Abstract
Introduction
Gastric cancer (GC) remains a major public health concern in Asia. Risk prediction tailored to regional biological features such as Helicobacter pylori (H. pylori) status and high-risk mucosal findings such as atrophic gastritis (AG) and intestinal metaplasia (IM) may help improve the screening workflow.
Methods
Using a large, real-world, nationwide screening cohort with available endoscopic AG/IM codes, we developed 2-year GC risk prediction models that integrate AG/IM with regional demographic and lifestyle factors. We compared a conventional Cox proportional hazards model (CPHM) with the following machine learning (ML) approaches: extreme gradient boosting (XGBoost), decision tree (DT), and logistic regression (LR). Discrimination and calibration were evaluated through internal and external validations. Model interpretability was assessed using Shapley Additive Explanations (SHAP).
Results
The XGBoost model demonstrated the best overall performance, achieving an AUROC of 0.764 (95% CI, 0.755–0.767), a sensitivity of 0.607 (95% CI, 0.560–0.650), and a specificity of 0.746 (95% CI, 0.744–0.750) in the internal validation. In the external validation cohort, XGBoost also showed the highest discrimination with an AUROC of 0.708 (95% CI, 0.682–0.884), a sensitivity of 0.666 (95% CI, 0.470–0.830), and a specificity of 0.597 (95% CI, 0.590–0.600). SHAP analysis consistently identified Helicobacter pylori infection, age, sex, smoking, and atrophic gastritis/intestinal metaplasia (AG/IM) as the major contributors to increased predicted gastric cancer risk.
Discussion
This externally validated and interpretable short-term GC risk model incorporating endoscopically ascertained AG/IM could provide a practical approach for informing risk-adapted screening workflows. The model could help identify individuals at a higher predicted risk for prospective evaluation and closer clinical review. In addition, SHAP clarifies the main contributors to each prediction by highlighting factors most strongly associated with a higher predicted risk.
Keywords: atrophic gastritis, gastric cancer, Helicobacter pylori, intestinal metaplasia, risk prediction, Shapley Additive Explanations
1. Introduction
Gastric cancer (GC) is the fifth leading cause of cancer mortality worldwide (1). Despite the poor prognosis of advanced GC, early detection yields markedly better outcomes, with 5-year survival exceeding 90% for early GC (2, 3). Endoscopy-based screening is widely recognized as the most effective strategy to lower GC mortality through early detection (4–6).
Normal gastric mucosa can progress to cancer through sequential precancerous stages (7), including gastritis, atrophic gastritis (AG), intestinal metaplasia (IM), and dysplasia. Therefore, more proactive management and consideration of shorter endoscopic intervals for high-risk individuals with AG or IM are an important public health priority in Korea, where GC burden is high (8). Most national guidelines recommend screening for high-risk individuals (9, 10). In Korea, since 2002, the National Cancer Screening Program has provided biennial screening endoscopy for all individuals aged ≥40 years.
However, because the National Cancer Screening Program uniformly offers biennial upper endoscopy to all eligible adults, the program’s scale (~7.76 million screening endoscopies in 2021 and 7.84 million in 2022) means that publicly financed expenditures for gastric-cancer screening clearly exceed KRW 250 billion annually (11, 12). There is a growing need to tailor screening intervals according to individual GC risks to reduce unnecessary examinations and concentrate resources on high-risk groups, thereby improving efficiency.
Here, using a large-scale cohort, we aimed to develop a Korean population-specific GC risk-prediction model that is clinically implementable and highly accurate by including not only family history and lifestyle factors but also endoscopically ascertained high-risk mucosal findings, such as AG, IM, and Helicobacter pylori (H. pylori) infection status as risk factors. In contrast to most prior GC risk models that rely on conventional statistical methods, our study applied machine-learning approaches to substantially enhance predictive performance.
2. Materials and methods
2.1. Study design and participants
This study was based on the Korean National Health Information Database (NHID), which integrates nationwide claims and health screening records and covers nearly the entire Korean population through the National Health Insurance (97%) and Medical Aid (3%) (13). Using this database, we identified adults aged 40–74 years who underwent gastroscopy as part of the National Cancer Screening Program (NCSP) in 2018, with a screening participation rate of 61.9% (14). We restricted the analysis to ages 40–74 to reflect the core NCSP screening population (screening offered from ≥40 years) and to reduce heterogeneity from advanced frailty/competing risks at older ages. Individuals with a prior diagnosis of GC or those who died during the 2-year follow-up period were excluded from the study. A total of 6,863,103 individuals met inclusion criteria. Participants were followed up until 2020 to identify patients with GC within 2 years of post-screening. Incident GC was defined according to the International Classification of Diseases, 10th revision (ICD-10) code C16, as a primary or secondary diagnosis in combination with an NHIS special registration code (V193 or V194), indicating eligibility for the expanded co-payment reduction program.
