A machine-learning model for staging autoimmune gastritis based on endoscopic and serological features

Area of Science:

• Gastroenterology
• Medical Informatics
• Oncology

Background:

• Autoimmune gastritis (AIG) lacks a standardized staging system, hindering clinical management.
• The AIG-atrophic stage (AIG-AS) is promising but subjective due to observer dependency.
• Accurate staging is crucial for predicting disease progression and associated complications.

Purpose of the Study:

• To develop and validate machine-learning models for objective staging of autoimmune gastritis.
• To compare the performance of different machine learning algorithms using endoscopic and serological features.
• To identify key predictors for accurate AIG staging.

Main Methods:

• A cross-sectional study of 203 AIG patients using integrated endoscopic (WLE, ME-NBI) and serological data.
• Seven machine learning algorithms were trained and evaluated using nested cross-validation.
• Patients were classified into three AIG-AS stages and a binary advanced vs. early-intermediate stage.

Main Results:

• A random forest model combining endoscopy and serology achieved the highest three-class AUC (0.878).
• A WLE+ME-NBI random forest model demonstrated high performance for binary staging (AUC 0.913).
• Key predictors included loss of gastric folds, gastrin-17, pepsinogen I, and crypt-opening grade.

Conclusions:

• Multi-modal machine learning provides accurate and interpretable staging for autoimmune gastritis.
• Endoscopic features are more predictive than serology alone, with combined models offering optimal performance.
• The high prevalence of Barrett esophagus in AIG underscores the need for routine esophageal surveillance.