AI Model Identifies Prediabetes Using Only ECG Data from Wristwatch Wearable Devices Measurements

A research group led by Lecturer Chikara Komiya, Graduate Student Ryo Kaneda, and Professor Tetsuya Yamada from the Department of Molecular Endocrinology and Metabolism, Graduate School of Medical and Dental Sciences, Institute of Science Tokyo, along with Researcher Daisuke Koga (currently Assistant Professor at Saga University), Lecturer Satoshi Ohno, and Professor Hideyuki Shimizu from the Department of AI Systems Medicine, Institute of Science Tokyo, announced that they have successfully developed a new AI model called "DiaCardia" in collaboration with Professor Hideki Katagiri from Tohoku University Graduate School of Medicine. The model can identify individuals with prediabetes with high accuracy using single-lead (lead I) electrocardiogram (ECG) data equivalent to that obtained from wristwatch-type wearable devices. This research is expected to contribute to extending healthy life expectancy through earlier prevention. Their findings were published in the journal Cardiovascular Diabetology on November 11.

It is estimated that there are 10 million people with diabetes in Japan, and another 10 million with prediabetes. Diabetes progresses without noticeable symptoms and can lead to various serious cardiovascular complications. Because it is difficult to cure once it has developed, prevention is considered essential. Diagnosis of both diabetes and prediabetes requires blood tests to measure blood glucose and HbA1c levels. However, HbA1c may be excluded from some health checkups, meaning prediabetes cases may be overlooked. Epidemiological studies have reported that cardiac changes can occur even in individuals with prediabetes, increasing the risk of heart failure. However, conventional ECG and echocardiography examinations have been unable to detect these changes.

The research team therefore investigated whether precise analysis of ECG data obtained during health checkups using an AI model could detect cardiac changes that occur in individuals with prediabetes. Subsequently, they developed an AI model for this purpose.

From approximately 18,000 health checkup records of individuals aged 18 and older who underwent examinations at a single clinic in fiscal year 2022, and who had information on 12-lead ECG, fasting plasma glucose (FPG), HbA1c, and diabetes treatment status, 16,766 records which did not violate the exclusion criteria were selected for analysis.

A total of 269 features were extracted from 12-lead ECGs, and supervised learning was performed using machine learning models for classification prediction.

Subjects meeting at least one of the criteria — fasting plasma glucose ≥ 110 mg/dL, HbA1c ≥ 6.0%, or undergoing diabetes treatment — were classified as "prediabetes/diabetes", totaling 1,447 records, while the remaining 15,319 records were classified as normoglycemia. The supercomputer "SHIROKANE" at the University of Tokyo's Institute of Medical Science was used to develop the algorithm.

As a result, the machine learning model "DiaCardia", built using gradient boosting decision trees, successfully detected "prediabetes/diabetes" with an AUC (AUROC) of 0.851 (sensitivity: 85.7%; specificity: 70.0%) (Figure 1).

Analysis of which features contributed to the prediction results showed that R-wave amplitude in lead aVL and heart rate variability were particularly influential. These are associated with increased left ventricular mass accompanying insulin resistance, confirming that the predictions are pathophysiologically reasonable. When the threshold values were varied incrementally, prediction accuracy increased with rising blood glucose and HbA1c levels (Figure 2), suggesting the model is able to reflect cardiac changes. It was also confirmed that prediction accuracy was maintained regardless of which of the two major ECG device manufacturers used in Japanese medical institutions was employed.

Furthermore, it was discovered that features with high contribution included a particularly large amount of information obtainable from lead I ECG. Using only ECG features from lead I among the 12 leads, detection was possible with high accuracy — sensitivity of 82% and specificity of 70%. Since lead I can be obtained from wristwatch-type wearable devices, there is potential for detecting prediabetes at home.

Going forward, the team aims to verify whether similar accuracy can be achieved with actual wristwatch-type wearable devices, working toward a practical implementation in society.

Shimizu: "Wearable devices and the like are expected to have a lot of noise, but there are many signal processing techniques to clean up noise. By applying these, we aim to work toward practical application in smart devices in the future."

Komiya: "I don't think AI is superior to clinicians in every way, but as I began this research, I became increasingly interested in how AI can accomplish things that clinicians cannot. I believe that using AI will advance medicine, not just in this research. I hope to further develop our research, including work with wearable devices."

Yamada: "I have been treating diabetes for many years, and I want to work toward eliminating diabetes, at least in Japan. The driving force of progress in medicine and life sciences is technological advancement, and I believe AI is one such innovative tool. We want to stay ahead of technological innovations and continue contributing to the advancement of life sciences."

Journal Information
Publication: Cardiovascular Diabetology
Title: Artificial intelligence identifies individuals with prediabetes using single-lead electrocardiograms
DOI: 10.1186/s12933-025-02982-4