This retrospective study aimed to construct and validate a predictive model for classifying disease stages along the H pylori–associated Correa cascade. Data were obtained from the laboratory information systems (LIS) of 2 independent, large-scale, grade 3A hospitals, Guangdong Provincial People’s Hospital (GDPH) and Shengli Oilfield Central Hospital (SOCH), between January 2022 and May 2025. The collected data included demographic information, complete blood counts, and serum biochemical markers. H pylori infection status was confirmed through gastric mucosal histological examination. Participants were classified into 5 groups, including healthy control (HC), nonatrophic gastritis (NAG), AG, IM, and GC. All diagnoses were independently determined by 2 experienced pathologists according to pathological standards. HCs were recruited from individuals undergoing routine health examinations who had normal upper gastrointestinal endoscopic findings and negative H pylori test results. The same exclusion criteria were applied to both HCs and patients across all disease stages, including a history of gastric resection or other gastrointestinal surgeries, coexisting malignancies, a history of organ transplantation, or severe heart, lung, liver, kidney, or hematologic diseases.

Features with more than 25% missing values were excluded. Missing values in the remaining features were imputed using 6 methods, including mean imputation, median imputation, k-nearest neighbor (KNN) imputation, linear interpolation, multiple imputation, and regression imputation. The imputation results were compared using box plots and histograms to evaluate the impact of different imputation methods on the data distribution and statistical characteristics. Furthermore, considering the potential influence of multicollinearity between features on prediction accuracy, features with a high correlation in the Spearman correlation analysis were excluded if they had a lower correlation with the target variable. Ultimately, the following features were retained, including gender, age, aspartate aminotransferase (AST), albumin (ALB), albumin/globulin ratio (A/G), total bilirubin (DBIL), alkaline phosphatase (AKP), gamma-glutamyl transferase (GGT), glucose (GLU), urea (UREA), creatinine (CREA), uric acid (UA), total cholesterol (TC), triglycerides (TG), high-density lipoprotein cholesterol (HDL), monocyte percentage (MONO%), basophil percentage (BASO%), neutrophil count (NEUT#), lymphocyte count (LYMPH#), monocyte count (MONO#), eosinophil count (EO#), red blood cell (RBC) count, mean corpuscular hemoglobin (MCH), mean corpuscular hemoglobin concentration (MCHC), platelet count (PLT), mean platelet volume (MPV), and platelet distribution width (PDW). To address the issue of class imbalance, we used the synthetic minority over-sampling technique (SMOTE) to generate synthetic samples. SMOTE was applied only to the internal training set, whereas the internal validation set and the 2 external validation cohorts were kept unaugmented. By reducing the bias toward the majority class during model training, this approach enables more stable learning of minority-class patterns and effectively improves the model’s sensitivity to rare but clinically meaningful cases [25,26]. In subsequent analyses, we compared and evaluated the effectiveness of this method.

This study was conducted using Python 3.9.7, along with the scikit-learn, catboost, lightgbm, and xgboost packages for training ML models. The 6 ML algorithms used included adaptive boosting (AdaBoost, AdaBoostClassifier), categorical boosting (CatBoost, CatBoostClassifier), decision tree (DT, DecisionTreeClassifier), light gradient boosting machine (LGBM, LGBMClassifier), random forest (RF, RandomForestClassifier), and eXtreme gradient boosting (XGBoost, XGBClassifier). The training cohort from 2 hospitals was divided into 2 subsets, with 80% allocated as the internal training set and 20% as the internal validation set. Before formal training, all models underwent hyperparameter tuning using GridSearchCV to fit all combinations of model parameters (Multimedia Appendix 1), followed by 5-fold cross-validation (CV) for optimal hyperparameter selection. Each model was then trained using the optimal parameter combination. Performance was evaluated via accuracy, macro precision, specificity, macro recall (sensitivity), and macro F₁-score. Receiver operating characteristic (ROC) curve and confusion matrix were also generated to evaluate discrimination and class-specific performance. All models followed the same construction and optimization process to assess the quality of the new data generated by SMOTE preprocessing, thus validating the feasibility of this approach. The best diagnostic model, selected after comparison of all models, was further evaluated using a calibration curve to determine whether the model’s predicted probabilities align with the actual outcomes. Additionally, the model’s potential clinical value was measured using decision curve analysis (DCA). Finally, an external validation cohort was simultaneously collected from both hospitals, with inclusion and exclusion criteria that matched those of the training cohort. The external validation data were entered into the best diagnostic model, and the same evaluation metrics were used to assess the model’s performance on unseen data.

