224 An actionable, explainable, and biologically plausible AI-ECG risk estimation platform for diabetes mellitus

Background

With the rising incidence of Type 2 Diabetes Mellitus (T2DM) and the number of undiagnosed cases, there is an urgent need for innovative strategies for early identification of individuals at higher risk. To address this, we explore the utility of deep learning applied to 12-lead electrocardiograms (ECGs) for predicting the risk of incident T2DM in non-diabetic individuals, offering a novel approach for early detection and risk stratification.

Methods

The AI-ECG model, developed on the Beth Israel Deaconess Medical Center (BIDMC) dataset of 1.1 million ECGs and externally validated in the UK Biobank (UKB, N = 65,606), employs a residual neural network architecture tailored for a discrete-time survival model. Model performance was evaluated using the concordance index (C-index), and its enhancement of traditional risk factors was assessed via likelihood ratio tests (LRT) and net reclassification index (NRI). We also explored associations with clinical and echocardiographic features through a phenome-wide association study (PheWAS), and with genetic loci through a genome-wide association study (GWAS).

Results

The model predicted future T2DM in non-diabetic outpatient individuals with a C-index of 0.666 (0.658–0.675) in BIDMC and 0.689 (0.663–0.715) in UKB. The model showed consistent performance in both sexes, across ethnic groups, and BMI categories, except for patients aged ≥ 65. An improved performance was noted in individuals aged < 65, with a C-index of 0.691 (0.681, 0.701) and 0.765 (0.730, 0.797) in UKB. Adding the AI-ECG model to age, sex, BMI, and ECG parameters significantly enhanced predictive accuracy in the BIDMC cohort (p < 0.0001). Similarly, adding the model to the American Diabetes Association (ADA) risk score in the UKB substantially improved predictive accuracy (p < 0.0001). The continuous Net Reclassification Improvement (NRI) was 0.30 (0.22–0.40) for the BIDMC and 0.35 (0.21–0.47) for the UKB. The PheWAS and echocardiographic analyses identified significant associations between model predictions and a range of cardiac and non-cardiac phenotypes, including lipid profiles, glycaemic control, blood pressure, as well as echocardiographic measures of cardiac structure and function. This was substantiated by the GWAS study, highlighting genes associated with left ventricular structure, left atrial function, myocardial mass, blood pressure, T2DM, and HbA1C.

Conclusion

We have developed an AI-ECG model capable of predicting the risk of future T2DM in non-diabetic outpatient populations, validated in both primary and secondary care cohorts. The model enhances T2DM risk prediction and stratification when integrated with traditional risk factors and scores. Its application in primary care settings holds promise for the early identification of individuals at higher risk of T2DM, enabling timely interventions and personalised management.

Conflict of Interest

None