Abstract Body (Do not enter title and authors here): Background
Undiagnosed diabetes and prediabetes present a significant global health challenge. Artificial Intelligence-enabled electrocardiography (AI-ECG) has shown promise in identifying subtle ECG changes in a wide range of subclinical diseases. Opportunistic ECG screening could identify prediabetic patients, enabling early interventions to prevent T2DM and adverse cardiovascular events.

Aims
To develop the AI-ECG Risk Estimator to diagnose prevalent T2DM and predict future T2DM (AIRE-DM)

Methods
AIRE-DM was trained on a real-world secondary care cohort from Beth Israel Deaconess Medical Center (BIDMC) of 1,163,401 ECGs and externally validated in the UK Biobank (UKB, N = 65,606). AIRE-DM employs a residual neural network architecture with a discrete-time survival loss function.

Results
AIRE-DM accurately identifies prevalent T2DM (AUROC: BIDMC – 0.712 (0.705-0.719), UKB - 0.731 (0.725 - 0.741) and predicts future T2DM (C-index: BIDMC - 0.666 (0.658-0.675), UKB 0.689 (0.663-0.715). In subjects without T2DM, the high-risk quartile shows a markedly increased risk of future T2DM (HR: BIDMC - 4.67 (4.01-5.45), UKB - 10.10 (5.87-17.40), adjusted for age and sex. Adding AIRE-DM to clinical risk factors in BIDMC and to the American Diabetes Association (ADA) score in the UKB significantly enhanced predictive accuracy for future T2DM (C-index improvement: BIDMC - 0.0359 (0.0354-0.0363), UKB: 0.0337 (0.0324-0.0350), continuous net reclassification index: BIDMC - 0.407 (0.360-0.445), UKB - 0.391 (0.259-0.503)).

Using phenome- and genome-wide association studies, we identified biologically plausible associations for AIRE-DM, including glucose regulation, cardiac morphology, diastolic dysfunction, arterial stiffness and lipid metabolism. We identified variants adjacent to CASQ2, TBX3, NOS1AP, TKT, VGLL2 and PRDM6, which are known regulators of cardiac morphology, arterial stiffness and glucose metabolism.

Conclusion
AIRE-DM can predict future T2DM in non-diabetics and enhances T2DM risk prediction when integrated with clinical risk scores. Its application holds promise for early identification of individuals at high risk of T2DM, enabling early lifestyle and pharmacological interventions.