<ul data-eligibleForWebStory="true"><li>Automated question answering (QA) over electronic health records (EHRs) is crucial for providing information to clinicians and patients.</li><li>Neural, a method developed for evidence-grounded clinical QA, was the runner-up in the BioNLP 2025 ArchEHR-QA shared task.</li><li>Neural's approach involves separating the task into sentence-level evidence identification and answer synthesis with explicit citations.</li><li>The method utilized DSPy's MIPROv2 optimizer to explore the prompt space and fine-tune instructions and few-shot demonstrations on the development set.</li><li>A self-consistency voting scheme was employed to enhance evidence recall without compromising precision.</li><li>On the hidden test set, Neural achieved an overall score of 51.5, ranking second while surpassing standard zero-shot and few-shot prompting by significant margins.</li><li>The results suggest that data-driven prompt optimization is a more cost-effective method than model fine-tuning for high-stakes clinical QA.</li><li>This advancement can enhance AI assistants' reliability in healthcare settings.</li></ul>

Neural at ArchEHR-QA 2025: Agentic Prompt Optimization for Evidence-Grounded Clinical Question Answering

Discover more