人工智能在复杂临床诊断中的应用：一例以隐匿性记忆力减退为首发表现的中枢神经系统生殖细胞瘤——基于大语言模型在多阶段复杂临床诊疗中的推理能力系统评估

Applying Artificial Intelligence to a Complex Diagnostic Challenge: A CNS Germ Cell Tumor Presenting as Insidious Memory Loss—A Systematic Evaluation of Large Language Models’ Reasoning Abilities across Multi-Stage Complex Clinical Workups

摘要: 目的：中枢神经系统生殖细胞瘤(CNS-GCTs)极少表现为单纯性认知症状，临床上极易导致漏诊与误诊。本文报道了一例相关罕见病例，并借此评估了五种人工智能(AI)大语言模型的临床诊断推理能力。方法：患者为16岁男性，因“进行性记忆力减退一年”就诊，初步检查提示严重高钠血症，最终确诊为鞍上中枢神经系统生殖细胞瘤。本研究将该病例三次就诊阶段的临床资料分别输入五种AI模型中，并由20名精神科医生采用修订版IDEA评分量表(Revised IDEA Score)对模型生成的推理过程进行量化评分。结果：放疗后患者肿瘤完全消退，但记忆损害表现持续存在。在AI模型评估中，尽管有三款模型展现出较高且稳定的临床推理质量，但整体诊断准确率仅为40.0%，最终仅有两款模型正确诊断出该肿瘤。结论：对于出现隐匿性认知功能减退的青少年患者，临床应高度警惕CNS-GCTs的可能。虽然先进的AI能够模拟高水平的临床推理，但在面对非典型病例时，其诊断准确率仍存在局限性。这表明在“复制推理逻辑”与“确保诊断正确”之间，AI仍存在关键技术鸿沟。未来的研究需进一步深挖人工智能在医疗场景中的应用，以期实现高质量、高精度的临床决策。

Abstract: Objective: Central nervous system germ cell tumors (CNS-GCTs) rarely present with isolated cognitive symptoms, causing significant diagnostic delays. We report such a case to evaluate the diagnostic reasoning of five leading artificial intelligence (AI) large language models. Methods: A 16-year-old male with a one-year history of progressive memory impairment was diagnosed with a suprasellar CNS-GCT after a workup revealed severe hypernatremia. Clinical data from three diagnostic stages were provided to five AI models. Twenty psychiatrists scored the AI-generated reasoning using the validated the Revised IDEA Score. Results: After radiotherapy, the patient’s tumor resolved, but his memory impairment persisted. In the AI evaluation, while three of the five models consistently demonstrated high-quality clinical reasoning, the overall diagnostic accuracy was only 40.0%, with just two models correctly identifying the tumor. Conclusion: CNS-GCTs should be suspected in adolescents with insidious cognitive decline. Although advanced AI can emulate high-level clinical reasoning, its diagnostic accuracy in atypical cases remains unreliable. This demonstrates a critical gap between replicating a reasoning process and achieving diagnostic certainty. Future research should further explore the specific applications of artificial intelligence in medical settings to enable high-quality and highly accurate clinical decisions.