Gemini与DeepSeek大型语言模型在伴胆脂瘤的慢性中耳炎乳突切除术手术决策支持中的性能比较

文献信息

中文摘要

目的： 比较两种最先进的大型语言模型（LLM）Gemini Pro v3.1与DeepSeek v3.2在使用专家验证的真实患者数据为乳突切除术提供手术决策支持和术后管理方面的临床准确性。
方法： 一项回顾性分析纳入121例因伴胆脂瘤的慢性中耳炎而实施的乳突切除术病例。向两个模型呈现临床病例情景，包括耳镜检查、纯音测听和影像学发现。由四名耳鼻咽喉科医师开发的100分评分系统在两个领域评估AI生成的回答：手术规划和术后管理。配对样本t检验比较模型性能，Mann-Whitney U检验和Kruskal-Wallis检验评估临床变量的影响。
结果： Gemini与DeepSeek在手术规划方面未发现显著差异（76.07% vs. 73.97%，p = 0.258）。然而，Gemini在术后管理方面显著优于DeepSeek（67.85% vs. 62.11%，p = 0.001）。两个模型在手术规划方面的得分均显著高于术后护理。DeepSeek在合并症患者（p = 0.004）和术后并发症患者（p < 0.001）中表现出显著更高的手术准确性。手术入路显著影响DeepSeek的性能（p = 0.001），显微病例的结果优于内镜病例。
结论： 两个LLM在手术规划方面表现良好，但在术后管理方面效果较差。Gemini在术后建议方面表现出更优性能，而DeepSeek受益于更清晰的疾病严重程度指标。这些发现支持LLM作为耳科手术决策辅助工具的潜力，同时强调需要专家临床监督。

英文摘要

OBJECTIVE: To compare the clinical accuracy of two state-of-the-art Large Language Models (LLMs), Gemini Pro v3.1 and DeepSeek v3.2, in providing surgical decision support and postoperative management for mastoidectomy using expert-validated real patient data.
METHODS: A retrospective analysis included 121 mastoidectomy cases performed for chronic otitis media with cholesteatoma. Clinical case scenarios, including otoscopy, pure-tone audiometry, and radiological findings, were presented to both models. AI-generated responses were evaluated using a 100-point scoring system developed by four otolaryngologists in two domains: surgical planning and postoperative management. Paired-sample t-tests compared model performance, while Mann-Whitney U and Kruskal-Wallis tests assessed the effects of clinical variables.
RESULTS: No significant difference was found between Gemini and DeepSeek in surgical planning (76.07% vs. 73.97%, p = 0.258). However, Gemini significantly outperformed DeepSeek in postoperative management (67.85% vs. 62.11%, p = 0.001). Both models achieved significantly higher scores in surgical planning than postoperative care. DeepSeek demonstrated significantly higher surgical accuracy in patients with comorbidities (p = 0.004) and postoperative complications (p < 0.001). Surgical approach significantly influenced DeepSeek's performance (p = 0.001), with better results in microscopic than endoscopic cases.
CONCLUSION: Both LLMs performed well in surgical planning but were less effective in postoperative management. Gemini demonstrated superior performance in postoperative recommendations, whereas DeepSeek benefited from clearer indicators of disease severity. These findings support the potential of LLMs as adjunctive tools in otologic surgical decision-making while emphasizing the need for expert clinical oversight.