Opportunistic detection of type 2 diabetes using deep learning from frontal chest radiographs

Abstract

Deep learning (DL) models can harness electronic health records (EHRs) to predict diseases and extract radiologic findings for diagnosis. With ambulatory chest radiographs (CXRs) frequently ordered, we investigated detecting type 2 diabetes (T2D) by combining radiographic and EHR data using a DL model. Our model, developed from 271,065 CXRs and 160,244 patients, was tested on a prospective dataset of 9,943 CXRs. Here we show the model effectively detected T2D with a ROC AUC of 0.84 and a 16% prevalence. The algorithm flagged 1,381 cases (14%) as suspicious for T2D. External validation at a distinct institution yielded a ROC AUC of 0.77, with 5% of patients subsequently diagnosed with T2D. Explainable AI techniques revealed correlations between specific adiposity measures and high predictivity, suggesting CXRs' potential for enhanced T2D screening.

Fig. 1. Flowchart depicting the selection process for cases and controls from the prospective, external validation, and retrospective k-fold cohorts.

Exclusions were made sequentially based on ICD codes as presented in the flowchart (patients may have multiple exclusionary diagnoses but were counted only once according to their first exclusionary diagnosis). In the prospective and retrospective cohorts, patients with two or fewer claims for evaluation and management codes (CPT codes 99202 to 99499) and five or fewer unique encounter dates were excluded, as they may have received care in another health system. Patients diagnosed with type 1 diabetes (ICD9: 250.x1, 250.x3; ICD10: E10.x) and gestational diabetes (ICD9: 648.80–648.84; ICD10: O24.4x) were also excluded from all cohorts as potential confounders. The final number represents unique patients with a single conventional frontal chest radiograph.

Fig. 2. Integration of clinical LR and CXR DL T2D predictors.

The combined clinical LR and CXR DL T2D predictor displays the odds ratios for diabetic patients (poorly controlled and controlled T2D, n = 1554) compared to patients without diabetes (n = 8126) using a logarithmic scale and 95% CI error bars. Self-reported race is expressed relative to White, and self-reported language preference is relative to English. The logarithmic transformation of odds ratios is used to enhance visualization. Source data are provided as a Source Data file.

Fig. 3. CXR DL model scores for prospective cohort positive and negative for T2D.

Box and whisker plots show DL model scores (y-axis) for (A) no T2D versus T2D and (B) no T2D, controlled T2D, and poorly controlled T2D. Boxes represent the IQR (25–75%) with the median noted by the horizontal bar within each box, whiskers extend from the box to the minimum and maximum values within 1.5 times the IQR, with circles representing outliers in the distributions (n = 9943). The two-sided Wilcoxon rank sum test was used to assess the difference in T2D DL model score and T2D diagnosis (P < 2.2 × 10⁻¹⁶). The Kruskal–Wallis test was used to compare differences between the three groups, which were significant (P < 2.2 × 10⁻¹⁶). Source data are provided as a Source Data file.

Fig. 4. Occlusion maps identifying key features in representative CXRs with high and low diagnostic scores.

Dark green pixels highlight significant features for model prediction, primarily associated with cardiomediastinal, upper abdominal, lower neck, and supraclavicular regions. Examples of CXRs with high and low diagnostic scores are presented.

Fig. 5. Random sampling of occlusion maps from both internal and external cohorts.

A A random selection of 48 occlusion maps from the DL CXR prediction model, with green regions indicating crucial features. B A composite image of the 49 occlusion maps showcasing positive attribution to the central mediastinum, lower neck and supraclavicular fossae, upper abdomen, and ribs. C A random sampling from the external validation dataset. D An averaged map from the external data exhibiting a distribution similar to (B). Source data are provided as a Source Data file.

Fig. 6. Gifsplanation using Latent Feature Autoencoder.

The color heatmap highlights areas of change, with DL predictors progressively increasing along the horizontal axis in top and bottom rows. The change in central mediastinal adiposity is a primary driver. High predictive values (rightmost) include changes in upper abdominal fat (arrow) and supraclavicular and rib attenuation (arrowhead) which are intense upon the heatmap. The animation can be viewed as Supplementary Movie 1, which highlights the changes dynamically.

Fig. 7. Scatterplot depicting the association between actual and predicted HbA1c values using the deep learning model.

For the retrospective cohort of patients from 2010 to 2021, HbA1c values were collected within a ±30-day window of the CXRs (n = 15,945). The x-axis represents actual HbA1c values, while the y-axis shows predicted values from the deep learning model. A solid line demonstrates the linear regression fit, yielding a regression equation of y = 0.105x + 5.497, an R² of 0.15, and a P value < 0.001. A two-sided linear regression test was conducted, and the 95% confidence interval is displayed as a light gray band. Source data are provided as a Source Data file.