Machine learning models using serum gastric biomarkers for the non-invasive prediction of atrophic gastritis: a comparative study

Background and aims The early, non-invasive detection of chronic atrophic gastritis (CAG), a precancerous lesion, remains a clinical challenge. While serological biomarkers are promising alternatives to endoscopy for screening, their predictive accuracy using conventional methods is suboptimal. This study aimed to identify key predictors of CAG and to comparatively develop multiple machine learning (ML) models, evaluating whether ML offers a definitive advantage and identifying a reliable model for triaging patients to endoscopy. Methods In this retrospective diagnostic study (conducted from January to October 2020), 222 subjects (CAG prevalence: 30.6%) were stratified randomly into a training set (80%) and an independent test set (20%). Feature selection was performed exclusively on the training set using multivariate logistic regression, which identified four independent predictors: PGI, the PGI/PGII ratio, age, and anti- H. pylori antibody status. Using these predictors, eight models—including Logistic Regression (as baseline), Elastic Net, Support Vector Machine, Neural Network, and tree-based ensembles—were trained and optimized via 5-fold cross-validation. Model performance was rigorously evaluated on the held-out test set using discrimination (AUC, sensitivity, specificity), calibration (Brier score), and clinical utility (Decision Curve Analysis). Results Multivariable analysis identified the four predictors, with anti- H. pylori antibody positivity associated with an approximately four-fold higher odds of CAG. On the independent test set, the Elastic Net (AUC = 0.823) and Logistic Regression (AUC = 0.810) models demonstrated the highest and most robust discriminative performance, showing excellent sensitivity (0.923) and negative predictive value (>0.95) for ruling out CAG. Statistical comparison confirmed that their AUCs were significantly higher than those of the severely overfitted tree-based models (e.g., Random Forest), but not significantly different from other complex models like Support Vector Machine. Decision Curve Analysis confirmed the superior net clinical benefit of the Elastic Net and Logistic Regression models across a wide range of decision thresholds. Conclusion Simple, interpretable linear models (Elastic Net and Logistic Regression) based on four routine clinical parameters provide a robust tool for the non-invasive identification of CAG in a clinical population referred for endoscopic evaluation. They show particular strength in ruling out disease, supporting their potential role as a triage tool. In this setting, they demonstrated more consistent performance than more complex machine learning algorithms. External validation in broader populations is warranted to confirm generalizability before clinical implementation.