Study on the application value of the large language model ChatGPT in electrical impedance tomography for lung ventilation imaging diagnosis
Pan Pan
Jiang Shuoran
Liu Yuhong
Zhao Mengchen
Gu Hongjun
Yang Qingyun
Xie Fei
Chen Qingcai
Abstract：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 1 215 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 0. 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.
Keywords：Electrical impedance tomographyLung ventilationHypoxemiaLarge language modelChatGPT
Publication Date：2023-10-10
Online Publishing Date：2025-08-15(First online date of this platform, not the publication date of the document)
Pages：8( 760-767 )