Development and Validation of an Interpretable Machine Learning Model for Staging Helicobacter pylori-Initiated Intestinal-Type Gastric Cancer in the Correa Cascade: Cross-Sectional Study.
Gastric cancer (GC) is one of the most common malignant tumors worldwide, with Helicobacter pylori-associated intestinal-type gastric cancer (IGC) being the most prevalent subtype, accounting for approximately 85% of cases. Because most patients are diagnosed at intermediate or advanced stages, early screening and accurate stage stratification of IGC progression remain major clinical challenges.
This study aimed to develop an interpretable machine learning (ML) model that leverages routine laboratory indicators to perform stage-specific diagnosis for patients across different stages of IGC.
Data from 2180 patients with known H pylori infection status were collected at 2 centers and included healthy controls (HCs), nonatrophic gastritis, atrophic gastritis, intestinal metaplasia, and GC. After excluding cases with severe (>25%) missing data, 1784 patients were included for model development and validation. Data imputation and feature selection were performed, and synthetic minority oversampling technique (SMOTE) augmentation was applied to the internal training dataset to improve the diagnostic performance of the model. Six ML algorithms were developed. Model performance and clinical decision-making utility were evaluated using multiple metrics and approaches, while Shapley Additive Explanations (SHAP)-based interpretability was used to identify key indicators and provide threshold reference values for them. Finally, a web-based tool was developed based on the Streamlit platform.
Through feature selection, 27 features were ultimately retained for final model construction. Among the 6 algorithms, CatBoost (categorical boosting) demonstrated the best performance, achieving an internal validation accuracy of 80.91%, sensitivity of 78.57%, and specificity of 95.27%. In the Guangdong Provincial People's Hospital (GDPH) and Shengli Oilfield Central Hospital (SOCH) external validation cohorts, CatBoost maintained robust performance, with accuracies of 79.96% and 83.37%, sensitivities of 76.82% and 84.91%, specificities of 93.14% and 95.73%, and area under the curves (AUCs) of 0.94 and 0.97, respectively. Confusion matrix analysis showed that the model was particularly reliable in identifying extreme disease states, including HCs and GC, whereas misclassifications mainly occurred between adjacent intermediate pathological stages. Calibration curves and Brier scores indicated good agreement between predicted and observed outcomes. Decision curve analysis (DCA) further confirmed the clinical net benefit across relevant threshold ranges. SHAP-based interpretability analysis identified monocyte count (MONO%), albumin/globulin ratio (A/G), basophil percentage (BASO%), platelet distribution width (PDW), total bilirubin (DBIL), neutrophil count (NEUT#), age, lymphocyte count (LYMPH#), creatinine (CREA), and aspartate aminotransferase (AST) as important contributors, reflecting inflammatory, hematological, nutritional, and metabolic changes during IGC progression. Based on these features, a lightweight predictive model was developed and deployed as a web-based application to facilitate translational and practical applications.
This study developed an interpretable ML model based on routine laboratory data for stage-specific prediction of H pylori-associated IGC progression, with promising applicability as an auxiliary diagnostic tool.
This study aimed to develop an interpretable machine learning (ML) model that leverages routine laboratory indicators to perform stage-specific diagnosis for patients across different stages of IGC.
Data from 2180 patients with known H pylori infection status were collected at 2 centers and included healthy controls (HCs), nonatrophic gastritis, atrophic gastritis, intestinal metaplasia, and GC. After excluding cases with severe (>25%) missing data, 1784 patients were included for model development and validation. Data imputation and feature selection were performed, and synthetic minority oversampling technique (SMOTE) augmentation was applied to the internal training dataset to improve the diagnostic performance of the model. Six ML algorithms were developed. Model performance and clinical decision-making utility were evaluated using multiple metrics and approaches, while Shapley Additive Explanations (SHAP)-based interpretability was used to identify key indicators and provide threshold reference values for them. Finally, a web-based tool was developed based on the Streamlit platform.
Through feature selection, 27 features were ultimately retained for final model construction. Among the 6 algorithms, CatBoost (categorical boosting) demonstrated the best performance, achieving an internal validation accuracy of 80.91%, sensitivity of 78.57%, and specificity of 95.27%. In the Guangdong Provincial People's Hospital (GDPH) and Shengli Oilfield Central Hospital (SOCH) external validation cohorts, CatBoost maintained robust performance, with accuracies of 79.96% and 83.37%, sensitivities of 76.82% and 84.91%, specificities of 93.14% and 95.73%, and area under the curves (AUCs) of 0.94 and 0.97, respectively. Confusion matrix analysis showed that the model was particularly reliable in identifying extreme disease states, including HCs and GC, whereas misclassifications mainly occurred between adjacent intermediate pathological stages. Calibration curves and Brier scores indicated good agreement between predicted and observed outcomes. Decision curve analysis (DCA) further confirmed the clinical net benefit across relevant threshold ranges. SHAP-based interpretability analysis identified monocyte count (MONO%), albumin/globulin ratio (A/G), basophil percentage (BASO%), platelet distribution width (PDW), total bilirubin (DBIL), neutrophil count (NEUT#), age, lymphocyte count (LYMPH#), creatinine (CREA), and aspartate aminotransferase (AST) as important contributors, reflecting inflammatory, hematological, nutritional, and metabolic changes during IGC progression. Based on these features, a lightweight predictive model was developed and deployed as a web-based application to facilitate translational and practical applications.
This study developed an interpretable ML model based on routine laboratory data for stage-specific prediction of H pylori-associated IGC progression, with promising applicability as an auxiliary diagnostic tool.
Authors
Tang Tang, Chen Chen, Zhang Zhang, Tay Tay, Marshall Marshall, Ma Ma, Wang Wang
View on Pubmed