An AI model can be right for the wrong reason. CLEAR makes the reason visible.
Today, our work introducing CLEAR (Concept-Level Embeddings for Auditable Radiology) was published open access in Nature Biomedical Engineering.
Most medical-imaging foundation models make predictions using hidden features that are difficult for clinicians and researchers to inspect. CLEAR instead represents a chest X-ray through 368,294 radiological observations, allowing each prediction to be decomposed into clinically meaningful concepts.
We trained CLEAR on more than 0.87 million image–report pairs from 239,391 patients and externally evaluated it across physician-annotated datasets from the United States, Europe, and Asia. The framework supports auditable zero-shot prediction, systematic identification of confounding associations, and concept bottleneck models.
One result illustrates why this matters. For enlarged cardiomediastinum, CLEAR exposed an inappropriate reliance on atelectasis-related concepts. Retaining clinically appropriate mediastinal concepts improved AUROC from 0.727 to 0.784 while the image encoder remained frozen and only the lightweight classifier was refitted.
This work moves medical AI from systems that only produce predictions toward systems whose reasoning can be inspected, challenged, and corrected.
Read and use CLEAR
- Open-access paper in Nature Biomedical Engineering
- Code on GitHub
- Model and concept resources on Hugging Face
- DOI: 10.1038/s41551-026-01741-4
The study was co-led by Tianyu Han and Riga Wu with collaborators across the University of Pennsylvania, RWTH Aachen University, Technical University of Munich, Technical University Dresden, Heidelberg University Hospital, and partner institutions.
Featured image: Figure 1 from Han et al., Nature Biomedical Engineering (2026), licensed under CC BY 4.0.