Because the 2-year GC incidence was low, we addressed class imbalance only in model training by applying the random oversampling example (ROSE) procedure to obtain an approximately 1:1 ratio of GC to non-GC; the validation/test sets retained the original outcome prevalence. We selected ROSE over alternative approaches such as synthetic minority over-sampling technique (SMOTE) or class weighting because it generates synthetic samples from a smoothed bootstrap distribution, better preserving the multivariate structure of the minority class and reducing the risk of creating noisy or borderline cases (15, 16). In contrast, class weighting does not modify the empirical feature distribution, which may limit model learning in extremely sparse outcomes (17, 18). Baseline characteristics between GC and non-GC groups were compared using χ² tests, as all variables were categorical.
2.2. Variables and datasets
Nineteen variables identified as risk factors for GC were incorporated into the prediction model, namely, age; sex; body mass index (BMI); health insurance premium; smoking status; alcohol consumption; family history of GC, colorectal cancer, or liver cancer; and personal history of hypertension, diabetes mellitus, myocardial infarction/angina pectoris, stroke, dyslipidemia, colorectal cancer, liver cancer, H. pylori infection, gastric adenoma, and AG or IM. These variables were pre-specified a priori based on guideline statements and epidemiologic evidence for gastric cancer risk in Asian screening contexts, supplemented by lifestyle and clinical-context variables available with high completeness/validity in NHID/NCSP. Specifically, we grouped predictors as: (i) guideline-anchored risk factors—H. pylori eradication history, AG/IM, age, sex, smoking, and family history of GC; (ii) lifestyle—BMI and alcohol; and (iii) contextual clinical variables—socioeconomic status (insurance premium), cardiometabolic comorbidities (hypertension, diabetes, dyslipidemia, myocardial infarction/angina, stroke), and prior cancers (colorectal, liver). Comorbidities and prior cancers were incorporated as non-causal context markers of baseline health and health-care pathways rather than as prespecified causal determinants (Supplementary Table 1).
Age groups were divided into 5-year intervals, and BMI was categorized as normal (<23 kg/m²), overweight (23.0–24.9), obese (25.0–29.9), and severely obese (≥30.0), consistent with Korean/Asia-Pacific criteria (19, 20). Economic status was classified into five groups based on health insurance premiums (medical aid beneficiaries and quartiles 1–4). Smoking status was categorized as nonsmoker, former smoker, or current smoker, and alcohol consumption was classified as nondrinker, drinking three or fewer times per week, or drinking four or more times per week. Alcohol exposure was available only as frequency categories in the 2018 NHID/NCSP screening dataset (nondrinker; ≤3 times/week; ≥4 times/week); per-occasion quantity (standard drinks) and grams of ethanol were not recorded in the analytic extract and thus could not be estimated. Family and medical histories were dichotomized as either present or absent. A history of H. pylori infection was defined based on records of the prescribed eradication regimens. Gastric adenoma was confirmed by gastroscopy and biopsy results from GC screening.
The dataset was divided into a training set and a test set in an 8:2 ratio. The Cox proportional hazard model (CPHM) and three machine learning-based prediction models, namely, logistic regression (LR), decision tree (DT), and extreme Gradient Boosting (XGBoost), were developed. Hyperparameter tuning for the LR and DT models was performed using five-fold cross-validation on the training set, whereas the boosting iteration for XGBoost was set to 500. To mitigate overfitting in the XGBoost model, we applied both L1 (reg_alpha = 1) and L2 (reg_lambda = 1) regularization, in addition to a low learning rate (η = 0.01), restricted tree depth (max_depth = 5), and subsampling (subsample = 0.8). Early stopping was implemented by monitoring the evaluation loss on a validation set (early_stopping_rounds = 50). Although more extensive hyperparameter optimization such as nested cross-validation could further improve model performance, it could not be implemented due to computational constraints. Accordingly, the maximum number of boosting rounds was fixed at 500, and the early-stopping procedure adaptively determined the effective number of iterations based on validation performance. Model performance was assessed using the accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUROC).
2.3. Analyses
We conducted three sensitivity analyses:
(1) multiple imputations of missing covariates using the Multivariate Imputation by Chained Equations (MICE) package (m=5); (2) feature selection using least absolute shrinkage and selection operator (LASSO) regression to remove potentially redundant factors and reduce the risk of data overfitting (21); and (3) missing imputation with feature selection.
External validation was conducted using the Cancer Screenee Cohort Study of the National Cancer Center, a dataset comprising of 31,285 individuals. This cohort was used to evaluate the generalizability and predictive performance of the models using an independent dataset.
To improve the interpretability of the top-performing model (with the highest AUROC value in the external validation set), the Shapley Additive Explanations (SHAP) algorithm was incorporated. This game theory-based method elucidates the contribution of each feature to model predictions, overcoming challenges related to the opacity of machine learning models (22). To support reproducibility and practical implementation, the full R code used to develop and validate the XGBoost model has been included in Supplementary Appendix 1.
