基于ChatGLM急性缺血性卒中大血管闭塞的识别与诊断 Identification and Diagnosis of Large Vessel Occlusion in Acute Ischemic Stroke Based on ChatGLM

Abstract

目的 探索基于急诊电子病历信息利用大语言模型ChatGLM在急性缺血性卒中患者中进行大血管闭塞识别及诊断的可行性和准确性。 方法 选取2016年1月—2024年1月连续就诊于清华大学附属北京清华长庚医院急诊神经科的发病在24 h内的急性缺血性卒中病例，根据头颈部血管检查（CTA/MRA/DSA）区分大血管闭塞和非大血管闭塞。依托于ChatGLM模型，利用提示词功能、指令微调、检索增强生成等技术给出推理过程和诊断结果，并探索不同的推理过程判断大血管闭塞的准确率、敏感性和特异性，由Python统一实现，用于表示模型的推理性能。 结果 共纳入连续就诊的急性缺血性卒中患者935例，其中大血管闭塞230例，采用零样本学习、零样本学习+思维链、生成式预训练模型+思维链、少样本学习及少样本学习+思维链推理诊断大血管闭塞的准确率分别为36.1%、52.1%、73.0%、72.6%和75.1%。 结论 大语言模型ChatGLM在病例诊断和推理方面具有一定的可行性，可基于电子病历文本进行急性大血管闭塞性卒中的识别和诊断，采用少样本学习的诊断准确性总体要显著高于零样本学习。 Abstract: Objective To explore the feasibility and accuracy of using the large language model ChatGLM for the identification and diagnosis of large vessel occlusion in patients with acute ischemic stroke based on emergency electronic medical record information. Methods This study selected patients with acute ischemic stroke that occurred within 24 hours and were consecutively treated at the Emergency Neurology Department of Beijing Tsinghua Changgung Hospital, Tsinghua University, from January 2016 to January 2024. Large vessel occlusion and non-large vessel occlusion were distinguished according to head and neck vascular examinations (CTA/MRA/DSA). Based on the ChatGLM model, the inference processes and diagnosis results were given by using techniques such as prompt function, instruction fine-tuning, and retrieval-augmented generation, and the accuracy, sensitivity, and specificity of different inference processes for determining large vessel occlusion were explored. This was uniformly implemented in Python to represent the model’s inference performance. Results A total of 935 patients with acute ischemic stroke were included, including 230 patients with large vessel occlusion. The diagnostic accuracy rates for large vessel occlusion using zero-shot learning, zero-shot learning + chain-of-thought, generative pre-trained transformer+chain-of-thought, few-shot learning, and few-shot learning + chain-of-thought were 36.1%, 52.1%, 73.0%, 72.6%, and 75.1%, respectively. Conclusions The large language model ChatGLM demonstrates a certain feasibility in case diagnosis and inference. It can identify and diagnose emergency large vessel occlusion strokes based on electronic medical record texts. The diagnostic accuracy using few-shot learning is significantly higher than zero-shot learning.

Keywords

• 缺血性卒中
• 大血管闭塞
• 大语言模型
• 推理
• ischemic stroke
• large vessel occlusion
• large language model
• inference