TAMU-NLP at ArchEHR-QA 2026: Grounded Clinical QA with Evidence Identification and Intent-Aware Answer Generation

Abstract

Electronic Health Records (EHRs) contain rich clinical information and provide an important data source for medical question answering. However, generating reliable answers grounded in patient-specific clinical evidence remains challenging. In this work, we participate in the ArchEHR-QA 2026 shared task and focus on Subtask 2 (Evidence Identification) and Subtask 3 (Answer Generation). For evidence identification, we explore both traditional learning-to-rank methods and large language models (LLMs), and propose a two-stage LLM framework that improves prediction stability through few-shot prompting and self-reflection reasoning. For answer generation, we design an intent-aware few-shot prompting framework to generate concise answers grounded in clinical evidence. Experimental results show that our approach achieves strong performance despite limited training data. On the official leaderboard, our system ranks 5th in Subtask 2 and 2nd in Subtask 3. These results demonstrate that combining evidence-driven reasoning with the generative capabilities of LLMs is an effective approach for EHR-based clinical question answering.

Resources

Details