結合知識檢索與生成式 AI 之健檢報告生成

Abstract

本研究旨在建立資料檢索導向的健康檢查報告生成系統。系統整合 ChatGPT 語言模型生成報告內容，同時以標準化健康檢查數據及結構化問卷調查作為檢索資料庫。透過多階段檢索架構，首先利用BGE-M3 embedding模型進行初步語義匹配，再透過 BGE-reranker 對查核項目（健檢建議）與檢索結果（檢查數據和問卷）進行精確相關性排序，篩選出高度相關的結果作為DeepSeek 語言模型查核的強化參考依據。此機制能有效識別健檢建議中的幻覺現象與邏輯謬誤，顯著提升報告整體可靠性與臨床參考價值。本研究主要貢獻為：（一）減少LLM在處理複雜健檢資料時的幻覺與誤判：運用 BGE-M3 進行多維度密集向量檢索獲取相關參考資料，再以 BGE-reranker 應用注意力機制重排檢索結果，精確篩選高相關資訊作為 LLM 查核依據，有效檢測幻覺的產生，未來可根據不同檢查項目特性與需求動態調整檢索參數與資料集；（二）拓展智慧醫學領域應用：實現高度客製化的自動健檢報告生成與幻覺查核，提高報告準確性、專業一致性與患者理解度，有效減輕醫師工作負荷，同時保持專業判讀品質，為精準醫療與人工智能醫療整合提供可擴展實證架構，推動醫療AI從輔助決策工具向智能協作夥伴的轉型。

This study proposes a retrieval-augmented system for generating personalized health checkup reports. By integrating the ChatGPT language model with standardized health examination data and structured questionnaires as the retrieval corpus, the system employs a multi-stage retrieval framework: BGE-M3 is used for initial semantic matching, followed by BGE-reranker to rank the relevance between checkup suggestions and retrieved content. The top-ranked results are then provided as contextual grounding for the DeepSeek language model to detect hallucinations and logical inconsistencies in the generated recommendations. Experimental results demonstrate that this approach effectively reduces hallucinations, enhances report reliability, and improves clinical value. The contributions of this study lie in reducing LLM-induced errors when processing complex medical data through dense retrieval and re-ranking, and in expanding the application of AI in precision healthcare by enabling automated, accurate, and explainable report generation.