All analyses were conducted using R software (version 3.6.4), adhering to the STROBE reporting guidelines.
2.4. Ethical considerations
This study was approved by the Institutional Review Board of the National Cancer Center (Approval number: Ncc2021-0141). The requirement for informed consent was waived because the data analyses were performed retrospectively using anonymized data.
3. Results
A total of 6,863,103 individuals were included in the study, comprising 2,624 with incident GC during the 2-year follow-up and 6,860,479 without incident GC (Figure 1). For model development, the cohort was randomly divided into training (80%) and test (20%) sets, yielding 5,490,482 individuals in the training set (2,099 GC cases and 5,488,383 non-GC cases) and 1,372,621 individuals in the test set (525 GC cases and 1,372,096 non-GC cases). The characteristics of the study participants are summarized in Table 1. Compared to those without incident GC, participants who developed GC were older and more often male (73.3% vs. 46.5%, p < 0.001). The BMI distribution differed, with obesity (≥30 kg/m²) being less common among patients with GC (3.6% vs. 5.3%, p < 0.001). The distribution of the health insurance premium categories did not differ significantly (p = 0.178). Smoking status varied markedly (current smoking 31.6% vs. 17.9%; ex-smoking 26.9% vs. 18.6%; p < 0.001), and drinking frequency also differed (≥4 times/week 6.6% vs. 8.7%, p < 0.001). The family history of GC was more frequent among patients with GC (13.9% vs. 9.7%, p < 0.001). Hypertension and hyperlipidemia were less prevalent among patients with GC (both p < 0.001), and the prevalence of diabetes was similar (p = 0.076). Preexisting myocardial infarction/angina, stroke, colorectal cancer, and liver cancer did not differ significantly between the groups (all p>0.3). In contrast, H. pylori infection, AG/IM, and gastric adenoma were more frequent among patients with GC (all p < 0.001).
Figure 1.
Flow chart of the study participants.
Table 1.
Baseline characteristics of the study population.
| Characteristics | Total (n=6,863,103) |
No gastric cancer (n=6,860,479) |
Gastric cancer (n=2,624) |
P-value |
|---|---|---|---|---|
| Age | <0.001 | |||
| 40–44 | 1,156,886 (16.9) | 1,156,739 (16.9) | 147 (5.6) | |
| 45–49 | 923,474 (13.5) | 923,282 (13.5) | 192 (7.3) | |
| 50–54 | 1,284,123 (18.7) | 1,283,702 (18.7) | 421 (16.0) | |
| 55–59 | 997,762 (14.5) | 997,270 (14.5) | 492 (18.8) | |
| 60–64 | 1,203,969 (17.5) | 1,203,348 (17.5) | 621 (23.7) | |
| 65–69 | 650,132 (9.5) | 649,770 (9.5) | 362 (13.8) | |
| 70–74 | 646,757 (9.4) | 646,368 (9.4) | 389 (14.8) | |
| Sex | <0.001 | |||
| Male | 3,193,616 (46.6) | 3,191,692 (46.5) | 1,924 (73.3) | |
| Female | 3,669,487 (53.4) | 3,668,787 (53.5) | 700 (26.7) | |
| BMI (kg/m2) | <0.001 | |||
| <23 | 2,433,275 (35.5) | 2,432,286 (35.5) | 989 (37.7) | |
| 23–24.9 | 1,739,760 (25.4) | 1,739,087 (25.3) | 673 (25.6) | |
| 25–29.9 | 2,329,619 (33.9) | 2,328,751 (33.9) | 868 (33.1) | |
| ≥30 | 360,449 (5.2) | 360,355 (5.3) | 94 (3.6) | |
| Health insurance premium | 0.178 | |||
| Medical Aid beneficiaries | 248,804 (3.7) | 248,719 (3.6) | 85 (3.2) | |
| ~25 percentile | 1,347,019 (19.6) | 1,346,512 (19.6) | 507 (19.3) | |
| ~50 percentile | 1,322,785 (19.2) | 1,322,251 (19.3) | 534 (20.4) | |
| ~75 percentile | 1,646,985 (24.0) | 1,646,324 (24.0) | 661 (25.2) | |
| ~100 percentile | 2,297,510 (33.5) | 2,296,673 (33.5) | 837 (31.9) | |
| Smoking | <0.001 | |||
| Non-Smoker | 4,357,937 (63.5) | 4,356,847 (63.5) | 1,090 (41.5) | |
