Prediction of Type 2 Diabetes Mellitus From Chest X-Rays Using a Suite of Previously Developed Chronic Disease Deep...

Screening for type 2 diabetes (T2D) is not optimal, leading to a large number of patients being undiagnosed. Recently, deep learning (DL) applied to chest radiographs (CXRs) has shown promise for opportunistic T2D prediction. A prior study in a predominantly suburban non-Hispanic White cohort achieved an area under the curve (AUC) of 0.84 for prevalence. In this study, we evaluate the performance and generalizability of this DL model in an urban cohort with greater racial diversity, higher social deprivation, and higher T2D prevalence. We further assess whether integrating DL predictions with BMI and demographic variables improves T2D prediction beyond demographics and BMI alone.

This study aims to externally validate a previously developed DL-based CXR model for T2D prediction in a diverse urban population, to assess its performance for both prevalent and incident T2D, and to determine whether combining DL predictions with demographics and BMI improves predictive performance.

We studied adults (2010-2020) from a tertiary academic medical center in Chicago with at least one ambulatory CXR. First, we performed external validation of a previously developed DL-CXR model by applying it directly to our cohort. Second, we evaluated whether combining the DL model output with additional data, demographics, BMI, and social deprivation index improved the performance. T2D prevalence was modeled using extreme gradient boosting, while incidence was assessed with Cox proportional hazards models. Model performance was compared using AUC and concordance, and feature contributions were evaluated using feature importance and odds ratios.

Among 39,908 patients (n=21,311, 53.4% non-Hispanic Black; n=9179, 23% Latino; and n=5587, 14% non-Hispanic White), 26% (n=10,376) had T2D at their first CXR. The previously developed DL-T2D model maintained discrimination for prevalent T2D in this diverse urban cohort, with similar performance across racial groups (Latino: 0.818; non-Hispanic White: 0.819; non-Hispanic Black: 0.790), supporting generalizability. Adding DL output to demographics and BMI improved prediction compared with clinical variables alone (AUC 0.808 vs 0.766; P<.001). For a 3-year incident T2D, the full model achieved an AUC of 0.709 with concordance of 0.707; individuals in the highest risk quartile had a 7-fold higher incidence.

In a diverse urban cohort, a previously developed DL model applied to CXRs provided significant incremental value beyond demographics and BMI for T2D risk prediction. Despite substantial differences in population characteristics compared with the derivation cohort, the DL model remained effective for T2D screening. Incidence prediction was less accurate than prevalence, highlighting the need for further refinement, potentially incorporating hemoglobin A_1c when available. Although racial disparities in prevalence exist, predictive performance was comparable across groups. These findings support the generalizability of CXR-based DL for opportunistic T2D screening in diverse populations.