To effectively interpret the model’s decision results, the study uses Shapley Additive Explanations (SHAP) for interpretability analysis. SHAP values for each feature are calculated using TreeExplainer from the SHAP library. Feature importance is ranked using the summary_plot, which provides a visual representation of the overall contribution of each feature to the model’s predictions. The dependence_plot function further analyzes the impact of individual features on model predictions, examining the trend of feature changes across different categories and their relationship with the SHAP values. The ridge plot displays the distribution of each feature across different groups, visually highlighting the differences between categories.

To facilitate the clinical application of the developed method, we integrate the best model, constructed from the top 10 features identified by SHAP, into a tool developed using the Streamlit framework. A web page deployment tool is created for predictive analysis at different stages of the Correa model. Users can input the corresponding feature values into the model and click the “Predict” button. The tool will return the prediction result and save it as the most recent prediction record.

This study was approved by the ethics review committee of Guangdong Provincial People’s Hospital (KY2025-1060-01). As this was a retrospective study, the requirement for informed consent was waived. To ensure privacy protection, all data were deidentified using participant codes, and no identifiable information was disclosed.

This retrospective analysis included 2180 patients at different stages of H pylori–associated GC along the Correa cascade, who were enrolled in the cohort for predictive model construction. During the study, GDPH and SOCH excluded 200 and 196 patients with extensive missing data, respectively. Ultimately, data from 1784 patients were used for model training and external validation. The training cohort included 1644 patients, consisting of 53 HC, 723 NAG, 181 AG, 606 IM, and 81 GC (Table 1). Baseline information after SMOTE augmentation is shown in Multimedia Appendix 2. At GDPH and SOCH, data from 50 and 90 patients were collected, respectively, as external validation cohorts. The overall framework of the study is illustrated in Figure 1, and the detailed study design workflow is provided in Multimedia Appendix 3.

This study included data from patients with IGC at different stages, collected at GDPH and SOCH. The data were divided into 5 groups: HC, NAG, AG, IM, and GC. The baseline demographic and clinical characteristics of the training cohort are shown in Table 1. Significant differences in gender distribution and age were observed among the 5 groups, with both variables showing statistical significance (P<.001). In the disease groups, the proportion of males was higher than in the HC group, particularly in the IM and GC groups. Age progressively increased with disease severity; the highest age values were observed in the GC group and relatively lower values in the NAG group. In contrast, indicators related to metabolism and liver function (AST, AKP, GGT, GLU, TC, and TG) did not show significant differences among the groups (P>.05). Several biochemical and hematological parameters exhibited significant differences across disease stages and demonstrated trends associated with different stages of IGC progression. As the disease progressed, ALB levels and the A/G ratio declined overall, reaching the lowest values in the GC group, while renal function indicators (including UREA and CREA) gradually increased, peaking in the GC group. UA levels increased in the early disease stages but decreased in the GC group. HDL levels were significantly lower in the disease groups compared with the HCs. Peripheral blood cell parameters also showed stage-related changes. As disease severity increased, NEUT# and MONO# progressively increased, reaching their highest values in the GC group, whereas LYMPH counts were relatively higher in the precancerous stages but decreased in the GC group. RBC counts were significantly reduced in the GC group, while MCH and MCHC showed mild increasing trends. PLT was relatively elevated in the GC group, while MPV and PDW decreased with disease progression; all these differences were statistically significant. The detailed number of patients in each group and the characteristics of the respective indicators for the 2 external validation cohorts are provided in Multimedia Appendices 4 and 5.

