文献速递乳腺癌支持与姑息治疗

DeepSeek人工智能模型在乳腺癌治疗决策中的应用分析

[Application analysis of the artificial intelligence model DeepSeek in breast cancer treatment decision-making].
PubMed 文献 · Zhonghua Zhong Liu Za Zhi · 2026-01-01

本研究纳入163例经MDT制定方案的乳腺癌患者,将匿名病历提交DeepSeek「深度思考R1」模式获取治疗方案,由2位专家与MDT方案对比评分。DeepSeek方案与MDT完全一致者占25.2%(41/163),在96例以辅助治疗决策为目的的患者中总体错误率为46.8%(45/96),提示其意见与MDT存在差异,需在专业人员指导下使用。

要点

  • DeepSeek与MDT方案完全一致率仅25.2%(41/163)。
  • 辅助治疗决策场景下DeepSeek总体错误率46.8%(45/96)。
  • 每例患者DeepSeek给出方案时间<20秒。

核心结论

  • DeepSeek对乳腺癌整体治疗决策的意见与MDT方案存在一定差异,需在专业人员鉴别指导下用于临床实践。

研究设计与样本

  • 单中心研究,纳入2025年2月11—24日山东省肿瘤医院经MDT制定方案并拟行抗肿瘤治疗的乳腺癌患者163例。
  • 病历资料匿名化后提交DeepSeek并选择「深度思考R1」模式获取治疗方案。
  • 由2位高级职称乳腺癌临床专家独立将DeepSeek方案与MDT方案比较一致性并评分。

主要发现

  • DeepSeek为每例患者给出治疗方案的时间<20秒。
  • 2位专家对DeepSeek与MDT方案一致性的最终评分为0分15例、1分28例、2分79例、3分41例。
  • DeepSeek方案与MDT方案完全一致的患者占25.2%(41/163)。
  • 96例以辅助治疗决策为目的的患者中,DeepSeek建议的总体错误率为46.8%(45/96)。

临床怎么用

  • DeepSeek可用于快速生成乳腺癌治疗参考意见,但不能替代MDT决策。
  • 临床采用DeepSeek建议时需由专业人员鉴别,尤其辅助治疗决策场景错误率较高。

局限性

  • 单中心研究,样本来自特定时间段参加MDT的患者。
  • 一致性评分由2位专家主观判断,缺乏客观金标准。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

原文

Objective: To evaluate the accuracy and feasibility of applying the DeepSeek artificial intelligence model in clinical decision-making for breast cancer by comparing its recommendations with the diagnosis and treatment plans formulated by multi-disciplinary team (MDT). Methods: A total of 163 breast cancer patients, who were given detailed treatment plans via MDT at Shandong Cancer Hospital and Institute between February 11, 2025 and February 24, 2025 and were scheduled to undergo anti-tumor treatment, were included in the study. The patients' medical records were meticulously organized to include details such as gender, age, menopausal status, imaging examination results, pathological diagnosis, molecular subtyping, previous treatment regimens and treatment course, recurrence, and metastasis. Personal identifiers, including patient names and medical record numbers, were anonymized, and the data were submitted to DeepSeek. The 'Deep Thinking R1' mode was selected to derive the treatment plans. Two senior breast cancer clinical experts independently compared the treatment plans provided by DeepSeek with the MDT treatment plans, assessed their consistency, and assigned scores accordingly. Results: DeepSeek provided treatment plans for each patient in under 20 seconds. Two experts assessed the final consistency between DeepSeek's treatment plans and those of the MDT, rating them as follows: 0 point for 15 cases, 1 point for 28 cases, 2 points for 79 cases, and 3 points for 41 cases. The percentage of patients for whom DeepSeek's treatment plans were entirely consistent with the MDT treatment plans was 25.2% (41/163). Among 96 patients for whom the purpose was to assist in treatment decision-making, the overall error rate of DeepSeek's recommendations was 46.8% (45/96). Conclusion: Compared with the MDT diagnosis and treatment plans, DeepSeek's opinions on the overall treatment decision of breast cancer have certain differences, and they need to be used in clinical practice under the guidance of professional identification. 目的: 通过与乳腺癌多学科诊疗(MDT)方案对比,探讨DeepSeek人工智能大模型在乳腺癌临床治疗决策方面应用的准确性和可行性。 方法: 收集2025年2月11—24日在山东省肿瘤医院参加MDT获得详细治疗方案并拟行抗肿瘤治疗的乳腺癌患者163例,将患者病历资料进行详细整理,包括患者性别、年龄、绝经情况、影像学检查结果、病理诊断、分子分型、既往治疗方案及治疗经过、复发转移等信息,隐去患者姓名、病案号等个人信息,提交给DeepSeek,选择“深度思考R1”模式,获取治疗方案。由2位具有高级职称的乳腺癌临床专家,就DeepSeek给出的治疗方案与MDT治疗方案进行比较,评估其一致性,分别独立给出评分。 结果: 每例患者DeepSeek给出治疗方案的时间<20 s。2位专家对DeepSeek治疗方案与MDT治疗方案一致性的最终评分为0分15例,1分28例,2分79例,3分41例,DeepSeek治疗方案与MDT治疗方案完全一致的患者占25.2%(41/163)。96例以辅助治疗决策为目的的患者中,DeepSeek建议的总体错误率为46.8%(45/96)。 结论: 针对乳腺癌的整体治疗决策,DeepSeek的意见与MDT诊疗方案相比存在一定的差异性,需要在专业人员的鉴别指导下才能用于临床实践。.

来源:Zhonghua Zhong Liu Za Zhi,查看原文。版权归原作者所有。

引用

  1. Shao H, Liu YB, Zhou ZZ, et al. [Application analysis of the artificial intelligence model DeepSeek in breast cancer treatment decision-making]. Zhonghua Zhong Liu Za Zhi. 2026. PMID: 42151091. doi:10.3760/cma.j.cn112152-20250319-00109

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。