EClinicalMedicine · Published 2026-08-01 · DOI 10.1016/j.eclinm.2026.104078
Theodorus Dapamede, Kéana Aitcheson, Pola Lydia Lagari, Awais Farooq, Peter Louis, Chad Robichaux, Frank Li, Bardia Khosravi, Mohammadreza Chavoshi, Aawez Mansuri, Rohan Satya Isaac, Beatrice Brown-Mulry, Hari Trivedi, William Galanter, Ayis Pyrros, Judy Gichoya
Summary: Background: Chronic kidney disease (CKD) affects 843 million people worldwide yet remains underdiagnosed. Current risk stratification, including the CKD Prognosis Consortium (CKD-PC) score, relies on laboratory data rarely evaluated outside nephrology settings. Chest radiographs (CXR) are among the most frequently performed imaging procedures globally, offering an opportunistic screening pathway. CXR-CKD5, a CXR-derived AI score for predicting 5-year incident CKD without laboratory data, was developed and externally validated. Methods: In this retrospective two-centre study conducted at two sites in the USA, 97,553 adults (≥18 years) without baseline CKD who underwent routine CXR were included: Emory University (development cohort, 2008–2021, n = 75,683) and University of Illinois Chicago (external validation cohort, 2010–2020, n = 21,870). Patients with a CKD diagnosis within 90 days of the index CXR were excluded. A convolutional neural network extracted 10 cardiopulmonary and metabolic risk features from routine CXRs, which were used to train an XGBoost accelerated failure time model. The primary outcome was 5-year incident CKD. Discrimination (C-index) and calibration were evaluated against a clinical base model and the CKD-PC score. Findings: CXR-CKD5 achieved apparent C-indices of 0.774 (95% CI: 0.768–0.780) in development (optimism-corrected: 0.737) and 0.713 (95% CI: 0.697–0.729) at external validation. A combined imaging-clinical model achieved the highest external validation discrimination (C-index 0.734, 95% CI: 0.719–0.751). Performance was strongest in non-diabetic patients (development C-index 0.783, 95% CI: 0.775–0.790; external validation 0.714, 95% CI: 0.689–0.739) and attenuated in diabetic patients across both cohorts (development C-index 0.658, 95% CI: 0.645–0.671; external validation 0.607, 95% CI: 0.580–0.634). Absolute risk was overestimated at external validation, indicating that calibration would be required before deployment at new institutions. A threshold of ≥15% predicted risk identified the top 20% at risk. Interpretation: CXR-CKD5 demonstrates that routine chest radiographs can identify patients at elevated CKD risk without laboratory data, offering a scalable opportunistic screening pathway. Recalibration and prospective evaluation assessing clinical outcomes and workflow integration are needed before deployment. Funding: NHLBI and the University of Illinois Chicago AI.Health4All Initiative.
Abstract from DOAJ. Public domain (CC0 1.0).
Read the article at the publisher →
Dapamede, T., Aitcheson, K., Lagari, P., et al. (2026). A deep learning model for opportunistic screening of chronic kidney disease using chest radiographs: a multicentre validation study in the USA. EClinicalMedicine. https://doi.org/10.1016/j.eclinm.2026.104078