To address missing values, 6 imputation methods were applied, and the effects on feature distributions were systematically compared. Using the A/G index as an example, the boxplot results showed similar medians and IQRs across all imputation methods, with no obvious systematic shifts observed (Figure 2A). Further density distribution analyses showed that mean and median imputation smoothed the distributions to some extent, whereas linear, multiple, and regression imputation introduced varying degrees of adjustment to the local density structure (Figure 2B). In contrast, the distribution obtained by KNN imputation was highly consistent with the original data, fully preserving the pronounced bimodal structure and introducing almost no additional artifacts. Therefore, KNN imputation was selected as the optimal method, and the KNN-imputed dataset was used in all subsequent analyses. After completing missing-value imputation, the effects of the original data and SMOTE-augmented data on model performance were further compared. Compared with the nonaugmented data, the model accuracy after SMOTE augmentation increased from approximately 63% to about 80%, indicating that this approach effectively improves the model’s ability to identify minority-class samples (Figure 2C). To assess whether data augmentation introduced potential distributional shifts, a comparative analysis of feature distributions between the original data and the SMOTE-augmented data was conducted (Figure 2D). The results demonstrate that the augmented data remained highly consistent with the original data in overall distribution shape, central tendency, and value range, with no apparent distributional distortion observed. Considering the potential presence of multicollinearity among clinical indicators, correlations among features were further calculated and visualized using a feature correlation heatmap (Figure 2E). Subsequently, the effects of different correlation coefficient thresholds on model performance were evaluated (Figure 2F), and the detailed performance metrics for each threshold are provided in Multimedia Appendix 6. The results indicate that when the correlation coefficient threshold was set to 0.65, the overall model performance reached its optimum, with all evaluation metrics stably maintained at approximately 80%, while retaining 27 features. Therefore, subsequent model construction and analyses were based on this selected feature set.

The results show that CatBoost performed the best on the internal training data (Table 2), achieving the highest accuracy of 0.9986 (95% CI 0.9978‐0.9993), sensitivity of 0.9997 (95% CI 0.9994‐0.9998), and specificity of 0.9986 (95% CI 0.9977‐0.9993). However, the mean accuracy across 5-fold cross-validation was 0.8087 (SD 0.0024), indicating overfitting and suggesting that the performance may be lower in actual applications. In contrast, AdaBoost performed the worst on the internal training data, with an accuracy of 0.5655 (95% CI 0.5543‐0.5765), which may reflect its limited ability to handle strong correlations among features. In the internal validation data, this overfitting phenomenon was further validated. Although its performance decreased, CatBoost still showed the best performance, with an accuracy of 0.8091 (95% CI 0.7972‐0.8211), specificity of 0.9527 (95% CI 0.9496‐0.9559), and sensitivity of 0.7857 (95% CI 0.7743‐0.7968), demonstrating ideal results. Other models, such as LGBM, also performed well in the internal validation, achieving an accuracy of 0.8067 (95% CI 0.7933‐0.8195), which is similar to that of CatBoost. During the external validation stage, data from GDPH and SOCH were used for initial transportability assessments to evaluate the model’s performance in independent cohorts. Although CatBoost showed some fluctuations in these 2 external validation cohorts, the accuracy for the GDPH cohort was 0.7996 (95% CI 0.6800‐0.9000), sensitivity was 0.7682 (95% CI 0.6866‐0.8518), and specificity was 0.9314 (95% CI 0.8877‐0.9678); for the SOCH cohort, the accuracy was 0.8337 (95% CI 0.7553‐0.9111), sensitivity was 0.8491 (95% CI 0.7812‐0.9082), and specificity was 0.9573 (95% CI 0.9351‐0.9768). The performance was similar to that in the internal validation data, providing preliminary evidence of model transportability for clinical application.

