Cancer Staging and Biomarker Extraction

The proposed framework aims to extract structured staging and biomarker variables while avoiding hallucination risk and extra post-processing from generative extraction. Unstructured pathology reports make large-scale registry curation difficult and create dependence on manual abstraction. Generative LLM extraction is presented as problematic because autoregressive decoding can produce invalid values, inconsistent formatting, stochastic outputs, and hallucinated categories. Breast cancer staging and biomarker profiling rely on pathology-report information such as TNM stage, histologic grade, ER, PR, and HER2 status.