Ishida Masanori, Gonoi Wataru, Nyunoya Keisuke, Abe Hiroyuki, Shirota Go, Okimoto Naomasa, Fujimoto Kotaro, Kurokawa Mariko, Katayama Akira, Takahashi-Mizuki Masumi, Inui Shohei, Saito Kazuhiro, Ushiku Tetsuo, Abe Osamu
Radiology, Tokyo Medical University Hospital, Tokyo, JPN.
Radiology, The University of Tokyo Hospital, Tokyo, JPN.
Cureus. 2025 May 8;17(5):e83721. doi: 10.7759/cureus.83721. eCollection 2025 May.
This study evaluates the diagnostic performance of a large language model (LLM) in determining causes of death by comparing three different information sources.
A total of 150 consecutive adult in-hospital cadavers underwent postmortem CT and pathological autopsy (2009-2013). The diagnostic accuracy of Claude 3.5 Sonnet (Anthropic, San Francisco, California) was evaluated in determining both underlying and immediate causes of death using three different information sources (clinical history alone, postmortem CT findings alone as documented by radiologists in their reports, and their integration). For each case, the LLM provided a primary diagnosis and two differential diagnoses. The autopsy result was used as the reference standard to assess accuracy.
For underlying causes, the integration of both sources achieved significantly higher accuracy (78.0%) compared with the clinical history alone (69.3%) or the CT findings alone (42.0%) (p<0.001). When considering either primary or differential diagnoses, the accuracy reached 84.7% with integrated sources, 78.0% with clinical history alone, and 58.7% with CT findings alone. For immediate causes, the integrated approach showed higher accuracy in the primary diagnosis (61.3%) than the clinical history alone (52.0%) and CT findings alone (46.7%) (p<0.001). Disease-specific diagnostic accuracy analyses revealed marked variations, with hematologic malignancies showing the most significant differences among information sources (clinical history: 78.9%, CT findings alone: 36.8%, integrated analysis: 85.7%; p=0.003).
Integrating postmortem CT findings with clinical history enhances LLM-based cause-of-death determination accuracy, demonstrating the value of multiple information sources while highlighting opportunities for disease-specific diagnostic optimization.
本研究通过比较三种不同的信息来源，评估了大语言模型（LLM）在确定死因方面的诊断性能。
共有150例连续的成年住院尸体接受了尸检CT和病理解剖（2009 - 2013年）。使用三种不同的信息来源（仅临床病史、放射科医生在报告中记录的仅尸检CT结果以及两者结合），评估了Claude 3.5 Sonnet（Anthropic，加利福尼亚州旧金山）在确定根本死因和直接死因方面的诊断准确性。对于每个病例，大语言模型提供一个初步诊断和两个鉴别诊断。尸检结果用作评估准确性的参考标准。
对于根本死因，与仅临床病史（69.3%）或仅CT结果（42.0%）相比，两种来源结合的准确性显著更高（78.0%）（p<0.001）。在考虑初步诊断或鉴别诊断时，结合来源的准确性达到84.7%，仅临床病史为78.0%，仅CT结果为58.7%。对于直接死因，综合方法在初步诊断中的准确性（61.3%）高于仅临床病史（52.0%）和仅CT结果（46.7%）（p<0.001）。特定疾病的诊断准确性分析显示出显著差异，血液系统恶性肿瘤在信息来源之间的差异最为显著（临床病史：78.9%，仅CT结果：36.8%，综合分析：85.7%；p = 0.003）。
将尸检CT结果与临床病史相结合可提高基于大语言模型的死因确定准确性，证明了多种信息来源的价值，同时突出了特定疾病诊断优化的机会。
