Risk of Bias in Chest Radiography Deep Learning Foundation Models

A chest-radiography foundation-model audit connects feature inspection with subgroup disease-detection evaluation. Authors report sex- and race-related performance disparities in the studied model.

L2Research workflow Source Peer-reviewed source Outcome Author-reported

Evidence & verification

Source level

Peer-reviewed source
Grounded in a peer-reviewed publication. This describes the source; it does not mean the reported results have been independently reproduced.

Outcome verification

Author-reported
Outcomes are reported by the team that built the system. The entry is source-checked against the team's own publications; not independently reproduced.

Sources checked

2026-11-01 (UTC)

AI source review: paper-excerpt. Institutional-hosted version-of-record layout, 10-page PDF; Glocker, Jones, Roschewitz and Winzeck; Radiology: Artificial Intelligence 2023 5(6):e230060; DOI10.1148/ryai.230060 visible; received Feb28/revised Aug7/accepted Aug24 2023. No alternate version inferred. Source class does not establish correctness or research benefit.

Source-checked does not mean independently reproduced. See the Method page for how these labels are assigned.

Sources (1)

Research applicability

Supports research design and analysis of bias in foundation-model features and downstream predictions. The study does not establish clinical safety or universal model behaviour.

Getting started

Read the primary methods and adapt only after checking source limitations; this is editorial guidance, not an executed recipe.

Maintenance

Only the cited snapshot was reviewed; present installation, maintained service availability and latest-version parity are unverified.

License

Check original licences, data rights and provider terms; source access is not reuse permission. No installation or rights audit performed.

Workflow & human oversight

Scoped L2: Retains prior L2, now supported by institutional PDF: researcher-selected image splits/models → feature extraction → dimension reduction/statistical inspection → downstream disease prediction and subgroup evaluation. Human team determines analysis and interprets results; no research-agent planning. Human control and gaps: Study design, model and statistical choices made by authors.

Workflow level is not an evidence score; higher is not better evidence.

Approach

Researchers define image splits and models, obtain model-generated features, inspect feature distributions, and compare downstream disease-detection performance across controlled subgroups.

Outcomes (as reported by the originating team)

Author-reported findings or documented capabilities only; no independent reproduction or expert validation was performed.

Limitations

Selected PDF excerpts reviewed, not a complete figure audit. Outcomes are author-reported and were not reproduced; current model/API availability and clinical safety are unverified.

Reproducibility

Source-text review with frozen provenance. No candidate software, models, data uploads or research outcomes were executed or reproduced.

Adoption requirements

Verify research scope, data rights, tool requirements and outputs before use; these editorial suggestions were not executed.

Comments

No comments yet.

Add a comment

Comments are held for moderation before appearing. Critical or corrective comments are welcome — especially evidence that an entry overstates what has been verified.