Not Just Answers, but the Basis for Judgment: CLEAR Enables Concept Auditing of Chest X-Ray AI

Even when chest X-ray artificial intelligence gives the correct answer, physicians may still not know whether it actually detected the lesion or picked up cues that happened to coexist in the data. A research team from the University of Pennsylvania School of Medicine and other institutions developed the CLEAR model in an effort to open this black box: in addition to outputting disease probabilities, it shows which radiological observations drove the judgment, allowing errors to be investigated rather than leaving users only able to accept or reject a score.

The research team trained CLEAR using MIMIC-CXR, CheXpert-Plus, and ReXGradient data, comprising a total of 873,342 chest X-ray and report pairs from 239,391 patients. The system extracted 368,294 observations from the reports, including descriptions such as cardiac enlargement, aortic calcification, or opacities in specific locations, and then used image and text encoders to map the features shown on X-rays to this concept space.

This design enables CLEAR to identify pneumonia, pulmonary nodules, pleural effusion, and skeletal or vascular abnormalities in a “zero-shot” setting without retraining for individual diseases, while also listing each concept’s contribution to the result. The study tested the model on external, physician-annotated data from the United States, Spain, and Vietnam. On VinDr-CXR, CLEAR achieved a mean AUROC of 0.782 across multiple findings, higher than the next-best comparison model’s 0.750; on PadChest, it achieved 0.700, compared with 0.668 for the comparison model. Performance still varied across datasets and lesions, so the averages cannot be regarded as accuracy across all clinical settings.

The most concrete value of auditability emerged when the model made mistakes. The team found that when a supervised classifier was used to identify an “enlarged cardiomediastinum,” the model relied heavily on concepts related to atelectasis, reflecting how frequently the two labels appeared together in the training data. By retaining only concepts genuinely related to the mediastinum and refitting a lightweight classifier, the researchers increased the AUROC from 0.727 to 0.784 without retraining the image encoder. This experiment shows that concept-based explanations are not merely post hoc displays, but may also be used to locate confounding factors and intervene to correct them.

However, being able to “state reasons” does not mean the model has achieved clinical safety. The concept library was derived from existing reports and may still inherit reporting practices, annotation biases, and population differences; the study also did not exhaustively cover all imaging protocols, patient populations, or real-world workflows. The statistical comparisons in the paper were not adjusted for multiple comparisons, and external benchmark testing cannot replace prospective clinical trials, particularly because it cannot by itself demonstrate that the model improves patient outcomes.

The team has released the model weights, code, and concept vocabulary, allowing other researchers to reproduce the concept scoring and auditing workflow. However, the public documentation explicitly states that CLEAR is currently for research use only, is not a medical device, and has not been validated for diagnosis, treatment, or triage. Before it can enter hospitals, regulators and purchasers must assess not only overall accuracy, but also whether concept attribution is stable, whether it can be reproduced across institutions, and how humans should handle situations in which the model’s stated reasons conflict with clinical judgment.