| Ex-Smoker | 1,273,595 (18.6) | 1,272,889 (18.6) | 706 (26.9) | |
| Smoker | 1,231,571 (17.9) | 1,230,743 (17.9) | 828 (31.6) | |
| Drinking | <0.001 | |||
| No drinking | 4,703,278 (68.5) | 4,701,446 (68.5) | 1,832 (69.8) | |
| ≤ 3 times/week | 1,564,668 (22.8) | 1,564,049 (22.8) | 619 (23.6) | |
| ≥ 4 times/week | 595,157 (8.7) | 594,984 (8.7) | 173 (6.6) | |
| Family history of gastric cancer | <0.001 | |||
| No | 6,199,918 (90.3) | 6,197,659 (90.3) | 2,259 (86.1) | |
| Yes | 663,185 (9.7) | 662,820 (9.7) | 365 (13.9) | |
| Family history of colorectal cancer | 0.317 | |||
| No | 6,595,230 (96.1) | 6,592,698 (96.1) | 2,532 (96.5) | |
| Yes | 267,873 (3.9) | 267,781 (3.9) | 92 (3.5) | |
| Family history of liver cancer | 0.662 | |||
| No | 6,553,544 (95.5) | 6,551,043 (95.5) | 2,501 (95.3) | |
| Yes | 309,559 (4.5) | 309,436 (4.5) | 123 (4.7) | |
| Hypertension | <0.001 | |||
| No | 6,317,180 (92.1) | 6,314,686 (92.0) | 2,494 (95.0) | |
| Yes | 545,923 (7.9) | 545,793 (8.0) | 130 (5.0) | |
| Diabetes | 0.076 | |||
| No | 6,195,035 (90.3) | 6,192,639 (90.3) | 2,396 (91.3) | |
| Yes | 668,068 (9.7) | 667,840 (9.7) | 228 (8.7) | |
| Hyperlipidemia | <0.001 | |||
| No | 6,231,934 (90.8) | 6,229,485 (90.8) | 2,449 (93.3) | |
| Yes | 631,169 (9.2) | 630,994 (9.2) | 175 (6.7) | |
| Myocardial infarction or angina pectoris | 0.351 | |||
| No | 6,770,177 (98.6) | 6,767,583 (98.6) | 2,594 (98.9) | |
| Yes | 92,926 (1.4) | 92,896 (1.4) | 30 (1.1) | |
| Stroke | 0.448 | |||
| No | 6,794,728 (99.0) | 6,792,134 (99.0) | 2,594 (98.9) | |
| Yes | 68,375 (1.0) | 68,345 (1.0) | 30 (1.1) | |
| Colorectal cancer | 0.353 | |||
| No | 6,839,054 (99.6) | 6,836,442 (99.6) | 2,612 (99.5) | |
| Yes | 24,049 (0.4) | 24,037 (0.4) | 12 (0.5) | |
| Liver cancer | 0.828 | |||
| No | 6,853,711 (99.9) | 6,851,091 (99.9) | 2,620 (99.8) | |
| Yes | 9,392 (0.1) | 9,388 (0.1) | 4 (0.2) | |
| Helicobacter pylori infection | <0.001 | |||
| No | 6,834,934 (99.6) | 6,832,378 (99.6) | 2,556 (97.4) | |
| Yes | 28,169 (0.4) | 28,101 (0.4) | 68 (2.6) | |
| Atrophic gastritis or intestinal metaplasia | <0.001 | |||
| No | 6,536,623 (95.2) | 6,535,105 (95.3) | 1,518 (57.9) | |
| Yes | 326,480 (4.8) | 325,374 (4.7) | 1,106 (42.1) | |
| Gastric adenoma | <0.001 | |||
| No | 6,845,012 (99.7) | 6,842,962 (99.7) | 2,050 (78.1) | |
| Yes | 18,091 (0.3) | 17,517 (0.3) | 574 (21.9) | |
Values are presented as number (%).
BMI, body mass index.
In the external validation cohort (Cancer Screenee Cohort, n = 31,285), demographic and endoscopic patterns among gastric cancer cases were consistent with those observed in the development dataset. Gastric cancer cases were predominantly in their 50s–60s and more frequently male. The prevalences of atrophic gastritis/intestinal metaplasia and gastric adenoma were also markedly higher among cases than non-cases. Detailed characteristics of the cohort are presented in Supplementary Table 2.
The performances of the models are summarized in Table 2. In the internal validation set, XGBoost had the highest AUROC (0.764; 95% confidence interval [CI], 0.755–0.767) (Figure 2) and specificity (0.746; 95% CI, 0.740–0.750). DT achieved the highest sensitivity (0.764, 95% CI, 0.730–0.800), whereas LR had the highest accuracy (0.770, 95% CI, 0.769–0.771). CPHM performed the worst across all metrics (AUROC 0.532, 95% CI, 0.528–0.536). In the external validation set, XGBoost had the highest AUROC (0.708; 95% CI, 0.682–0.884). In the sensitivity analyses, the results were consistent across approaches, with XGBoost retaining the highest AUROC (Supplementary Figure 1).
Table 2.
