Conformal uncertainty aware prototype discovery for auditing deep chest radiograph clustering

Abstract Deep clustering can support exploratory organization of large chest-radiograph repositories, but deterministic assignments can conceal model ambiguity and apparently coherent clusters may reflect acquisition or optimization effects rather than clinically validated phenotypes. We present CUPID-Med, an uncertainty-aware prototype clustering framework combining distributional latent learning, learnable prototypes, anti-collapse regularization, and patient-disjoint split-conformal cluster sets. The evaluated implementation operates on frozen 1,024-dimensional ImageNet-pretrained DenseNet121 embeddings and uses a leakage-controlled 60/10/15/15 Train/Validation/Calibration/Test protocol on 10,000 CheXpert radiographs from 10,000 distinct patients. Model selection is restricted to Validation, and the frozen benchmark evaluates ten matched seeds in a common train-only PCA-64 geometry. On Test, CUPID-Med achieved shared-space Silhouette $$0.0403\pm 0.0041$$ , compared with 0.0440 for Spectral-kNN and 0.0429 for k -means, while yielding higher mean Silhouette than the evaluated deep representation-learning baselines. None of the 63 CUPID-Med-versus-baseline matched-seed comparisons remained significant after Holm correction, so performance differences are interpreted descriptively. At nominal 90% confidence, the primary stochastic APS procedure achieved $$0.9753\pm 0.0070$$ empirical coverage with mean set size 5.61; assignment entropy aligned with cross-seed instability ( $$\rho =0.618$$ , bootstrap 95% CI 0.581–0.652), and frozen weak-view perturbations changed hard assignments in 3.36% of Test cases on average. External evaluation on 112,120 NIH ChestXray14 images retained conservative model-relative coverage under measurable distribution shift. Across ten seeds, the metadata audit identified substantial association between cluster membership and the recorded acquisition-projection field (Cramér’s $$V=0.614\pm 0.039$$ , range 0.515–0.641). The uniformly small internal Silhouette values indicate weak absolute separation for all evaluated methods. CUPID-Med should therefore be interpreted as an exploratory framework for auditing cluster structure and assignment ambiguity, not as a diagnostic or clinically validated phenotype-discovery system.