关键词: 大语言模型, 语义网络, LDA主题建模, 词频-逆文档频率（TF-IDF）分析, 丹红注射液

Abstract: Objective To construct a four-step extraction and analysis method for adverse drug reactions (ADR) of traditional Chinese medicine injections based on large language model (LLM)-based few-shot information extraction and multi-dimensional text mining, and to validate the feasibility of this method combination in extracting ADR signals from unstructured narratives in spontaneous reporting systems more fully and accurately. Methods A total of 4 124 spontaneous ADR reports of Danhong injection were collected that were submitted in 2023-2024. The DeepSeek-V3 model was employed with few-shot prompting to retrieve data on underlying diseases, symptoms, and outcomes, which was then concatenated into standardized sentences. Three-dimension text mining was subsequently performed: ① term frequency and TF-IDF analysis to identify high-frequency and highly discriminative symptom terms; ② semantic network analysis to construct symptom association networks based on co-occurrence matrices; and ③ LDA topic modeling to uncover latent topic structures. Results The top five high-frequency symptoms were pruritus (1 111 occurrences), rash (621), dizziness (612), chest tightness (584), and nausea (422). The semantic network identified four symptom clusters that centered on cardiovascular disease, cerebrovascular disease, non-specific systemic symptoms, and skin reactions, respectively. LDA extracted four topics in the following proportions: skin and allergic reactions (22.1%), infusion-related systemic reactions (21.1%), cardiovascular and cerebrovascular comorbidities (21.5%), and non-specific systemic and cardiovascular reactions (35.4%). The results of the two unsupervised methods—semantic network and LDA—were highly consistent and mutually validating. Conclusion The LLM-based few-shot extraction and multi-dimensional text mining method established in this study can effectively extract semantic features and latent topic structures from unstructured ADR narratives. The four-step pipeline—LLM denoising and structuring, TF-IDF for identifying high-frequency and characteristic symptoms, semantic network for revealing symptom co-occurrence associations, and LDA topic modeling for summarizing latent thematic structures—forms a hierarchical analytical framework that progresses from individual terms (“points”) through co-occurrence patterns (“lines”) and to thematic structures (“surfaces”). The feasibility of this pipeline has been validated using Danhong injection as a case study, which has established a method for post-marketing safety evaluation of traditional Chinese medicine injections that makes fuller use of free-text information.

Key words: Large Language Model, Semantic Network, LDA Topic Modeling, TF-IDF Analysis, Danhong Injection