sebis at CRF Filling 2026: A Two-Stage Local LLM Pipeline for Medical CRF Filling

The article presents a two-stage local pipeline for filling Case Report Forms (CRFs) using the MedGemma-27B model, addressing challenges in extracting structured clinical data from unstructured electronic health records (EHRs). The architecture separates binary presence classification from value extraction, ensuring outputs are strictly based on textual evidence, which results in a macro-F1 score of 0.55, placing second in the official English test track among local, open-source submissions. This approach highlights the feasibility of privacy-preserving, on-premise LLM solutions that can rival proprietary models in clinical natural language processing.