@inproceedings{krishnasamy-wihardjo-2026-gigitai-archehr,
    title = "{G}igit{AI} at {A}rch{EHR}-{QA} 2026: Prompting Strategies and Constitutional {AI} for Clinical Question Answering",
    author = "Krishnasamy, Saran  and
      Wihardjo, Inez",
    editor = "Gupta, Deepak  and
      Thompson, Paul  and
      Ananiadou, Sophia  and
      Demner-Fushman, Dina",
    booktitle = "Proceedings of the Third Workshop on Patient-Oriented Language Processing ({CL}4{H}ealth) @ {LREC} 2026",
    month = may,
    year = "2026",
    address = "Palma, Mallorca (Spain)",
    publisher = "ELRA Language Resources Association (ELRA)",
    url = "https://preview.aclanthology.org/test-year-match/2026.cl4health-1.53/",
    doi = "10.63317/57f49ajj4svw",
    pages = "565--577",
    abstract = "Answering patient questions from electronic health records requires identifying relevant evidence in lengthy clinical notes and generating faithful, patient-friendly answers. We present a systematic study of LLM prompting strategies for both tasks, evaluating 21 evidence identification methods and 13 answer generation methods across 7 language models. For evidence identification, we find that LLM prompting outperforms traditional retrieval (BM25, SBERT, BioLinkBERT) by 19 F1 points, and that prompt framing alone controls precision{--}recall trade-offs: inclusive framing achieves 90{\%} recall on dev while balanced framing reaches 67{\%} precision. For answer generation, we introduce a Constitutional AI pipeline that critiques and revises answers against five clinical faithfulness principles, improving BLEU and ROUGE over the constrained baseline. Our analysis reveals that chain-of-thought effectiveness is strongly model-dependent, and that simple well-designed prompts outperform complex multi-step pipelines. We evaluate our approaches on the ArchEHR-QA 2026 shared task at CL4Health, achieving 58.0 F1 for evidence identification and 31.8 overall for answer generation."
}