PO-04-210 NEW-ONSET DIABETES SCREENING USING ARTIFICIAL INTELLIGENCE-ENHANCED ELECTROCARDIOGRAM

Abstract

Diabetes has a long asymptomatic period which can remain undiagnosed for years. Glycated hemoglobin (HbA1c) enables the discovery of asymptomatic diabetes. Yet, undiagnosed diabetes is present in 23.0% of all US adults with Diabetes and 3.4% of all US adults. American Diabetes Association (ADA) guidlines promote the use of risk tests that help identify candidates for diagnostic testing. To develop and evaluate an artificial intelligence (AI)-enhanced electrocardiogram algorithm (ECGAI-DM) for screening of Diabetes. We trained a neural network to estimate HbA1c using a 12-lead ECG and readily available demographics (ECG model), and evaluated its ability to screen for Diabetes in a complete outpatient population. The dataset comprised of patients (aged ≥18 years) with paired 12-lead, 10-second ECG and HbA1c test acquired during an outpatient encounter. Our primary outcome was the performance of the AI-enhanced ECG at classifying new-onset diabetes (HbA1c ≥ 6.5 in patients without a prior history of diabetes), which was assessed using the area under the receiver operator curve (AUC) and positive predictive value (PPV) at the sensitivity defined by the ADA risk test, with 95% bootstrap CIs. Population bias was addressed by reconstruction of a pseudopopulation with adjusted individual patient weights. We trained on 198,857 ECGs from 160,788 patients, and our test set comprised 34,106 ECGs from 30,593 patients, recorded between January 1, 2013 and September 17, 2021. The prevalence of Diabetes (HbA1c ≥ 6.5) was 4.9%. The ECGAI-DM significantly outperformed the ADA Risk test (AUC, 0.80 [95% CI, 0.79-0.81] vs. AUC, 0.68 [95% CI, 0.68-0.69]; p-value <0.001). The ECGAI-DM model restricted to using a single lead (Lead-I) outperformed the ADA risk test as well (AUC, 0.78 [95% CI, 0.78-0.79] vs. 0.68 [95% CI, 0.68-0.69]; p-value <0.001). Patients with false positive prediction by ECGAI-DM had a 3-fold risk for Diabetes over a 1 year follow up as compared to the true negative population, suggesting unmasking of latent high risk. In a separate test set, ECGAI-DM significantly outperformed 2 human Cardiologists (AUC 0.82 vs. 0.61, p<0.001) assessed on a class balanced 100 randomly sampled ECGs. An AI-enhanced ECG can automate the screening of patients for Diabetes with fewer false positives than questionnaire-based assessment. This also applies to single-lead ECGs commonly found on wearable devices, which suggests an avenue for community-wide opportunistic Diabetes screening.