應用多代理檢索增強生成自動擷取急性淋巴性白血病亞型診斷特徵

Abstract

急性淋巴性白血病（Acute Lymphoblastic Leukemia, ALL）具有高度分子異質性，不同亞型的預後與治療策略存在根本性差異——精準的亞型診斷直接決定是否需要加入靶向藥物，誤判可能使患者錯失關鍵治療或承受不必要的毒性。然而，ALL 的診斷特徵廣泛分散於大規模生醫文獻中，且歷史命名異質性嚴重阻礙自動化知識整合。現有的基因資訊擷取技術仍高度仰賴傳統文獻檢索與人工整理，缺乏具擴展性與自動語意理解能力的工具；而直接依賴大型語言模型（LLM）參數化記憶的生成方式，則在臨床場景中存在不可接受的幻覺風險。 針對上述挑戰，本研究設計一套多代理檢索增強生成（Multi-Agent Retrieval-Augmented Generation, Multi-Agent RAG）系統，以「文本可溯源擷取」作為核心設計約束，要求每項擷取結果均須追溯至原始文獻的明確文字證據。系統整合倒數排名融合混合檢索與交叉編碼器重排序，實現雙階段精煉檢索；多代理人協作框架包含文件評判器、生成代理、幻覺偵測器與深度推理代理，並搭配依亞型診斷邏輯差異設計的動態提示詞路由。 實驗結果顯示，Agentic RAG 策略的整體命中率由無檢索增強基準線的 17.84% 大幅提升至 95.69%。相較於前沿商用大型語言模型 Claude Opus 4.8 與 GPT-5.5 之純參數知識（No-RAG）——兩者雖具高召回率（78.8%、73.6%），卻因過度生成而精確率偏低（29.9%、27.4%）——本系統在文本可溯源擷取的約束下兼具高召回與高精確（F1 91.4%），顯示其效能源自可溯源設計而非過度生成；切塊消融實驗揭示切塊粒度對擷取效能呈倒 U 型影響，此模式可能同時反映證據粒度、檢索排序與上下文利用率的綜合效應。 本研究的主要貢獻在於提出一套面向 ALL 多亞型診斷特徵擷取、以文本可溯源性為核心約束的 Multi-Agent RAG 框架，為生醫知識的自動化擷取提供一套可溯源且可透過語料庫更新持續擴充的系統性方法。

Acute Lymphoblastic Leukemia (ALL) exhibits high molecular heterogeneity, with fundamentally different prognoses and treatment strategies across subtypes — accurate subtype diagnosis directly determines whether targeted therapies are indicated, and misclassification may deprive patients of critical treatment or expose them to unnecessary toxicity. Yet diagnostic features are dispersed across large-scale biomedical literature and further complicated by substantial historical naming heterogeneity. Existing geneinformation extraction techniques remain heavily reliant on traditional literature search and manual curation, lacking scalable tools with automatic semantic understanding; meanwhile, generation that relies solely on the parametric memory of large language models (LLMs) poses unacceptable hallucination risks in clinical contexts. To address these challenges, this study presents a Multi-Agent Retrieval-Augmented Generation (Multi-Agent RAG) system constrained by evidence-grounded extraction as a core design principle, requiring every extracted result to be traceable to explicit textual evidence in source documents. The system integrates Reciprocal Rank Fusion (RRF) hybrid retrieval with cross-encoder reranking for two-stage refinement, and employs a multi-agent framework comprising a Document Grader, Generator, Hallucination Detector, and Deep Reasoner, complemented by dynamic prompt routing tailored to differences in diagnostic logic across subtype groups. Experimental results show that the Agentic RAG strategy improves overall hit rate from a no-RAG LLM baseline of 17.84% to 95.69%. Compared with the parametric knowledge (No-RAG) of frontier commercial LLMs Claude Opus 4.8 and GPT-5.5 — which attain high recall (78.8% and 73.6%) but low precision (29.9% and 27.4%) due to over-generation — the proposed system achieves both high recall and high precision (F1 = 91.4%) under its evidence-grounded design, indicating that its performance stems from traceable retrieval rather than overgeneration; chunking ablation experiments reveal an inverted-U pattern in extraction performance across chunk sizes, likely reflecting joint effects of evidence granularity, retrieval ranking, and context utilization. The key contribution of this study is a Multi-Agent RAG framework for extracting diagnostic features across multiple ALL subtypes under an evidence-grounded constraint. This framework provides a traceable and systematically extensible approach to automated biomedical knowledge extraction through continued corpus updates.