This study utilized comprehensive datasets from three sources. Firstly, data from the Surveillance, Epidemiology, and End Results (SEER) Program (https://seer.cancer.gov/) provided population-based information on cancer incidence, mortality, and clinical outcomes from 17 U.S. registries spanning 2000 to 2019, extracted via SEER*Stat software (version 8.4.4). Secondly, HTSeq gene expression data from The Cancer Genome Atlas Stomach Adenocarcinoma (TCGA-STAD) cohort (https://www.cancer.gov/tcga), including 412 tumor samples and 36 normal controls, were downloaded via the TCGAbiolinks package for molecular analyses. Lastly, clinical data from 171 patients diagnosed and treated at the Affiliated Hospitals of Sun Yat-sen University (SYSU) between 2000 and 2024 served as an external validation cohort. Ethical approval for SEER and SYSU data usage was obtained as required; full ethical details are provided in the Ethics Statement section.

Eligible patients met the following six conditions: (1) aged between 19 and 85 years; (2) confirmed diagnosis of GC; (3) survival months ≥ 1 month; (4) primary tumors located in the cardia, fundus, body, antrum, and pylorus (ICD-O-3 codes 16.0–16.4); (5) histological subtype of adenocarcinoma or signet ring cell carcinoma; and (6) disease-specific death attributed to GC.

Patients were excluded on the basis of the following seven conditions: (1) age ≤ 18 years or > 86 years; (2) diagnosed with gastrointestinal stromal tumors, late-onset distal gastric cancer or neuroendocrine tumors; (3) survival months < 1 month; (4) incomplete pathological information, including TNM stage, tumor differentiation, tumor size, or number of positive lymph nodes; (5) missing demographic data such as marital status, race, or household income; (6) nongastrectomy or unknown treatment modalities; and (7) incomplete follow-up information.

All the statistical analyses and visualizations were conducted via R software (18) (version 4.4.1; R Foundation for Statistical Computing, Vienna, Austria). The following R packages were used: survminer, MatchIt, survival, rms, compareGroups, foreign, dcurves, survivalROC, pROC, DynNom, tidyverse, DESeq2 (19), pheatmap, and clusterProfiler. All tests were two-tailed, and a P value < 0.05 was considered to indicate statistical significance.

Propensity score matching (PSM) was performed via a logistic regression model (20). A 1:1 nearest-neighbor matching method was applied, and matching quality was evaluated via standardized differences. Three patient groups were derived: EONDGC, early-onset DGC (EODGC), and late-onset NDGC (LONDGC). Baseline characteristics were compared before and after PSM. Subgroup analyses were performed within the EONDGC group.

Continuous variables are presented as means ± standard deviations or medians with interquartile ranges, depending on distribution. Categorical variables are summarized as counts and percentages. Group comparisons were performed using the t-test, Wilcoxon rank-sum test, or chi-square test as appropriate (21).

The study was conducted according to the workflow illustrated in Figure 1 . After screening and excluding 124,244 patients from the SEER database, 4,810 eligible cases were included: 535 with EONDGC, 3,312 with LONDGC, and 963 with EODGC.

Following PSM ( Table 1 ), 264 patients remained in both the EONDGC and EODGC groups. Most baseline variables were well balanced, except for primary site (P < 0.001) and type of operation (P < 0.001), which remained significantly different. No significant differences were observed in vital status (14.4% vs. 16.3%, P = 0.618) or survival months (40.5 vs. 51.0, P = 0.629) after matching.

GC cases were identified from the SEER database (2001-2019) under predefined inclusion/exclusion criteria, yielding a primary cohort (n=4,810). The left branch presents high-risk factor analyses: EONDGC (n=535) was propensity-matched 1:1 to LONDGC and to EODGC, followed by survival comparisons and subgroup analyses within EONDGC. The right branch shows model development within EONDGC: SEER cases were split into training (n=376) and internal-validation (n=159) sets, with an external SYSU cohort (n=171) for validation; variables were selected by Cox and LASSO and integrated into a prognostic nomogram. Using TCGA cases meeting the same conditions (n=10), risk scores were calculated to define a high-risk group, from which differentially expressed genes were identified and evaluated for prognostic value. ARSB, a candidate from this screen, underwent in-vitro validation by RT-qPCR, scratch (migration area at 24/48 h), and Transwell migration assays.

In the comparison between the EONDGC and LONDGC groups ( Table 2 ), each group included 506 matched patients. There were significant differences in survival months (41.0 vs. 30.0, P < 0.001) and vital status (57.9% vs. 65.8%, P = 0.001). These findings underscore the prognostic distinctions between early- and late-onset, as well as distal and nondistal gastric cancer. To further explore the heterogeneity within EONDGC, we next performed subgroup analyses based on anatomical subsites.

