Patient-LITE：利用大型語言模型，打造以患者為中心的通用型臨床試驗資格篩選問卷

Abstract

傳統的臨床試驗招募方式通常以站點為中心，依賴患者資格篩查。這種方式 效率較低，且往往難以招募足夠的參與者。患者一般處於被動等待試驗匹配的狀 態，這使他們無法有效比較並選擇最適合的試驗。現有的以患者為中心的系統雖 然簡化了資格標準並提供問卷，但仍需大量人工來處理標準重疊或重複的問題。

為了改善這一情況，我們推出了 Patient-LITE，一個基於大型語言模型 (LLMs)技術的系統，能夠生成對於病患而言更易懂的問卷。它能將來自多個試 驗的資格標準整合成一份簡潔的表單，達到 74-78% 的準確率。其技術創新在於混合專家方法 (mixure-of-experts)，將多個基於 GPT 的模型與優化策略相結合。 主要特徵包括:(1)獨創的「分解後組合」資料管道流程(data pipeline)，(2) 經過微調(fine-tuning) 的小型模型，用於高質量且具成本效益的問題生成，以及 (3)提示工程以協助文字擷取與邏輯推理。

Patient-LITE 是首個基於 LLM 的跨試驗問卷設計框架，並可根據不同疾病進 行靈活調整。我們提供了評估指標和微調數據集，對試驗資格的表示、分類和驗 證文獻做出了貢獻。這項研究期待為後續醫學研究者提供啟發，協助其開發技術， 促進醫學知識的普及。

The traditional site-centric approach to clinical trial recruitment, which relies on patient eligibility screening, is often inefficient and struggles to enroll sufficient participants. Patients typically wait passively for trial matches, limiting their ability to compare and select the most suitable trials. Existing patient-centric systems simplify complex criteria and provide questionnaires, but still require significant manual effort to navigate overlapping or duplicated criteria.

We introduce Patient-LITE, a tool that uses large language models (LLMs) to generate patient-friendly questionnaires consolidating eligibility criteria from multiple trials into one concise form, achieving a correctness rate of 74-78%. The technical innovation lies in a mixture-of-experts approach, combining multiple GPT-based models and optimization strategies. Key features include: (1) a break-then-assemble pipeline for cross-trial criterion extraction, (2) fine-tuned smaller models for high-quality, cost-efficient question generation, and (3) prompt engineering for context-aware logic and attribute extraction.

Patient-LITE is the first LLM-based framework for generating inter-trial questionnaires, with adaptable prompts for various diseases. We provide evaluation metrics and a fine-tuned dataset, contributing to the literature on trial eligibility representation, classification, and verification. Our work can inspire researchers in the broader medical field to adopt similar methods, enhancing patient comprehension and democratizing medical knowledge.