CACRC-Pan: A Concept Annotated Case Report Corpus and Processing Pipeline for Rare Pancreatic Diseases

We introduce CACRC-Pan, a corpus of case reports for six rare pancreatic diseases (AIP1-2, CF, HCP, PS and SDS) annotated at two concept levels: token-level symptom labels (HPO-coded spans) and document-level disease labels. This dual annotation enables both standard natural language processing (NLP) tasks and studies on concept bottleneck models, linking interpretable symptom concepts to diagnostic outcomes. We also provide a baseline NLP pipeline combining PubMedBERT for symptom extraction with a lightweight voting classifier for disease inference. Results demonstrate the effectiveness of this simple, interpretable approach, and all resources are released openly to support reproducible research on rare pancreatic diseases.