Performance of CPHM and different machine learning models without missing imputation.
| Accuracy | Sensitivity | Specificity | AUROC | |
|---|---|---|---|---|
| Internal validation set | ||||
| CPHM | 0.534 (0.531–0.537) | 0.531 (0.529–0.533) | 0.511 (0.510–0.512) | 0.532 (0.528–0.536) |
| LR | 0.770 (0.769–0.771) | 0.607 (0.560–0.650) | 0.700 (0.693–0.704) | 0.689 (0.668–0.709) |
| DT | 0.695 (0.674–0.717) | 0.764 (0.730–0.800) | 0.534 (0.530–0.540) | 0.695 (0.674–0.717) |
| XGBoost | 0.746 (0.744–0.750) | 0.607 (0.560–0.650) | 0.746 (0.740–0.750) | 0.764 (0.755–0.767) |
| External validation set | ||||
| CPHM | 0.631 (0.628–0.634) | 0.556 (0.555–0.557) | 0.706 (0.703–0.709) | 0.611 (0.599–0.613) |
| LR | 0.425 (0.420–0.431) | 0.700 (0.669–0.708) | 0.425 (0.420–0.430) | 0.512 (0.423–0.602) |
| DT | 0.655 (0.578–0.736) | 0.800 (0.610–0.920) | 0.483 (0.480–0.490) | 0.655 (0.578–0.736) |
| XGBoost | 0.597 (0.592–0.603) | 0.666 (0.470–0.830) | 0.597 (0.590–0.600) | 0.708 (0.682–0.884) |
Values in parentheses represent 95% confidence intervals. AUROC, area under the receiver operating characteristic curve; CPHM, Cox proportional hazards model; LR, logistic regression; DT, decision tree; XGBoost, extreme Gradient Boosting.
Figure 2.
AUROCs of the prediction model. AUROC, area under the receiver operating characteristic curve; CPHM, cox proportional hazards model; LR, logistic regression; DT, decision tree; XGBoost, eXtreme Gradient Boosting.
Figure 3 presents the SHAP explanations for the final XGBoost model, which is the model showing the best discrimination in the external validation set. Panel (A) shows a beeswarm plot of individual SHAP values (x-axis) with feature values encoded by color; positive values increase and negative values decrease the predicted GC risk. Panel (B) ranks the predictors by global importance using the mean absolute SHAP value. The largest contributions were from H. pylori infection, age, sex, smoking status, and AG/IM, followed by BMI and a family history of GC. Specifically, H. pylori infection, older age, male sex, current smoking, and the AG/IM ratio showed positive SHAP contributions, indicating a higher predicted GC risk. In contrast, the presence of hypertension showed negative SHAP contributions, consistent with a lower predicted risk.
Figure 3.
SHAP summary plots for predictors of incident gastric cancer. (a) Beeswarm plot of SHAP values for 19 predictors in the external validation set. Each dot represents one participant, with SHAP values on the x-axis and feature values encoded by color. Positive SHAP values increase the predicted risk, and negative values decrease the predicted risk. (b) Global importance based on mean absolute SHAP. Longer bars indicate larger average contributions.
4. Discussion
In a large real-world Asian screening cohort, we showed that a 2-year GC risk model explicitly integrating AG/IM with demographic and lifestyle factors may inform risk-adapted endoscopic surveillance. We compared CPHM with machine-learning approaches (i.e., XGBoost, DT, and LR). XGBoost achieved the highest AUROC in both the internal and external validations, whereas CPHM showed the lowest performance. Beyond discrimination, SHAP quantified each predictor’s contribution and rendered the model clinically transparent. A history of H. pylori infection, age, sex, smoking, and AG/IM were the dominant positive contributors. At the individual level, SHAP generated explanation profiles showing how specific features increased or decreased an individual’s predicted risk, which could support targeted counseling and triage in future implementations. By drawing on mucosal pathology aligned with the Correa cascade, this study advances implementable, biology-informed stratification for high-burden Asian settings.
GC has a high disease burden, ranking 5th worldwide for both cancer incidence and mortality (1). Recent global projections highlight the substantial and growing disease burden of GC, with a projected lifetime burden for the 2008–2017 birth cohorts of approximately 15.6 million cases (~76% H. pylori-attributable). With a burden expected to double in the Americas and rise six-fold in sub-Saharan Africa, there would be potential value in pairing endoscopic screening with risk-stratified, personalized strategies anchored to H. pylori and high-risk endoscopic findings (AG, IM) (23, 24). High-performance prediction models can help identify individuals at elevated risk and support tailored surveillance and intervention plans. In our study, the XGBoost model maintained strong discrimination in external validation, underscoring its potential clinical utility.
