Gladiator at MEDIQA-SYNUR 2026: Contextual Clinical Extraction: Integrating Foundation Models with Domain-Specific Validation Rules
Year: 2026
Type: conference-paper
Abstract: We present a hybrid system for schema-guided clinical information extraction from synthetic nursing dictations, developed for the MEDIQA-SYNUR shared task at ClinicalNLP 2026.Our approach combines a large language model (Claude Opus 4) with comprehensive prompt engineering, multi-layered post-processing pipelines, and rule-based extraction using over 400 domain-specific regex patterns.The system extracts structured observations across 193 clinical concepts spanning vital signs, neurological assessments, respiratory status, and functional evaluations.On the development set, our system achieves 74.27% precision, 84.32% recall, and 78.98% F1 score.Detailed error analysis reveals that secondary diagnosis over-extraction (32 false positives) and cognitive status classification (12 value errors) represent primary challenges.We provide comprehensive methodology documentation to enable reproducibility, including prompt templates, validation rules, and regex patterns.We discuss limitations including cost-performance tradeoffs of proprietary LLMs and deployment considerations for healthcare settings.
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: hybrid
OpenAlex ID: W7166537735