under the receiver operating characteristic curve (AUROC) of 0.851 in recognising prediabetes, based solely on ECG data. The model also continued to perform well in external validation with data from another institution, without retraining. This indicates a high degree of generalisability.
Using Shapley additive explanations (SHAP), the researchers analysed which ECG characteristics contributed most to the predictions. Among other things, higher R-peak amplitudes in specific ECG leads and reduced heart rate variability proved to be important indicators. These characteristics are known from the literature and are associated with insulin resistance and autonomic neuropathy, which supports the physiological plausibility of the model.
Future of home screening
A striking result is that DiaCardia delivers virtually comparable performance with single-channel ECG data (lead I) as with 12-lead ECGs. This is relevant for applications outside the hospital, for example via wrist-worn wearables. This creates a real possibility for home screening of prediabetes without blood sampling.
‘This is the first robust, interpretable, and generalisable AI model capable of identifying individuals with prediabetes using either 12-lead or single-lead ECG data,’ says Komiya. He expects DiaCardia to make prediabetes screening ‘scalable, accessible and available at any time, without blood tests’. According to the researchers, widespread use of this technology could ultimately contribute to better diabetes prevention and reduce the burden of disease.
Breath test
Last year, researchers developed an innovative breath test that could potentially detect diabetes and prediabetes easily. The technology measures the acetone content in exhaled air, an important biomarker for disturbances in fat metabolism. The sensor offers a fast, non-invasive and affordable alternative to blood tests and makes early screening more accessible and better suited for population screening and self-monitoring.
The sensor is based on laser-induced graphene combined with zinc oxide, which allows acetone to be reliably measured even in moist breath. Acetone concentrations are often elevated in diabetes; values above 1.8 ppm may indicate this. The researchers are working on integration into portable applications, such as wearables, to enable continuous and personalised metabolic monitoring.