In the subgroup analyses ( Table 3 ), significant differences remained in the pathological pattern (P = 0.012), pathological grade (P < 0.001), and American Joint Committee on Cancer (AJCC) stage (P = 0.020) after matching. However, the prognostic differences were not statistically significant: the median survival months were 40.0 and 43.0 (P = 0.909), with mortality rates of 59.4% and 51.9% (P = 0.333), respectively.

After confirming that anatomical subsites within the EONDGC did not significantly influence prognosis, we proceeded with model development. A statistical comparison of the baseline variables between the SEER-training set (n = 376) and SEER-validation set (n = 159) was conducted ( Supplementary Table S1 ), confirming that their distributions were not significantly different. Baseline characteristics of the SEER-training cohort and the SYSU-validation cohort were summarized in Supplementary Table S2 .

Univariate and multivariate Cox proportional hazards regression analyses were conducted to identify prognostic factors associated with overall survival (OS) in patients with EONDGC ( Table 4 ).

In the univariate analysis, poor pathological grade (Grade IV: hazard ratio [HR] = 12.93, 95% confidence interval [CI]: 3.42–48.97, P < 0.001), AJCC stage IV (HR = 14.51, 95% CI: 7.75–27.19, P < 0.001), T4 stage (HR = 8.13, 95% CI: 3.95–16.71, P < 0.001), N3 stage (HR = 7.49, 95% CI: 4.52–12.42, P < 0.001), M1 stage (HR = 3.83, 95% CI: 2.71–5.42, P < 0.001), and a high lymph node ratio (LNR; HR = 9.58, 95% CI: 6.44–14.25, P < 0.001) were all significantly associated with worse prognosis. Additionally, larger tumor size (HR = 1.01, 95% CI: 1.01–1.01, P < 0.001) and receipt of chemotherapy (HR = 1.84, 95% CI: 1.26–2.68, P = 0.002) were also linked to increased mortality. Conversely, being of “Other” racial background (HR = 0.62, 95% CI: 0.41–0.93, P = 0.026) and being married (HR = 0.72, 95% CI: 0.55–0.94, P = 0.017) were associated with reduced mortality risk.

In the multivariate analysis, independent predictors of poor survival included T4 stage (HR = 4.64, 95% CI: 2.05–10.52, P < 0.001), N3 stage (HR = 2.42, 95% CI: 1.31–4.48, P = 0.005), M1 stage (HR = 2.56, 95% CI: 1.78–3.69, P < 0.001), and LNR (HR = 4.03, 95% CI: 2.35–6.89, P < 0.001). Meanwhile, chemotherapy was independently associated with improved survival (HR = 0.50, 95% CI: 0.33–0.77, P = 0.002), as were “Other” racial background (HR = 0.59, 95% CI: 0.38–0.92, P = 0.019) and married status (HR = 0.74, 95% CI: 0.56–0.98, P = 0.037).

Given the large number of prognostic factors identified by Cox regression, we applied LASSO regression to reduce variable dimensionality and enhance the model’s predictive efficiency. LASSO regression showed a variable selection range of 2 to 28 predictors ( Figure 2 ). Seven variables were ultimately retained on the basis of their prognostic significance: race, pathological grade, T stage, N stage, M stage, LNR, and chemotherapy.

These variables were integrated into a nomogram to visualize the EONDGC prognostic model ( Figure 3A ). In the nomogram, each variable was assigned a point value ( Supplementary Table S3 ), and the total score was calculated by summing across variables. This total risk score was then mapped to predict 1-, 3-, and 5-year OS probabilities. Furthermore, an interactive web-based dynamic nomogram was developed, allowing users to input specific clinical information and to obtain individualized survival predictions (https://zhangzhq79sysu.shinyapps.io/EONDGC/, Figure 3B ).

The C-index was 0.758 (95% CI: 0.725–0.791) in the SEER-training set, 0.718 (95% CI: 0.663–0.773) in the SEER-validation set, and 0.762 (95% CI: 0.719–0.805) in the SYSU-validation set. In the SEER-training set ( Figure 4A ), the AUCs at 1, 3, and 5 years were 0.802 (95% CI: 0.738–0.866), 0.817 (95% CI: 0.769–0.854), and 0.837 (95% CI: 0.784–0.866), respectively. In the SEER-validation set ( Figure 4B ), the corresponding AUCs were 0.760 (95% CI: 0.632–0.877), 0.768 (95% CI: 0.682–0.834), and 0.756 (95% CI: 0.668–0.821). In the SYSU-validation set ( Figure 4C ), the model achieved AUCs of 0.812 (95% CI: 0.760–0.890) at 1 year, 0.819 (95% CI: 0.757–0.885) at 3 years, and 0.836 (95% CI: 0.761–0.886) at 5 years.

