MIDAS_SYNUR at MEDIQA-SYNUR 2026: A Prompting Study for Clinical Observation Extraction from Nurse Dictation Transcriptions

Abstract

This paper describes MIDAS_SYNUR, a system developed for the MEDIQA-SYNUR task at ClinicalNLP 2026 on observation extraction from nurse dictations. The primary system adopts a single-prompt, field-rich few-shot strategy using GPT-5.2, jointly generating all schema fields in one structured output. Few-shot demonstrations are curated and grouped by value type, with five examples per type, promoting consistency across heterogeneous value distributions while leveraging global context to resolve cross-field dependencies. To analyze design trade-offs, this holistic strategy is compared against a field-wise decomposed prompting baseline, where each schema field is extracted independently using explicit positive and NULL demonstrations to improve absence detection and reduce cross-field interference. Zero-shot variants of both approaches are also evaluated to isolate the contribution of in-context examples. The results highlight inference-time prompting as a simple, reproducible, and competitive baseline for large-scale clinical observation extraction from conversational nurse dictations. Keywords: Prompt Engineering, Few-Shot Prompting, In-Context Learning, Structured Output

Resources

Details