Development and validation of an interpretable machine learning model for predicting chronic atrophic gastritis in...

Chronic atrophic gastritis (CAG) is a key precancerous condition, but its screening in the elderly is hindered by the invasiveness, high cost, and limited accessibility of endoscopy. This study aimed to develop and validate interpretable machine learning (ML) models for predicting CAG in elderly patients and identify key predictive factors.

We adopted a retrospective cohort study design using data from the Endoscopy Department of Shanghai Tenth People's Hospital. Two independent cohorts were included: a development cohort (January 2023 to December 2024, n = 1,268) for model construction and internal validation, and a temporal validation cohort (January 2025 to October 2025, n = 544) for external validation. A total of 28 candidate variables were collected, covering demographic characteristics, lifestyle and dietary habits, past medical history/medication history/family history, psychological status, and clinical symptoms.

Through multi-dimensional feature selection integrating univariate/multivariate logistic regression, recursive feature elimination (RFE), Boruta algorithm, and clinical expert review, eight key variables were identified. Nine ML models were constructed and optimized, with performance evaluated using metrics including area under the curve (AUC), sensitivity, specificity, and F1-score. The Multilayer Perceptron (MLP) model demonstrated optimal performance, with an area under the receiver operating characteristic curve (AUC) of 0.826 (95% CI: 0.788-0.864) in internal validation and 0.780 (95% CI: 0.745-0.815) in temporal validation. SHAP analysis revealed that Helicobacter pylori infection, age, smoking status, and high-salt pickled food intake were the top risk factors, while fruit and vegetable intake was a protective factor.

This study developed and validated an interpretable ML model for predicting CAG in elderly patients. The model, incorporating eight readily available clinical variables, provides a non-invasive, accessible tool for risk stratification and may facilitate early screening and targeted intervention in primary care settings.