SemAnTICA Lab at MediQA-SYNUR 2026: Route, Extract and Verify – An LLM-gated Ensemble for Parsing Nurse Dictations

Abstract

We describe the Semantic Analysis of Text to Inform Clinical Action (SemAnTICA) Lab’s system for the MediQA-SYNUR 2026 shared task on extracting structured clinical observations from nurse dictation transcripts. The task requires mapping observations from disfluent conversational text to a large, fixed ontology and producing strictly normalized outputs, where small amounts of concept over-selection severely degrade micro-F1 score. Our approach evolved from a full-schema in-context baseline to a pipeline that explicitly separates concept selection from value extraction. We first preprocess transcripts, then generate transcript-specific concept candidates using hybrid sparse–dense retrieval. The candidates are then pruned with an evidence-based filter. For extraction, we adopt a system-level mixture-of-experts design with an online LLM router that selects a subset of domain-specialized experts per transcript. Each expert operates over a constrained schema partition to reduce spurious predictions. We enhance robustness with agreement-gated ensembling and targeted adjudication for ambiguous cases. Finally, we intersect complementary high-recall and high-precision runs to produce the best submission. Our system ranked first on the official test leaderboard with F1 = 0.814, P = 0.826, R = 0.801.

Resources

Details