New-Onset Diabetes Assessment Using Artificial Intelligence-Enhanced Electrocardiography

TL;DR

This study develops an AI model that detects new-onset diabetes using a combination of 12-lead ECG signals and demographic data, explicitly adjusting for selection bias via inverse probability weighting to reflect the target population of ECG recipients. The authors extend FastSHAP to multi-modal inputs to enable efficient interpretability of the ECG-demographics model. The ECG-based approach achieves superior discrimination (AUC ≈ 0.80) and precision (PPV ≈ 0.11 at a matched threshold) compared with ADA risk and QDiabetes, and external validation confirms generalizability. Prospective analyses show higher future diabetes incidence among ECG-model high-risk individuals, supporting both screening and short-term prediction utility. The framework offers a scalable path for community and wearable-based diabetes screening and provides methodological guidance for addressing selection bias in retrospective clinical AI studies.

Abstract

Diabetes has a long asymptomatic period which can often remain undiagnosed for multiple years. In this study, we trained a deep learning model to detect new-onset diabetes using 12-lead ECG and readily available demographic information. To do so, we used retrospective data where patients have both a hemoglobin A1c and ECG measured. However, such patients may not be representative of the complete patient population. As part of the study, we proposed a methodology to evaluate our model in the target population by estimating the probability of receiving an A1c test and reweight the retrospective population to represent the general population. We also adapted an efficient algorithm to generate Shapley values for both ECG signals and demographic features at the same time for model interpretation. The model offers an automated, more accurate method for early diabetes detection compared to current screening efforts. Their potential use in wearable devices can facilitate large-scale, community-wide screening, improving healthcare outcomes.

Figures (9)

• Figure 1: Diagram of pseudo-population construction. The electronic health record provides a large amount of data with which to train and evaluate an AI-enhanced ECG to estimate HbA1c. However, for many patients ECGs or HbA1c tests are not performed. In order to understand how well the AI-enhanced ECG will work in practice, one needs to estimate the performance on the complete population. This diagram shows that by modeling the probability of ordering an HbA1c, the observed population can be re-weighted to represent the complete population.
• Figure 2: Kaplan-Meier analysis. High-risk group identified by the ECG Model showed significant higher cumulative incidence compared to the high-risk group identified by the ADA Risk Test.
• Figure 3: Sensitivity analysis. Even with significant violations made to MAR, the ECG model still outperformed both baselines.
• Figure 4: Shapley values quantifying feature contributions to model predictions.(a) Distribution of Shapley values for demographic features across test set, ranked by mean absolute value. Each point is one patient; color indicates feature value (red=high, blue=low). Positive values increase predicted diabetes probability. (b) Shapley values for three highest-scoring ECG samples. Top 20% absolute values (0.1s superpixels, all leads) overlaid on Lead I. Red segments increase predicted probability; blue decrease it.
• Figure A1: ECG model architecture. A 34-layer ResNet CNN (16 residual connections) processes 8-lead $\times$ 2,500 ECG input through temporal convolutions with batch normalization, ReLU, dropout, and max-pooling. The 128-dimensional ECG embedding is concatenated with demographic features and fed through three fully-connected layers (hidden-size 1000, 1000, 1000) to a softmax output layer predicting four HbA1c bins.