In external validation, the CatBoost model demonstrated good discriminative performance across 2 independent centers. In the GDPH external validation cohort, the ROC curve (area under the curve [AUC]) was 0.94 (Figure 3A), while the AUC of the SOCH external validation cohort was 0.97 (Figure 3B), indicating promising discriminative ability of the model at both centers. The confusion matrices illustrated the detailed classification performance across different disease stages. In the GDPH cohort, the correct classification rate for AG was 87%, with approximately 13% of AG samples misclassified into the adjacent IM stage (Figure 3C). The IM category showed the best recognition performance, achieving a 100% correct classification rate. The correct classification rate for GC was 70%, with the remaining 30% mainly misclassified as AG (20%) and IM (10%). In the SOCH cohort, the classification accuracies for HC and GC both reached 100% (Figure 3D). The correct classification rates for NAG and AG were 80% and 90%, respectively, with a small number of samples showing cross-misclassification between NAG and AG. The correct classification rate for IM was 65%, with the remaining 35% misclassified as NAG. Misclassifications in this cohort were concentrated in intermediate pathological stages, suggesting that the model was capable of identifying extreme states (HC and GC). The consistency of probability predictions in the external validation cohorts was assessed using calibration curves and Brier scores. In the GDPH cohort (Figure 3E), the calibration curve closely approximated the ideal reference line, with a Brier score of 0.0257. In the SOCH cohort, the Brier score was 0.0830 (Figure 3F), indicating good agreement between predicted probabilities and observed outcomes. DCA demonstrated that the CatBoost model provided potential clinical utility in both external validation centers. In the GDPH cohort (Figure 3G), when the threshold probability ranged from approximately 0.15 to 0.9, the net benefit of the model consistently exceeded that of the “treat-all” and “treat-none” strategies, with stable positive gains observed in the low-to-moderate threshold range. In the SOCH cohort (Figure 3H), the model maintained a significant net benefit advantage across a wider threshold probability range (approximately 0.02‐0.98), suggesting its potential net benefit advantage. Collectively, these findings suggest that the CatBoost model has the potential to provide effective support for clinical diagnosis.

SHAP explainability analysis is used to gain further insight into the model’s decision-making process. Figure 4A presents the top 20 most important features in the model’s decision-making process, quantifying the significance of these features at different stages. For instance, MONO% plays a crucial role at the IM and GC stages, while DBIL is particularly important at the GC stage. Figure 4B shows how the top 10 features influence the decision boundary of the predicted outcomes. The results reveal that different features exhibit distinct threshold effects across different disease stages. For example, the threshold range for MONO% is between 3 and 7.5, which closely aligns with its clinical reference range (3-10), while the threshold range for DBIL is between 2 and 6, also close to its clinical reference range (0‐6.8). These findings suggest that diagnostic models built on these indicators have the potential to assist in disease screening and staging. The density distribution plot in Figure 4C further illustrates the distribution of the top 10 features across different stages. For example, indicators such as MONO%, BASO%, and NEUT# show significant differences in distribution and intensity between the HC and GC groups, confirming that these features play a crucial role in distinguishing between different disease stages. Additionally, since age is among the top 10 most important features, we further evaluated the model’s performance across different age strata in 2 external validation cohorts. The results showed that the CatBoost model continued to effectively differentiate patient data across various age strata. Detailed results are provided in Multimedia Appendix 7.

To maximize efficient use of the model and enhance its translational potential, we embedded a model constructed from the top 10 features ranked by SHAP into the Streamlit platform (Figure 5). Users need to enter only the 10 required features, and the application will automatically predict the stage of gastric disease progression for a given patient. This web tool is freely accessible through direct access to the Streamlit platform [27].

This study developed and validated an explainable ML-based model for predicting different stages in the progression of H pylori–initiated IGC. The model demonstrated robust performance across internal validation and external cohorts, suggesting that the ML-based model may serve as a potential tool for IGC staging. It supports early screening and facilitates informed clinical decision-making for subsequent evaluation and treatment.

Currently, most diagnostic models for gastric diseases based on routine laboratory indicators focus on binary classification at a single disease stage. Previous studies have primarily focused on discriminating between IM and NAG [28], the prediction of future GC risk in initially negative individuals [29], the assessment of laboratory parameter changes before and after treatment in patients with GC [30], or differentiation between GC and precancerous lesions [31]. In contrast, dynamic changes in hematological parameters across different stages of H pylori–initiated IGC remain poorly understood, as well as the feasibility of using these parameters for multistage prediction models also remains unclear.