According to the Correa hypothesis, intestinal-type GC occurs preferentially in the presence of AG and IM (7). Therefore, appropriate treatment and longitudinal management of these precursor lesions are important public health priorities in Korea, where the incidence of GC is high. In addition, the current Korean endoscopy guidelines recommend biennial screening endoscopy for adults aged ≥40 years; however, there is an ongoing debate as to whether patients with IM and/or AG should undergo surveillance at intervals shorter than 2 years. The 2020 AGA Clinical Practice Guidelines (Gastroenterology) estimated the annual risk of progression from IM to GC to be 0.124%, whereas a Korean study reported a higher annual progression risk of 0.187% in the presence of IM (25). Furthermore, a study that followed up patients with IM found progression to adenoma or GC after a median of 11–32 months, with incomplete-type IM tending to progress more rapidly (26). Collectively, these findings suggest that surveillance at intervals shorter than 2 years may benefit patients with incomplete-type IM; however, confirmatory evidence from Korean cohorts is currently lacking.
Based on the first national screening dataset in which AG/IM endoscopic codes were available, our study incorporated AG/IM findings into a 2-year GC risk-prediction model and identified these features as risk-increasing factors. These findings may inform the consideration of a shift from uniform biennial endoscopy to biology-informed, risk-stratified scheduling in high-burden Asian screening programs. The results provide external validation and SHAP-based interpretability for programmatic adoption and patient-facing communication. Although our analysis could not directly evaluate the optimal surveillance interval, future studies are warranted to determine whether endoscopic screening at intervals of less than 2 years in patients with AG/IM reduces GC mortality.
The proposed models could be used to (1) flag high-risk individuals in routine screening programs and (2) inform personalized risk communication and management. SHAP-based explanations may guide tailored interventions by highlighting modifiable contributors (e.g., H. pylori eradication and smoking cessation) and contextualizing AG/IM-driven risk in shared decision-making regarding surveillance timing.
A key strength of this study is the large real-world dataset derived from the National Health Insurance Service program, which includes robust demographics, diagnostic codes, medication use, and endoscopic results implicating high-risk findings such as AG, IM, and gastric adenoma. Notably, AG or IM coding has been available since 2018, enabling one of the first short-term GC prediction analyses to explicitly incorporate these high-risk endoscopic features. Another strength is our use of SHAP to mitigate the “black-box” limitation of machine-learning models by clarifying the direction and magnitude of each variable’s effect.
The primary limitation of this study is the relatively short follow-up period, which restricts inferences about short-term (2-year) GC risk. Longer follow-up and prospective evaluations are needed to assess calibration drift, clinical impact, and whether intensified surveillance for AG/IM reduces GC incidence and mortality. Because survival-based machine-learning models could not be implemented within the NHIS secure analysis environment, we were unable to perform a conceptually consistent comparison between Cox and survival-ML approaches. Therefore, we compared the Cox proportional hazards model with available binary classification algorithms, using Cox as the traditional and most widely used benchmark in clinical and epidemiologic prediction research. These analytical constraints also influenced our handling of censoring, and we excluded individuals who died within two years after screening to minimize outcome misclassification. Although this approach reduced misclassification of early deaths, it may have introduced selection or survivorship bias and informative censoring, potentially attenuating estimated risks.
Another limitation is that we restricted the cohort to individuals aged 40–74 years, which limits generalizability to adults ≥75 years. In addition, our dataset captured alcohol frequency but not per-occasion quantity or ethanol grams, precluding dose–response analyses and potentially attenuating alcohol-related effects. Moreover, our evaluation focused primarily on discrimination metrics such as AUROC, accuracy, sensitivity, and specificity; although this approach is consistent with prior gastric cancer prediction studies (27), the lack of calibration assessment and clinical-utility evaluation using decision curve analysis (DCA) remains a limitation.
Finally, the closed and fixed-version computational environment substantially limited the scope of hyperparameter tuning. This restriction may have constrained achievable discrimination, calibration performance, and overall model stability; broader hyperparameter optimization may yield improved performance in settings with more flexible computational infrastructure.
5. Conclusions
Using a large, national screening cohort including high-risk endoscopic findings, we developed a machine-learning model, XGBoost, with higher discrimination than that of conventional approaches for predicting 2-year GC risk. Beyond identifying high-risk individuals, the SHAP-based interpretability facilitates personalized counseling and targeted interventions. Although our results support the rationale for risk-adapted surveillance, potentially with shorter intervals for patients with AG/IM, prospective studies are warranted to determine the optimal surveillance frequency and its effect on GC outcomes.
Funding Statement
The author(s) declared that financial support was received for this work and/or its publication. This study was supported by a grant from the National Cancer Center of Korea (Grant number: 2111060).
Footnotes
Edited by: Yutong He, Fourth Hospital of Hebei Medical University, China
Reviewed by: Christian Cotsoglou, IRCCS San Gerardo dei Tintori Foundation, Italy
Hongwei Liu, Henan Provincial Cancer Hospital, China
Data availability statement
Data cannot be shared publicly because the data are from the Korean National Health Insurance Service (NHIS) health screening cohort and are subject to access restrictions. Data are available from the NHIS Institutional Data Access/Ethics Committee for qualified researchers who meet the criteria for access to confidential data. Contact: URL: https://nhiss.nhis.or.kr/; Address: 32 Gungang-ro, Wonju-si, Gangwon-do 26464, Republic of Korea. Requests to access the datasets should be directed to https://nhiss.nhis.or.kr.
