心電図の測定だけで糖尿病予備群を発見～血液検査なしで早期発見を可能にする新たなAI技術～

2025-11-112025-12-02

2025-11-11 東京科学大学

東京科学大学と東北大学の共同研究チームは、一般的な心電図データのみから糖尿病予備群を高精度に検出できるAIモデル「DiaCardia」を開発した。2022年に実施された健診データ16,766件を用いて構築し、勾配ブースティング決定木により学習。12誘導心電図から得られる269項目の特徴量のみで、糖尿病予備群/糖尿病をAUROC0.851(感度85.7%、特異度70.0%)の精度で識別した。別施設データによる外部検証でもAUROC0.785を示し、モデルの汎用性が確認された。さらに、腕時計型ウェアラブル端末で取得可能なI誘導心電図でも同等の精度(AUROC0.844)を達成し、血液検査を行わず「いつでも・どこでも」スクリーニングできる可能性を示した。AIはaVL誘導R波の高さや心拍変動を重要因子として判定しており、生理学的にも妥当性が確認された。本成果は糖尿病予防・早期発見の新たな手段として期待され、国際誌『Cardiovascular Diabetology』に掲載された。

図1.DiaCardiaは高い精度で糖尿病予備群を検出

<関連情報>

人工知能が単誘導心電図を用いて糖尿病前症の個人を特定する Artificial intelligence identifies individuals with prediabetes using single-lead electrocardiograms

Daisuke Koga,Ryo Kaneda,Chikara Komiya,Satoshi Ohno,Akira Takeuchi,Kazunari Hara,Masato Horino,Jun Aoki,Rei Okazaki,Ryoko Ishii,Masanori Murakami,Kazutaka Tsujimoto,Kenji Ikeda,Hideki Katagiri,Hideyuki Shimizu & Tetsuya Yamada
Cardiovascular Diabetology Published:11 November 2025
DOI:https://doi.org/10.1186/s12933-025-02982-4

Abstract

Background

Early detection of prediabetes is crucial for diabetes prevention, yet it remains challenging due to its asymptomatic nature and low screening rates. This study aimed to develop and rigorously validate artificial intelligence (AI) models to identify individuals with prediabetes solely using electrocardiograms (ECGs).

Methods

We defined prediabetes/diabetes based on fasting plasma glucose ≥ 110 mg/dL, hemoglobin A_1c ≥ 6.0%, or ongoing diabetes treatment. From a primary cohort of 16,766 health checkup records, 269 ECG features were extracted to develop a novel AI model. The final model was subsequently evaluated using an internal held-out test dataset and an independent external validation cohort (n = 2,456). SHAP (SHapley Additive exPlanations) was applied to assess feature importance and clinical interpretability.

Results

The best-performing model, a LightGBM-based algorithm we termed DiaCardia, achieved an area under the receiver operating characteristic curve (AUROC) of 0.851 in the internal test dataset (sensitivity: 85.7%, specificity: 70.0%). The model demonstrated robust generalizability, achieving an AUROC of 0.785 in the external validation cohort. Furthermore, DiaCardia maintained substantial predictive ability (AUROC: 0.789) after adjustment for six major confounders using propensity score matching. Higher R-wave amplitude in leads aVL and I, and smaller peak interval dispersion were prominent predictors. Notably, a version of DiaCardia using only single-lead (lead I) ECG data achieved a comparable AUROC of 0.844 (sensitivity: 82.3%; specificity: 70.2%).

Conclusions

This study establishes that an AI model, DiaCardia, can accurately identify individuals with prediabetes from an ECG alone, with performance that is robust across different patient cohorts and independent of major clinical confounders. Our highly generalizable, single-lead DiaCardia model offers a promising solution for scalable prediabetes screening via wearable devices, potentially enabling early, home-based detection and transforming diabetes prevention strategies.

Research insights

What is currently known about this topic?

Prediabetes is common, progresses unnoticed, and remains difficult to recognize early. ECG reflects systemic effects on the heart. AI enables advanced ECG analysis, but prediabetes detection is unproven.

What is the key research question?

Can an interpretable AI detect prediabetes from an ECG alone for broad clinical use?

What is new?

First demonstration of an interpretable AI model detecting prediabetes solely using ECGs. Single-lead version achieved accuracy comparable to 12-lead ECG. Key predictive features reveal cardiac pathophysiology associated with impaired glucose homeostasis.

How might this study influence clinical practice?