Clinical and pathological data of patients with LOGA who underwent radical gastrectomy with lymphadenectomy were collected from four sources: (1) the SEER database, (2) the TCGA database, (3) The First Affiliated Hospital of Anhui Medical University, (4) The Fourth Affiliated Hospital of Anhui Medical University.

Patients were excluded based on the following criteria: (1) age <50 or >80 years at diagnosis; (2) incomplete clinical or pathological data; (3) presence of multiple primary tumors; (4) missing lymph node dissection data. A total of 10,361 patients from SEER, 135 from TCGA, 99 from The First Affiliated Hospital, and 153 from The Fourth Affiliated Hospital were included in the present study ( Figure 1 ). Patients from the SEER database were randomly divided into a training set(n=7252, 70%) and a testing set(n=3109, 30%). The training set was used to develop the prognostic prediction models. Data from the testing set and the two hospitals were used as independent external validation groups for examination of the prognostic prediction model.

Data from SEER and TCGA were exempt from ethical approval. This retrospective study was conducted using previously recorded clinical data, without involving any personally identifiable information. The study protocol was reviewed and approved by the Ethics Committee of The First Affiliated Hospital of Anhui Medical University (J 2025-03-72) and conducted in accordance with the Declaration of Helsinki. Given the non-interventional and anonymized nature of the study, the requirement for informed consent was waived by the Ethics Committee.

The following clinicopathological variables were included in this study: age, gender, primary site, grade, T-stage, N-stage, M-stage, TNM-stage, radiation, chemotherapy, size, ELN, PLN, LNR, LODDS, overall survival (OS), cause-specific survival (CSS), survival months. LNR was defined as the ratio of PLN to ELN. LODDS was calculated using the formula: log[(PLN + 0.5)/(ELN - PLN + 0.5)]. Overall survival (OS) and cause-specific survival (CSS) were the primary and secondary endpoints, respectively, and were derived from the SEER variables “COD to site record” and “SEER cause-specific death classification.”

Age and tumor size were treated as continuous variables. The primary site was classified into nine subgroups: cardia, fundus of stomach, body of stomach, gastric antrum, pylorus, lesser curvature of stomach, greater curvature of stomach, overlapping lesion of stomach, stomach(NOS). Grade was divided into four subgroups: well, moderate, poor, undifferentiated/anaplastic. Based on the 8th edition of the AJCC staging system, T stage was divided into five categories: T1, T2, T3, T4a and T4b. N stage was divided into five categories: N0, N1, N2, N3a and N3b. M stage was divided into two categories: M0 and M1. TNM stage was divided into eight categories: IA, IB, IIA, IIB, IIIA, IIIB, IIIC, IV. Radiation and chemotherapy were recorded as None/Unknown or YES.

The optimal cut-off values for LODDS, LNR, and PLN were determined using X-tile software (version 3.6.1). Based on these thresholds, patients were divided into three groups for each metric:

A new staging system was developed by integrating LODDS with T and M stages, following the structure of the AJCC 8th edition. This model initially included 18 subgroups, which were subsequently combined into 9 final stages ( Figure 2 ):

Similar grouping methods were applied to PLN and LNR-based models.

Cox regression analysis was performed using the training set. Variables with statistical significance in univariate analysis were included in multivariate analysis. Four prognostic models were constructed based on TNM, PLN, LNR, and LODDS classifications. Model performance was compared using the Akaike Information Criterion (AIC), Concordance Index (C-index), and Area Under the Curve (AUC). AIC is a statistical measure that aims to optimize the balance between model fit and complexity, penalizing excessive parameters to prevent overfitting. The C-index evaluates a model’s ability to discriminate risk rankings, particularly in survival analysis, where it measures concordance between predicted and observed event sequences. The AUC is a metric that assesses the performance of a binary classification model by quantifying the separation between classes across all possible decision thresholds. It is notable for its resilience to class imbalance. Collectively, these metrics comprehensively address model parsimony, predictive consistency, and discriminative power across a range of applications.

Based on the training set, nomograms were developed to predict OS and CSS by integrating the most predictive variables. The evaluation of the applicability of the nomograms was performed using the C-index, time-dependent AUC, and calibration curves in the training and testing sets. Decision Curve Analysis (DCA) was performed to assess clinical utility by calculating net benefit across different threshold probabilities.

Patients were categorized into high- and low-risk groups according to the calculated survival risk scores derived from the nomogram. Survival differences between the two groups were analyzed using Kaplan-Meier curves and the log-rank test.

