Encoding Versus Linear Use of Patient Characteristics in Chest X-Ray Foundation Models on MIMIC-CXR
Open Access
- 29 June 2026
- journal article
- research article
- Published by MDPI AG in Diagnostics
- Vol. 16 (13) , 2030
Abstract
Background: Chest X-ray (CXR) foundation models can predict patient demographic categories (sex, age, race) from images alone by linear probing, but whether encoded attributes drive finding prediction has not been tested at scale. Methods: On MIMIC-CXR (230,697 images, 60,518 patients), we measured attribute dependence (AUROC drop after residualizing an attribute from a frozen embedding) across 24 patient attributes (four demographics and 20 ICD-coded comorbidities), 10 thoracic findings, and 6 overlap-free foundation models ( triplets), with 3 additional CXR-pretrained models (RAD-DINO, CheXzero, CheSS) for encoding and fairness analyses. Dependence was regressed on attribute-finding odds ratios (ORs), encoding strength, and model-level factors. Results: Encoding and dependence dissociated. Sex (AUROC 0.942) contributed ; race (0.83) contributed (rank 14/24); heart failure (0.774) showed the largest dependence (). explained 50.6% of dependence variance (, ); model factors added no detectable contribution (, ). Residualizing the top three high-OR attributes reduced AUROC by 0.026 without narrowing sex or age subgroup gaps (minimum detectable effect size (MDES) = 0.0019). Across 9 models, four-category race subgroup gaps (mean 0.069) were 30–75× larger than race residualization drops (mean 0.0015); CheXzero showed the same decoupling. Conclusions: Encoding, residualization-sensitive dependence, and subgroup bias are three separable quantities on the same model. Pre-deployment audits on inpatient-skewed cohorts can prioritize attributes by local OR; jointly residualizing race and its cardiac correlates does not narrow the race subgroup gap, which instead tracks group-wise finding base rates. Cross-institutional transfer remains open: no public CXR cohort currently links comorbidity electronic health records for external validation of the OR-dependence relationship.
Keywords
Funding Information
- Korea Health Industry Development Institute
This publication has 25 references indexed in Scilit:
- Domain Shift Analysis in Chest Radiographs Classification in a Veterans Healthcare Administration PopulationJournal of Imaging Informatics in Medicine, 2025
- The limits of fair medical imaging AI in real-world generalizationNature Medicine, 2024
- Detecting shortcut learning for fair medical AI using shortcut testingNature Communications, 2023
- Algorithmic encoding of protected characteristics in chest X-ray disease detection modelsEBioMedicine, 2023
- CheSS: Chest X-Ray Pre-trained Model via Self-supervised Contrastive LearningJournal of Digital Imaging, 2023
- MIMIC-IV, a freely accessible electronic health record datasetScientific Data, 2023
- Expert-level detection of pathologies from unannotated chest X-ray images via self-supervised learningNature Biomedical Engineering, 2022
- AI recognition of patient race in medical imaging: a modelling studyThe Lancet Digital Health, 2022
- AI for radiographic COVID-19 detection selects shortcuts over signalNature Machine Intelligence, 2021
- MIMIC-CXR, a de-identified publicly available database of chest radiographs with free-text reportsScientific Data, 2019