面向医学大语言模型决策支持的模式依赖性提示组件消融：应用于耳鼻咽喉科生物制剂筛查

文献信息

中文摘要

背景： 大语言模型（LLMs）在临床决策支持方面显示出前景，但隐私限制和不可靠的自由文本决策限制了其部署。我们提出Medotor，一个本地部署的系统，用于筛查伴有鼻息肉的慢性鼻窦炎（CRSwNP）患者是否适合生物制剂治疗，并在163例回顾性病例（132例合格和31例不合格）上评估提示设计。方法：五种模型在相同患者上测试五种变体：最小提示（P0）、不含检索的完整角色/安全性/结构化九点提示（P1）、移除角色/安全性模块（P2）、移除九点模式（P4），以及在P1基础上加入一个所有患者共享的固定MemoRAG衍生知识块（M0）。每个臂均保留全部163次尝试；不可判定或格式错误的最终决策计为错误。平衡准确率是主要指标，采用患者级bootstrap区间、精确配对McNemar检验和Holm校正。结果：观察到的最佳平衡准确率为DeepSeek V4 Flash/P0的0.720、Qwen 3.5 397B/P4的0.713、GPT-OSS 120B/P4的0.704、本地Qwen 2.5/P2的0.657和本地Gemma 3n/P0的0.549。移除九点模式使Qwen 3.5相对于P1有所改善（$p_{\text{Holm}}=0.015$），而完整提示显著降低Gemma 3n性能并产生84个不可判定输出。校正后，M0未显著改善任何模型的P1。结论：因此，提示效应强烈依赖于模型，更长或检索增强的提示并非普遍更安全或更准确。由于M0既未个性化也未在患者之间更新，未运行标准RAG臂，且医师幻觉标签是定性的而非针对每个输出，该研究支持关于分类和输出可靠性的结论——而非测量的幻觉减少或MemoRAG优于标准RAG。

英文摘要

Large language models (LLMs) show promise for clinical decision support, but privacy constraints and unreliable free-text decisions limit deployment. We present Medotor, a locally deployed system for screening patients with chronic rhinosinusitis with nasal polyps (CRSwNP) for biologic therapy, and evaluate prompt design on 163 retrospective cases (132 eligible and 31 ineligible). Five models were tested on the same patients under five variants: a minimal prompt (P0), the full role/safety/structured nine-point prompt without retrieval (P1), removal of the role/safety block (P2), removal of the nine-point schema (P4), and addition to P1 of one fixed MemoRAG-derived knowledge block shared by all patients (M0). Every arm retained all 163 attempts; undecidable or malformed final decisions were errors. Balanced accuracy was the primary metric, with patient-level bootstrap intervals, exact paired McNemar tests, and Holm correction. The best observed balanced accuracies were 0.720 for DeepSeek V4 Flash/P0, 0.713 for Qwen 3.5 397B/P4, 0.704 for GPT-OSS 120B/P4, 0.657 for local Qwen 2.5/P2, and 0.549 for local Gemma 3n/P0. Removing the nine-point schema improved Qwen 3.5 relative to P1 ($p_{\text{Holm}}=0.015$), whereas the full prompt sharply degraded Gemma 3n and produced 84 undecidable outputs. M0 did not significantly improve P1 for any model after correction. Thus, prompt effects were strongly model-dependent, and a longer or retrieval-augmented prompt was not universally safer or more accurate. Because M0 was neither personalized nor updated between patients, no standard-RAG arm was run, and physician hallucination labels were qualitative rather than per output, the study supports conclusions about classification and output reliability-not measured hallucination reduction or superiority of MemoRAG over standard RAG.