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

In three linesTwo-stage local pipeline using MedGemma-27B for extracting structured clinical information from unstructured EHR notes. Separates binary presence classification from value extraction, no external API calls or fine-tuning. Macro-F1 score of 0.55 on CRF 2026 test set, second place among open-source local submissions.

Summary generated by Claude — human-verified