Can artificial intelligence improve the diagnosis and management of patients with eosinophilic esophagitis?

Table 3 Comparative table of the main artificial intelligence models applied to endoscopic diagnosis of esophagitis diagnosis

Model	Training dataset	Population size	Validation status	Key limitations	Methodological rigor
AI-EoE-EREFS (DL/CNN)[48]	Endoscopic images, EREFS-labeled, multi-center	484 images from 134 patients	Retrospective, some external validation	Retrospective data, limited diversity, mainly binary classification, limited real-world scenarios	High accuracy (AUC = 0.992), but generalizability and real-world robustness uncertain
CNN[47]	Endoscopic images (EoE vs controls)	1192 characteristic endoscopic images of 108 patients	Tested against patients and controls	Dataset size not reported, unclear diversity, retrospective design	High accuracy (95%), but limited transparency and unclear reproducibility
CNN with EREFS integration[49]	Endoscopic images annotated with EREFS	200 WLIs, including 100 WLIs from EoE patients and 100 WLIs of normal esophagus	Compared to human experts	Training/test data overlap possible, unclear external validation	Sensitivity improved to 85%, specificity 95% with EREFS
Random decision forest[50]	Clinical + endoscopic data	Not specified	Real-world data, AUC up to 0.94	Complexity, integration challenges, population diversity not detailed	Robust with multi-source inputs, but needs more transparent validation
General endoscopic AI (DL/CNN)	Mixed datasets (WLI, NBI, CT images)	Varies; often small	Mostly retrospective, single/multi-center	Data heterogeneity, lack of multi-center, prospective validation	High reported accuracy, but shallow validation, risk of bias due to dataset curation

AI: Artificial intelligence; EoE: Eosinophilic esophagitis; DL: Deep learning; EREFS: Eosinophilic esophagitis endoscopic reference score; CNN: Convolutional neural network; WLI: White light image; NBI: Narrow band imaging; CT: Computed tomography; AUC: Area under the curve.