Categorical variables were reported as frequencies (n) and percentages (%). Continuous variables with normal distribution were expressed as mean ± standard deviation (SD) and compared using independent-sample t-tests. Non-normally distributed variables were presented as median and interquartile range (IQR), and compared using the Mann-Whitney U test. All statistical tests were two-sided, and a p value <0.05 was considered statistically significant. Given that all the analyses were based on predefined hypotheses and clinically relevant variables rather than exploratory multiple testing, no formal adjustment for multiple comparisons was applied. All the statistical analyses in this study were conducted using R software (version 4.4.2) and GraphPad Prism (version 10.2.3).

A total of 10,361 patients diagnosed with LOGA extracted from the SEER database were randomly divided into training and testing sets. As presented in Table 1 , no statistically significant differences in clinicopathological parameters were observed between the two groups (p>0.05).

Firstly, 18 clinicopathologic variables were included in a univariate COX regression analysis, 16 of which were significantly associated with OS and 17 with CSS ( Table 2 ). Subsequently, to address potential multicollinearity, certain variables were consolidated. In the analysis of OS, age, gender, grade, size and chemotherapy were included in the multivariate COX regression analysis in combination with one of TNM-stage, LNR-stage, PLN-stage and LODDS-stage, respectively. In the analysis of CSS, age, gender, grade, size, radiation and chemotherapy were included in the multifactorial COX regression analysis in combination with one of TNM-stage, LNR-stage, PLN-stage and LODDS-stage, respectively. The results showed that all included variables were identified as independent prognostic factors for OS ( Table 3 ), and all but radiation were independent factors for CSS ( Table 4 ). In conclusion, it’s consistent on both OS and CSS that age, gender, grade, size and chemotherapy in combination with one of TNM-stage, LNR-stage, PLN-stage and LODDS-stage were independent influence factors.

AIC, C-index, and AUC were used to compare the performance of the four staging-based prognostic prediction models in the training group. In the analysis of OS, LODDS-stage model demonstrated the lowest AIC value, the highest C-index and time-dependent AUC values at 1, 3, 5, and 10 years in comparison with the other three models. In the analysis of CSS, LODDS-stage model outperformed the other three staging systems in terms of C-index and time-dependent AUC across all evaluated time points ( Table 5 ), indicating superior overall predictive performance.

In conjunction with the above studies, the prognostic prediction model integrating LODDS-stage was found to be more comprehensive and accurate. Therefore, this model was utilized to construct the nomograms for predicting 1-, 3-, 5-, and 10-year OS ( Figure 3A ) and CSS ( Figure 4A ) in LOGA patients. Subsequently internal validation based on the training set and external validation based on the testing set, the TCGA dataset, and the hospital dataset, the calibration curves demonstrated excellent agreement between predicted and observed survival probabilities ( Figures 3B-D ; Figures 4B-D ).

Based on the time-dependent AUC curves, the LODDS-stage model outperformed all individual variables ( Figure 5 ) and other staging systems ( Figure 6 ) for both OS and CSS in both the training and testing sets. The predictive efficacy of the LODDS staging-based predictive model for OS ( Figures 7A, B ) and CSS ( Figures 7C, D ) at different time points (1, 3, 5, and 10 years) was evaluated using DCA in the training and testing sets. The results showed that the LODDS-stage model was consistently above the All and None curves at most risk thresholds, suggesting its high clinical applicability. In conclusion, the LODDS-stage model shows good predictive benefits at multiple time points, which is expected to provide a reference basis for clinical individualized treatment decision-making.

The TCGA dataset and the combined dataset of the two hospitals were used as two external validation cohorts for the constructed predictive models. Both the time-dependent AUC ( Figures 8A, B ) and DCA curves ( Figures 8C-F ) were utilized to demonstrated the accuracy and clinical utility of constructing prediction models based on LODDS-stage.

To further validate the utility of the prognostic prediction model, patients were equally stratified into two groups based on the median model score: high- and low-risk. Survival analyses were subsequently performed using Kaplan-Meier survival curves for both groups of patients in all datasets(training ( Figures 9A, E ) and testing sets ( Figures 9B, F ) from the SEER database, external validation sets from the TCGA database ( Figures 9C, G ) and dataset consisting of data from two hospitals ( Figures 9D, H )). The analysis revealed a significantly poorer survival outcomes in the high-risk group compared to the low-risk group, with a statistically significant difference (p < 0.05). Throughout the follow-up period, individuals in the low-risk group consistently exhibited better survival outcomes. These findings suggest that the LODDS model demonstrates the effective discriminatory ability and holds promise as a valuable tool for predicting long-term survival outcomes.