大语言模型ChatGPT在电阻抗断层显像(EIT)肺通气诊断中的应用价值

Study on the application value of the large language model ChatGPT in electrical impedance tomography for lung ventilation imaging diagnosis

目的使用ChatGPT对电阻抗断层显像(EIT)肺通气参数进行分析并提供诊断。方法回顾中国人民解放军总医院呼吸与危重症医学中心2020年9月至2023年6月的EIT数据,使用GPT-3.5和GPT-4进行EIT数据的自动判断,采用零样本学习为基础,并使用上下文学习测试以提高诊断准确度。在统计方面使用敏感度和特异度来评估模型性能。结果ChatCPT对EIT具有一定的知识储备。本研究纳人530例患者的1215份EIT检查数据,显示ChatGPT对EIT数据的自动判断效果并不理想:零样本数据提示,与人工诊断相比一致性不佳;上下文学习的方法提示,GPT-4虽然比GPT-3.5的逻辑推理能力有了显著提升,但GPT4对指令的遵循能力却显著下降。GPT-4最终对肺通气诊断准确率只有27.02%,在“轻度通气功能缺失”上的预测最高,但也只达到53.58%。“无特殊”“重度通气功能缺失”的敏感度更底,分别为21.72%和13.71%。遗憾的是,“中度通气功能缺失”的敏感度为0。结论ChatCPT并没有掌握EIT测量参数来诊断通气功能缺失。尽管上下文学习能够提高模型的诊断能力,但结果仍然不够理想,未来需要纳人更多的特征和更复杂的模型来提高自动判断的准确度。

Objective To analyze electrical impedance tomography(EIT)ventilation parameters and provide the diagnosis by using ChatGPT.Methods This study reviewed EIT data from the Respiratory and Critical Care Medicine Center of the Chinese PLA General Hospital.GPT-3.5 and GPT-4 were used for automated EIT data interpretation,based on zero-shot learning,and context learning tests were employed to improve diagnostic accuracy.Sensitivity and specificity were used for statistical evaluation.Results ChatGPT had a certain knowledge base for EIT.The study included 1215 EIT examination data from 530 patients.The results showed that ChatGPT's automated interpretation of EIT data was suboptimal.The zero-shot data prompted there was poor consistency in comparison to manual diagnosis.Although context learning methods indicated significant improvements in logical reasoning abilities for GPT-4 compared to GPT-3.5,GPT-4's compliance with instructions significantly decreased.The prediction for"mild ventilatory dysfunction"was the highest,but it only reached 53.58%.The sensitivities of"no special"and"severe ventilatory dysfunction"are even lower,21.72%and 13.71%respectively.Unfortunately,the sensitivity of"moderate ventilatory dysfunction is O.Conclusions ChatGPT does not possess the knowledge to diagnose ventilation dysfunction based on EIT measurement parameters.Although context learning can enhance model diagnostic capabilities,the results are still suboptimal,and future work will require the inclusion of additional features and more complex models to improve the accuracy of automated interpretation.