Ethics statement
The studies involving humans were approved by Institutional Review Board of the National Cancer Center. The studies were conducted in accordance with the local legislation and institutional requirements. The ethics committee/institutional review board waived the requirement of written informed consent for participation from the participants or the participants’ legal guardians/next of kin because The requirement for informed consent was waived because the data analyses were performed retrospectively using anonymized data.
Author contributions
HO: Validation, Writing – original draft. CK: Data curation, Formal analysis, Visualization, Writing – review & editing. JJ: Conceptualization, Methodology, Writing – review & editing. MS: Conceptualization, Methodology, Writing – review & editing. KC: Conceptualization, Methodology, Writing – review & editing. IC: Conceptualization, Methodology, Writing – review & editing. BP: Funding acquisition, Supervision, Writing – review & editing.
Conflict of interest
The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Generative AI statement
The author(s) declared that generative AI was not used in the creation of this manuscript.
Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
Publisher’s note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
Supplementary material
The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fonc.2026.1732072/full#supplementary-material
References
- 1. Bray F, Laversanne M, Sung H, Ferlay J, Siegel RL, Soerjomataram I, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA: A Cancer J For Clin. (2024) 74:229–63. doi: 10.3322/caac.21834. PMID: [DOI] [PubMed] [Google Scholar]
- 2. Correa P. Gastric cancer: overview. Gastroenterol Clinics North America. (2013) 42:211. doi: 10.1016/j.gtc.2013.01.002. PMID: [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3. Tan YK, Fielding JW. Early diagnosis of early gastric cancer. Eur J Gastroenterol Hepatol. (2006) 18:821–9. doi: 10.1097/00042737-200608000-00004. PMID: [DOI] [PubMed] [Google Scholar]
- 4. Jun JK, Choi KS, Lee H-Y, Suh M, Park B, Song SH, et al. Effectiveness of the Korean National Cancer Screening Program in reducing gastric cancer mortality. Gastroenterology. (2017) 152:1319–28:e7. doi: 10.1053/j.gastro.2017.01.029. PMID: [DOI] [PubMed] [Google Scholar]
- 5. Hibino M, Hamashima C, Iwata M, Terasawa T. Radiographic and endoscopic screening to reduce gastric cancer mortality: a systematic review and meta-analysis. Lancet Regional Health–Western Pacific. (2023) 35:1–12. doi: 10.1016/j.lanwpc.2023.100741. PMID: [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6. Shah SC, Wang AY, Wallace MB, Hwang JH. AGA clinical practice update on screening and surveillance in individuals at increased risk for gastric cancer in the United States: expert review. Gastroenterology. (2025) 168:405–16:e1. doi: 10.1053/j.gastro.2024.11.001. PMID: [DOI] [PubMed] [Google Scholar]
- 7. Correa P. Human gastric carcinogenesis: a multistep and multifactorial process—first American Cancer Society award lecture on cancer epidemiology and prevention. Cancer Res. (1992) 52:6735–40. [PubMed] [Google Scholar]
- 8. Kang MJ, Jung K-W, Bang SH, Choi SH, Park EH, Yun EH, et al. Cancer statistics in Korea: incidence, mortality, survival, and prevalence in 2020. Cancer Res Treatment: Off J Korean Cancer Assoc. (2023) 55:385–99. doi: 10.4143/crt.2023.447. PMID: [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9. Shah SC, Wang AY, Wallace MB, Hwang JH. AGA clinical practice update on screening and surveillance in individuals at increased risk for gastric cancer in the United States: expert review. Gastroenterology. (2024) 168(2):405–416.e1. doi: 10.1053/j.gastro.2024.11.001, PMID: [DOI] [PubMed] [Google Scholar]
- 10. Dinis-Ribeiro M, Libânio D, Uchima H, Spaander MC, Bornschein J, Matysiak-Budnik T, et al. Management of epithelial precancerous conditions and early neoplasia of the stomach (MAPS III): European Society of Gastrointestinal Endoscopy (ESGE), European Helicobacter and Microbiota Study Group (EHMSG) and European Society of Pathology (ESP) Guideline update 2025. Endoscopy. (2025) 57:504–54. doi: 10.1055/a-2529-5025. PMID: [DOI] [PubMed] [Google Scholar]
