De‑identified High-Risk Pathologic Features from Unstructured Gastrointestinal Reports for PNI Extraction Benchmarking

Title:De‑identified Gastrointestinal Pathology Reports for PNI Extraction Benchmarking Description:This dataset contains 445 de‑identified gastrointestinal surgical pathology reports (colorectal 48.5%, gastric 27.8%, pancreatic 23.7%) annotated for perineural invasion (PNI) status. Expert pathologist consensus (κ = 0.85) serves as ground truth, with three classes: Present, Absent, and Indeterminate. Reports were split 70/10/20 into training, validation, and test sets (stratified by PNI status). A subset of 34 reports (7.6%) is flagged as Gold Standard Uncertainty due to ambiguous language. The data support benchmarking of large language models and traditional NLP methods for high‑risk feature extraction from unstructured pathology narratives. Usage notes: All reports are de‑identified under HIPAA Safe Harbor. Use the provided split (random seed 42) for reproducible comparisons. The “Gold Standard Uncertainty” flag allows sensitivity analyses. Format:JSON lines (.jsonl) with fields: report_id, text, pni_ground_truth, split, gold_standard_uncertainty.