Machine Learning Models Using Serum Gastric Biomarkers for the Non-Invasive Prediction of Atrophic Gastritis: A Comparative Study

Abstract

Background & Aims: Chronic atrophic gastritis (CAG) is a precancerous lesion for which early, non-invasive detection is clinically needed. We aimed to identify key predictors of CAG and to compare machine learning (ML) models to establish a robust, non-invasive prediction tool, specifically evaluating whether ML offers a definitive advantage in this setting. Methods: In this retrospective diagnostic study, 222 subjects were split into training (80%) and independent test (20%) sets. Feature selection on the training set identified four independent predictors: PGI, PGI/PGII ratio, age, and anti-H. pylori status. Eight models were trained and optimized via cross-validation, with performance rigorously evaluated on the test set. Results: Elastic Net (AUC=0.823) and Logistic Regression (AUC=0.810) demonstrated the highest and most robust discriminative performance on the test set, showing excellent sensitivity (0.923) for ruling out CAG. Their performance was significantly better than severely overfitted tree-based models. Decision Curve Analysis confirmed their superior net clinical benefit across a wide threshold range. Conclusions: Simple, interpretable linear models based on four routine parameters provide a robust tool for non-invasive CAG identification in a clinical population referred for endoscopy. They are particularly strong for ruling out disease, supporting a potential triage role. In this setting, they demonstrated more consistent performance than more complex ML algorithms. External validation is warranted before clinical implementation.