In this study, we constructed a multiclass ML model for predicting different developmental stages of IGC. Among the evaluated models, CatBoost achieved the best diagnostic performance, with an accuracy of 0.8091 in the internal validation cohort and accuracies of 0.7996 and 0.8337 in 2 independent test sets, respectively. DCA further supported its potential clinical applicability. However, the relatively large number of features included in the current model may limit its practical implementation. Therefore, it will be important to develop more lightweight models and identify a more concise set of key features while maintaining predictive performance.

Given the lack of unified guidelines for feature selection in predictive models, SHAP analysis was used to provide both global and local explanations of the model, thereby elucidating the contribution of individual features to the model’s predictions [32]. This study demonstrates that routine laboratory indicators, including MONO%, A/G, BASO%, PDW, and DBIL, play important roles in distinguishing different disease stages. Higher MONO% was associated with an increased predicted risk of GC, which is consistent with previous findings showing that elevated monocyte proportion is associated with poorer prognosis and reduced overall survival [33]. In contrast, lower A/G levels markedly increased the likelihood of GC classification, in agreement with epidemiological evidence linking low A/G levels to increased mortality across multiple malignancies, including GC [34]. In addition, age was ranked among the top 10 most important features. Given that the incidence of GC increases significantly with age [35], we further evaluated the model’s predictive performance across different age strata. The results demonstrated that the CatBoost model was able to consistently identify patients at different disease stages across age groups. Furthermore, based on the top 10 key features and the Streamlit framework, we developed a user-friendly online prediction platform to enhance its clinical accessibility.

This study has the following limitations. First, it was based on populations at 2 hospitals in China, which may have been influenced by factors such as demographic structure and lifestyle habits. Therefore, the generalizability of the results needs further validation. However, the study provides preliminary evidence for diagnosing IGC based on routine laboratory hematological indicators. Second, although the model can distinguish between different stages of IGC progression to some extent, its relatively high specificity and slightly lower sensitivity may lead to missed or incorrect classifications, especially in late-stage patients, which is unacceptable for practical applications. Therefore, the current method should be used only as an auxiliary diagnostic tool and should be combined with routine clinical methods. Additionally, the construction of ML models requires sufficiently large, balanced, and representative clinical datasets. This study exhibits biases in the HC, AG, and GC groups. Although clinical data augmentation using the SMOTE method has been widely applied [25,26], its assumption of a uniform feature space means the synthetic data generated may not fully reflect real clinical heterogeneity or maintain complete clinical interpretability, potentially increasing the risk of overfitting. The relatively small external validation cohorts may also be insufficient to comprehensively assess the model’s generalizability. Therefore, future research should expand the sample size and include data from more collaborative centers to enhance the model’s generalizability. Moreover, the testing platforms differ across hospitals and may affect routine laboratory indicators and thus may influence model transferability. Although the study did not demonstrate the specific effects of platform differences on diagnostic outcomes in the results section, interinstrument and interlaboratory variability remains an important practical barrier to broader application. Therefore, promoting mutual recognition of test results and standardization of clinical reference ranges at regional and national levels is crucial. Finally, this study focuses only on IGC caused by H pylori infection, and its applicability to other types of GC, such as diffuse GC or signet-ring cell carcinoma, has not been clarified. Future studies should expand the scope of research to evaluate the prospects of using routine laboratory indicators in different types of GC.

In conclusion, this study developed and validated an interpretable ML model based on routine clinical laboratory data to predict different stages in the progression of H pylori infection-associated IGC. The model demonstrated promising predictive performance in the internal validation and provided preliminary evidence of transportability in 2 external cohorts. As a potential auxiliary diagnostic tool, it shows promising applicability in remote areas and health care settings with limited medical resources, helping to optimize resource allocation and support early clinical intervention.