In both the SEER-training ( Figure 4D ) and SEER-validation sets ( Figure 4E ), the curves at all time points closely aligned with the ideal 45-degree reference line. In the SYSU-validation set ( Figure 4F ), the 1-, 3-, and 5-year calibration curves showed moderate deviation at intermediate predicted probabilities but remained close to the ideal line in the high-probability range.

To further evaluate the clinical decision-making value of the EONDGC model, we compared its net benefit with that of the traditional TNM staging system using DCA. As shown in Supplementary Figures S1A-C for the SEER-training set and Supplementary Figures S1D-F for the SEER-validation set, the EONDGC model consistently provided a wider range of clinical net benefit than did the traditional Tumor-Node-Metastasis (TNM) staging system across all time points. In the SYSU-validation set ( Supplementary Figures S1G-I ), the model exhibited higher net benefit than the TNM staging system at 1-, 3-, and 5-year time points across a wide range of threshold probabilities.

Significant differences in OS were observed between the high- and low-risk groups in the SEER-training ( Figure 5A ), SEER-validation ( Figure 5B ), and SYSU-validation set ( Figure 5C ; P < 0.0001), indicating consistent and reproducible stratification performance of the model across internal and external datasets.

A total of 10 EONDGC patients from the TCGA cohort were included. Using the EONDGC model, we calculated risk scores for each patient and identified those in the high-risk group for further molecular characterization. DEGs analysis in high-risk patients was visualized using a volcano plot ( Figure 6A ), identifying 73 significantly upregulated genes. These genes were subsequently subjected to KEGG pathway enrichment analysis ( Figure 6B ), and their expression profiles are displayed in a heatmap ( Figure 6C ). A global comparison of DEGs analysis between early-onset and late-onset NDGC was presented in Supplementary Figures S2A, B . Based on the upregulated genes ( Supplementary Table S4 ), univariate and multivariate Cox regression analyses were performed in the TCGA-NDGC cohort to evaluate their prognostic significance.

In the univariate analysis, genes such as JCAD (HR = 2.02, 95% CI: 1.21–3.37, P = 0.007), WWC3 (HR = 1.88, 95% CI: 1.14–3.12, P = 0.014), TNC (HR = 1.7, 95% CI: 1.04–2.79, P = 0.034), and ARSB (HR = 2.23, 95% CI: 1.33–3.72, P = 0.002) were associated with a greater risk of death. In the multivariate analysis, ARSB (HR = 2.40, 95% CI: 1.38–4.17, P = 0.002) and PDCD1 (HR = 0.50, 95% CI: 0.27–0.91, P = 0.024) remained statistically significant. To further validate their prognostic stratification potential and explore their combined impact, we conducted Kaplan-Meier survival analyses based on ARSB, PDCD1, and their expression ratio.

As shown in Figure 6D , patients with high ARSB expression had significantly worse OS than those with low expression (18.7 vs. 70.0 months, P < 0.05). In Figure 6E , higher PDCD1 expression was associated with a better prognosis, although the difference did not reach statistical significance (21.9 vs. 25.9 months, P = 0.064). Figure 6F showed that patients with an ARSB/PDCD1 ratio (APR) >0.86 had significantly worse survival than those with an APR ≤ 0.86 (25.9 vs. 36.5 months, P = 0.041). In addition, APR was inversely correlated with the IPS ( Supplementary Figure S2C ). ROC analysis further showed that APR discriminated IPS status with an AUC of 0.729 ( Supplementary Figure S2D ).

To enable gain-of-function testing, an ARSB overexpression plasmid was constructed ( Figure 7A ) and transiently introduced into MGC-803 and AGS cells. RT-qPCR confirmed robust ARSB mRNA upregulation relative to empty-vector controls ( Figures 7B, C ). In wound-healing assays, representative images at 0/24/48 h are shown ( Figures 7D, F ), and quantification demonstrated significantly greater wound closure at 24 and 48 h in ARSB-overexpressing cells versus controls in both lines ( Figures 7E, G ). In Transwell migration assays using uncoated inserts, representative micrographs revealed more migrated cells upon ARSB overexpression ( Figure 7H ), with corresponding counts significantly increased for both MGC-803 and AGS ( Figures 7I, J ). Collectively, these results indicate that forced ARSB expression enhances the migratory capacity of GC cells in vitro.