- 11. Kim YI. Performance of the national cancer screening program for gastric cancer in Korea. Korean J Helicobacter Up Gastrointest Res. (2024) 24:231–7. doi: 10.7704/kjhugr.2024.0039. PMID: [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12. Bae S, Lee H, Her EY, Lee K, Kim JS, Ahn J, et al. Cost utility analysis of national cancer screening program for gastric cancer in Korea: A Markov model analysis. J Korean Med Sci. (2025) 40:1–14. doi: 10.3346/jkms.2025.40.e43. PMID: [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13. Kim MK, Han K, Lee S-H. Current trends of big data research using the Korean National Health Information Database. Diabetes Metab J. (2022) 46:552–63. doi: 10.4093/dmj.2022.0193. PMID: [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14. Service KSI . National health screening statistics. Daejeon: Ministry of Data and Statistics. Available online at: https://kosis.kr/statHtml/statHtml.do?orgId=350&tblId=DT_35007_N009&vw_cd=MT_ZTITLE&list_id=350_35007_A002 (Accessed October 1, 2025). [Google Scholar]
- 15. Chawla NV, Bowyer KW, Hall LO, Kegelmeyer WP. SMOTE: synthetic minority over-sampling technique. J Artif Intell Res. (2002) 16:321–57. doi: 10.1613/jair.953 [DOI] [Google Scholar]
- 16. He H, Garcia E. Learning from imbalanced data. IEEE Trans Knowl Data Eng. (2009) 1263–84. doi: 10.1109/TKDE.2008.239, PMID: 41116384 [DOI] [Google Scholar]
- 17. Chen W, Yang K, Yu Z, Shi Y, Chen CP. A survey on imbalanced learning: latest research, applications and future directions. Artif Intell Rev. (2024) 57:137. doi: 10.1007/s10462-024-10759-6. PMID: 41784122 [DOI] [Google Scholar]
- 18. Zhang J, Chen L. Clustering-based undersampling with random over sampling examples and support vector machine for imbalanced classification of breast cancer diagnosis. Comput Assisted Surg. (2019) 24:62–72. doi: 10.1080/24699322.2019.1649074. PMID: [DOI] [PubMed] [Google Scholar]
- 19. Region WWP . The Asia-Pacific perspective: Redefining obesity and its treatment. Sydney: Health Communications Australia; (2000). 17–8. [Google Scholar]
- 20. Shiwaku K, Anuurad E, Enkhmaa B, Kitajima K, Yamane Y. Appropriate BMI for Asian populations. Lancet. (2004) 363:1077. doi: 10.1016/S0140-6736(04)15856-X. PMID: [DOI] [PubMed] [Google Scholar]
- 21. Mahmoudian M, Venäläinen MS, Klén R, Elo LL. Stable iterative variable selection. Bioinformatics. (2021) 37:4810–7. doi: 10.1093/bioinformatics/btab501. PMID: [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22. Messalas A, Kanellopoulos Y, Makris C. (2019). Model-agnostic interpretability with shapley values, in: 2019 10th International Conference on Information, Intelligence, Systems and Applications (IISA). IEEE. 1–7. doi: 10.1109/IISA.2019.8900669, PMID: [DOI] [Google Scholar]
- 23. Gu J, He F, Clifford GM, Li M, Fan Z, Li X, et al. A systematic review and meta-analysis on the relative and attributable risk of Helicobacter pylori infection and cardia and non-cardia gastric cancer. Expert Rev Mol Diagn. (2023) 23:1251–61. doi: 10.1080/14737159.2023.2277377. PMID: [DOI] [PubMed] [Google Scholar]
- 24. Park JY, Georges D, Alberts CJ, Bray F, Clifford G, Baussano I. Global lifetime estimates of expected and preventable gastric cancers across 185 countries. Nat Med. (2025) 31:3020–7. doi: 10.1038/s41591-025-03793-6. PMID: [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25. Gawron AJ, Shah SC, Altayar O, Davitkov P, Morgan D, Turner K, et al. AGA technical review on gastric intestinal metaplasia-natural history and clinical outcomes. Gastroenterol. (2020) 158(3):705–731.e5 doi: 10.1053/j.gastro.2019.12.001, PMID: [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26. Dinis-Ribeiro M, Lopes C, da Costa-Pereira A, Guilherme M, Barbosa J, Lomba-Viana H, et al. A follow up model for patients with atrophic chronic gastritis and intestinal metaplasia. J Clin Pathol. (2004) 57:177–82. doi: 10.1136/jcp.2003.11270. PMID: [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27. Fan Z, He Z, Miao W, Huang R. Critical analysis of risk factors and machine-learning-based gastric cancer risk prediction models: A systematic review. Processes. (2023) 11:2324. doi: 10.3390/pr11082324. PMID: 41725453 [DOI] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
Data cannot be shared publicly because the data are from the Korean National Health Insurance Service (NHIS) health screening cohort and are subject to access restrictions. Data are available from the NHIS Institutional Data Access/Ethics Committee for qualified researchers who meet the criteria for access to confidential data. Contact: URL: https://nhiss.nhis.or.kr/; Address: 32 Gungang-ro, Wonju-si, Gangwon-do 26464, Republic of Korea. Requests to access the datasets should be directed to https://nhiss.nhis.or.kr.