Artificial Intelligence-Enhanced Electrocardiography for the Prediction of Future Type 2 Diabetes Mellitus: a model-development and multicentre validation study

Figures

• 

  Figure 1. Performance summary of AI-ECG Risk Estimator for Diabetes Mellitus (AIRE-DM) in detecting prevalent type 2 diabetes mellitus (T2DM) and predicting future incident T2DM. Panel (A) illustrates AIRE-DM’s performance in predicting incident T2DM in individuals without T2DM, quantified by Harrell’s concordance index (C-index). Panel (B) shows its performance in detecting prevalent T2DM, quantified by the area under the receiver operating characteristic curve (AUC). In both panels, results are shown for the 12-lead and single-lead AIRE-DM models. Panels (C)–(E) display Kaplan–Meier curves for Beth Israel Deaconess Medical Center (BIDMC), UK Biobank, and ELSA-Brasil participants without T2DM, respectively, stratified into quartiles of predicted risk. In all three cohorts, individuals in the highest risk quartile had a significantly higher cumulative incidence of T2DM compared with those in the lowest quartile. Hazard ratios were adjusted for age and sex. The log-rank test was used to derive the P value for the differences in cumulative incidence of future T2DM across the risk quartiles.

• 

  Figure 2. Comparison of predictive performance for incident type 2 diabetes mellitus (T2DM): AI-ECG Risk Estimator for Diabetes Mellitus (AIRE-DM) vs. American Diabetes Association Diabetes Risk Test (ADA DRT) across demographic and ethnic subgroups. Panels (A)–(C) display Harrell’s concordance index (C-index) (95% confidence interval) for incident T2DM prediction by AIRE-DM (green), ADA DRT (blue), and the combined AIRE-DM and ADA DRT (red) across different demographic and ethnic subgroups in the Beth Israel Deaconess Medical Center (BIDMC) cohort, UK Biobank, and ELSA-Brasil, respectively.

• 

  Figure 3. Comparison of predictive performance of AI-ECG Risk Estimator for Diabetes Mellitus (AIRE-DM) and glycated haemoglobin (HbA1c) for incident type 2 diabetes mellitus (T2DM) across glycaemic strata and cumulative incidence by risk quartile. Panels (A) and (C) display Harrell’s concordance index (C-index) (95% confidence interval) for incident T2DM prediction by AIRE-DM alone (green), HbA1c alone (blue), and the combined model (red) across standard HbA1c categories in the Beth Israel Deaconess Medical Center (BIDMC) test set (A) and the ELSA-Brasil cohort (C). Nested Cox models were compared by likelihood-ratio test, and non-nested models by partial likelihood-ratio test. Panels (B) and (D) show Kaplan–Meier curves for participants with baseline HbA1c < 5.7%, stratified by AIRE-DM risk quartile in BIDMC (B) and ELSA-Brasil (D), adjusted for continuous HbA1c, age, and sex. Differences between risk quartiles were evaluated by the log-rank test.

• 

  Figure 4. AI-ECG Risk Estimator for Diabetes Mellitus (AIRE-DM) is independently associated with type 2 diabetes mellitus (T2DM)-related adverse outcomes. In adjusted Cox models, the AIRE-DM score independently predicts T2DM-related adverse outcomes in subjects from the Beth Israel Deaconess Medical Center (BIDMC) cohort (A), regardless of whether they have prevalent T2DM. For the UK Biobank (UKB) cohort (B), the score is predictive for individuals without prevalent T2DM, but not significant for those with prevalent T2DM. Covariates for BIDMC analysis included age, sex, systolic and diastolic blood pressures, smoking status, prevalent hypertension, hyperlipidaemia, and ethnicity. UKB analyses additionally included body mass index as a covariate. Hazard ratio refers to one standard deviation of AIRE-DM score.

• 

  Figure 5. Explainability analysis. Two explainability approaches were applied to explore the electrocardiogram (ECG) morphologies associated with AI-ECG Risk Estimator for Diabetes Mellitus (AIRE-DM) predictions. Variational autoencoder-derived latent features were analysed using linear regression to predict AIRE-DM outputs, with the three most significant features shown, highlighting abnormalities in QRS morphology and T wave morphology, including T wave inversion and biphasic T waves, as being most associated with high diabetes mellitus risk. Additionally, median ECG waveforms (mean ± SD, shaded region) were generated for the 10 000 highest and lowest predicted survival ECGs from the Beth Israel Deaconess Medical Center test set, identifying poor precordial R wave progression as the feature most strongly associated with a high risk of future type 2 diabetes mellitus (T2DM).

• 

  Figure 6. Phenotypic associations of AI-ECG Risk Estimator for Diabetes Mellitus (AIRE-DM) predictions. Panel (A) illustrates the associations between AIRE-DM risk predictions and echocardiographic parameters in Beth Israel Deaconess Medical Center test set and panel (B) depicts the associations with cardiovascular magnetic imaging parameters in the UK Biobank. LA, left atrium; LV, left ventricle; LVEDD, LV end-diastolic diameter; LVEF, LV ejection fraction; LVESD, LV end-systolic diameter; LVOT, LV outflow tract; MV, mitral valve; TR, tricuspid regurgitation.

• 

  Figure 7. Genotypic associations of AI-ECG Risk Estimator for Diabetes Mellitus (AIRE-DM) predictions. A Manhattan plot of genomic loci associated with AIRE-DM scores. Nearest genes to significant single nucleotide polymorphisms are shown. The red line depicts the genome-wide significant threshold (P value < 5e-8).