SHAP-explained machine-learning model for high-risk gastric cancer identification.

Area of Science:

• Gastroenterology
• Oncology
• Data Science

Background:

• Gastric cancer (GC) is a significant health issue, particularly in Asia.
• Effective screening strategies are needed, considering regional factors like Helicobacter pylori (H. pylori) infection and precancerous lesions such as atrophic gastritis (AG) and intestinal metaplasia (IM).

Purpose of the Study:

• To develop and validate a short-term (2-year) gastric cancer risk prediction model.
• To integrate endoscopic findings (AG/IM) with demographic and lifestyle factors for improved risk assessment.
• To compare the performance of machine learning models against conventional methods.

Main Methods:

• Utilized a large, real-world, nationwide screening cohort with AG/IM endoscopic codes.
• Developed and compared risk prediction models using Cox proportional hazards model (CPHM), extreme gradient boosting (XGBoost), decision tree (DT), and logistic regression (LR).
• Evaluated model performance through internal and external validation, assessing discrimination and calibration. Employed Shapley Additive Explanations (SHAP) for model interpretability.

Main Results:

• The XGBoost model exhibited superior performance in both internal (AUROC 0.764) and external (AUROC 0.708) validation.
• SHAP analysis identified H. pylori infection, age, sex, smoking, and AG/IM as key predictors of gastric cancer risk.
• The model demonstrated good discrimination and calibration, highlighting its potential clinical utility.

Conclusions:

• An interpretable, externally validated 2-year GC risk model incorporating AG/IM findings offers a practical tool for risk-adapted screening.
• This model can identify high-risk individuals for targeted surveillance and clinical review.
• Understanding the key contributing factors through SHAP enhances the model's transparency and clinical applicability.