CAPRA Framework Detects Hidden Subgroup Biases in Medical AI Without Metadata
New AI method reveals failure modes masked by strong aggregate performance when demographics are missing.
Medical imaging models often fail silently on specific subgroups (e.g., certain ethnicities or device types) when metadata like demographics or acquisition parameters are not recorded. Without this information, standard auditing tools cannot detect these hidden failure modes. In the paper "Beyond Metadata: CAPRA for Hidden Subgroup Analysis under Missing Metadata in Medical Imaging" (arXiv:2607.09102, July 2026), Yawen Li, Yan Li, Zhe Xue, Yingxia Shao, Meiyu Liang, and Guanhua Ye introduce CAPRA—a calibrated proxy-axis framework that identifies and characterizes hidden subgroups directly from images. CAPRA first predicts image-derived semantic axes (e.g., disease severity, tissue texture) using a neural network. It then calibrates the posteriors of these axes using a small, metadata-labeled subset via a patient-level cross-fitting technique. The result is an interpretable, calibrated subgroup interface that organizes images into meaningful partitions without needing subgroup labels at deployment.
Across three medical imaging domains—fundus photography, dermoscopy, and chest radiography—CAPRA revealed disparity patterns that were missed by traditional metadata-only slicing methods. For example, it detected performance drops on images with specific noise patterns or subtle pathological features that correlated with demographic groups, even when those demographics were unknown. The method remained informative under dataset shift (e.g., from different hospitals or scanners) and produced subgroup partitions that aligned more closely with explicit failure axes than baselines using image-only clustering or latent variable models. Furthermore, the same subgroup interface can be reused by downstream robust learning algorithms to improve worst-group accuracy, though gains were domain-dependent. CAPRA turns the challenge of hidden subgroup analysis under missing metadata into a practical, reusable tool for deployment-time failure analysis and fairness auditing.
- CAPRA predicts semantic axes from images alone, then calibrates using a small metadata-labeled subset via patient-level cross-fitting to create a subgroup interface.
- Tested on fundus, dermoscopy, and chest X-ray datasets, it uncovers disparity patterns missed by metadata-only slicing and remains robust under dataset shift.
- Provides an interpretable, reusable subgroup partition that can be fed into downstream robust learning methods to improve worst-group accuracy.
Why It Matters
Enables fairness auditing of medical AI in real-world settings where metadata is often missing, preventing silent failures on vulnerable subgroups.