关键词: 大语言模型, DeepSeek, 术前决策, 知识库, 改良心脏风险指数评分

Abstract: Objective To explore the application effectiveness of different generation modes of large language models(such as DeepSeek, ChatGPT, etc.)in the field of preoperative management and their value in assisting decision-making processes for junior physicians. Methods A total of 100 medical history records of orthopedic inpatients at Qilu Hospital of Shandong University were randomly selected from January to August 2025. Patients who were scheduled to undergo Grade I, II, III surgeries and non-joint replacement surgeries were excluded, resulting in the inclusion of total 87 patients. Guidelines related to perioperative management were retrieved from databases such as PubMed and UpToDate. After text processing and vectorization, these guidelines were used to build a perioperative management knowledge base, providing external knowledge support for subsequent model calls and question-answering tasks. The anonymized patient records were uploaded to different versions of the DeepSeek model ［DeepSeek Chat version(V3), DeepSeek Chat + knowledge base version, DeepSeek Deep Thinking version(R1), and DeepSeek R1 + knowledge base version］, and questions were posed under the identical “Instruction-Context-Input-Output(ICIO)” prompt framework. The model outputs were evaluated both objectively and subjectively. Results The DeepSeek R1 model achieved accuracy rates of 75.86% and 78.16% in the Revised Cardiac Risk Index(RCRI)scoring and risk classification tasks, respectively, significantly outperforming the Chat series models. All four model versions showed moderate accuracy in the American Society of Anesthesiologists(ASA)physical status classification and surgical feasibility judgment, with the R1 version performing slightly better. The introduction of the knowledge base slightly improved RCRI scoring accuracy only in the Chat version(+4.6%)but reduced performance in the R1 version. Subjective evaluation results indicated that junior physicians generally considered the R1 series models answers to be of greater clinical reference value, with an average score(4.19±0.72)significantly higher than that of the Chat series(Chat version: 3.06±0.06; Chat + knowledge base version: 2.97±0.03). This suggested that the R1 model has stronger practicality and acceptability in preoperative decision support(P<0.05). Conclusion The DeepSeek R1 model demonstrates good application potential in orthopedic preoperative anesthesia risk assessment and clinical decision support. However, knowledge base building and task adaptation require further optimization to enhance the models reliability and generalizability in real clinical scenarios.

Key words: Large language model, DeepSeek, Preoperative decision-making, Knowledge base, Revised cardiac risk index score