TOPLINE:
An artificial intelligence (AI) model integrating both clinical information and endoscopic ultrasonographic (EUS) images can aid in the differential diagnosis of solid pancreatic lesions, new research shows.
METHODOLOGY:
- Researchers in China used clinical information and EUS images from 439 patients with solid pancreatic lesions to train and validate an AI model’s ability to distinguish cancer from noncancerous pancreatic lesions.
- In a randomized crossover trial, 12 endoscopists with varying levels of expertise from four centers across China diagnosed solid pancreatic lesions with or without assistance from the AI model.
- Researchers tested the model’s performance internally and externally using retrospective datasets with 628 patients and a prospective dataset with 130 patients. They compared diagnostic accuracy of endoscopists with and without AI assistance.
TAKEAWAY:
- The AI model demonstrated robustness across internal and external cohorts, with the area under the curve of 0.996 in the internal test dataset and between 0.924 and 0.976 in external testing.
- The diagnostic accuracy of novice endoscopists was significantly enhanced with AI assistance, increasing from 0.69 to 0.90 (P < .001).
- Expert and senior endoscopists exhibited a greater tendency than novice endoscopists to reject the AI model’s predictions. However, after supplementing the results of interpretability analyses, the acceptance of AI predictions increased among more experienced endoscopists, becoming generally comparable to that of novice endoscopists.
IN PRACTICE:
“This study suggests that endoscopists of varying expertise can efficiently cooperate with this multimodal AI model, establishing a proof-of-concept study for human-AI interaction in the management of solid lesions in the pancreas,” the authors wrote.
SOURCE:
The study, with first author Haochen Cui, MD, Department of Gastroenterology and Hepatology, Tongji Hospital, Wuhan, China, was published online in JAMA Network Open.
LIMITATIONS:
The crossover trial was conducted in a simulated environment and the superior performance of the model cannot be directly applied to clinical practice. The prospective dataset was relatively small. Further research is needed in a larger and more diverse patient pool to further assess the clinical applicability of the model.
DISCLOSURES:
The study was supported by grants from the National Natural Science Foundation of China. Two of the authors reported being employed by Wuhan EndoAngel Medical Technology Co, with no other relevant conflicts of interest noted.