This study analyzed EONDGC using the SEER database, comparing it with LONDGC and EODGC, alongside subgroup analyses within EONDGC. We found that tumor location did not significantly affect survival in early-onset cases, whereas age was a key prognostic factor, suggesting location influences clinical presentation more than survival. By integrating clinical and molecular factors, this study pioneers a prognostic model for EONDGC, introducing the novel APR index to enhance risk stratification.

Unlike most previous models for young or early-onset GC, which were primarily based on TNM staging, tumor size, or location (26, 27), our EONDGC model incorporates seven clinically relevant variables, including LNR, race, and chemotherapy. LNR has been shown to outperform conventional N staging, with reported C-index improvements from 0.665 to 0.773 when included (28). Although a previous model included external validation, our study was specifically designed as a prognosis-oriented comparative analysis, ensuring clear clinical relevance and methodological rigor (26). This prognosis-driven approach, together with the inclusion of LNR, treatment factors, and multi-cohort validation, enhances the model’s methodological robustness and clinical applicability.

To explore molecular mechanisms, we identified ARSB and PDCD1 as independent prognostic biomarkers. PDCD1, a key immune checkpoint receptor, has been reported to correlate with favorable outcomes in immune-active GC, particularly when expressed on CD8⁺ T cells. However, such studies primarily reflect immune status without addressing tumor-intrinsic biology (29–31). In contrast, ARSB is a tumor-derived enzyme linked to Wnt/β-catenin signaling, upregulated in EONDGC and associated with poor prognosis (32, 33). To integrate these opposing effects, we proposed the APR, which effectively stratified survival risk. Unlike existing signatures that focus solely on immune or metabolic markers, APR provides a concise, biologically interpretable index reflecting tumor–immune interaction. To complement these clinical findings, we performed in-vitro gain-of-function assays in AGS and MGC-803 cells: transient ARSB overexpression accelerated wound closure at 24/48 h and increased Transwell migration counts, indicating that ARSB enhances gastric cancer cell motility and providing experimental support for its association with poorer survival in EONDGC.

To the best of our knowledge, this study is the first to specifically investigate EONDGC, an understudied GC subtype with distinct clinical features. By separating this population, we identified prognostic patterns that enabled the development of a site-specific model outperforming conventional TNM staging. Furthermore, we propose the APR as a novel molecular biomarker. As the first index combining ARSB and PDCD1 expression, APR shows promise for individualized treatment, particularly in immunotherapy, due to its relevance to the PD-1/PD-L1 axis (30, 34). The concordant increase in migration readouts with ARSB overexpression suggests that APR may partly capture motility-associated risk, supporting its potential utility for refined surveillance and treatment planning.

Despite these advances, several limitations should be acknowledged. First, as a retrospective study, our analysis is inherently subject to selection bias and potential misclassification resulting from incomplete clinical information—such as insufficient lymph node harvest—although PSM was applied to reduce observable confounding and approximate the conditions of a randomized controlled trial. Second, due to the limited number of EONDGC cases in TCGA (n = 10), NDGC cases (n = 176) were used for validation. This small sample size resulted from strict inclusion criteria and the requirement for complete clinical and transcriptomic data. While this substitution was necessary, we demonstrated comparable ARSB/PDCD1 expression patterns between EONDGC and NDGC, supporting the feasibility of this approach. Nevertheless, potential biological heterogeneity cannot be ruled out (35, 36). Third, limited follow-up in the SYSU cohort (n = 171) may affect long-term metrics, yet 1-, 3-, and 5-year predictions remained robust. To address these limitations, future studies should incorporate broader datasets—such as Gene Expression Omnibus and institutional cohorts (e.g., SYSU)–and adopt prospective, multi-center designs to validate the APR and explore its immunotherapeutic potential. Our wet-lab validation was intentionally minimal—transient overexpression in two cell lines with qPCR confirmation only, scratch and Transwell migration assays, and no reciprocal knockdown/rescue or protein-level assays—so future work should incorporate loss-of-function/rescue experiments, protein validation, and pathway readouts to strengthen causal inference.

In summary, this study provides a comprehensive prognostic evaluation of EONDGC by integrating large-scale clinical data with molecular analysis. These findings contribute to the growing understanding of age-specific tumor biology and underscore the need for prospective, multicenter studies to further validate the APR index and explore its implications in personalized therapy and immuno-oncology.