Prognostic value of a modified pathological staging system for gastric cancer based on the number of retrieved lymph nodes and metastatic lymph node ratio

Study population and data collection

Clinical data from the US Surveillance, Epidemiology, and End Results (SEER) Program from 2010–2015 (https://seer.cancer.gov/) was extracted and analyzed as training set, data from 2016–2017 was adopted as internal validation set. Data from The Cancer Genome Atlas Program (TCGA) (https://portal.gdc.cancer.gov/) and prognosis data from the gastrointestinal surgery department, Third Affiliated Hospital of Sun Yat-sen University were applied as external validation sets.

Screening criteria for gastric cancer cases were as follow: exclusion of cases with only autopsy or death certificate, cases where initial tumor location was not stomach, patients with stage 0 and stage IV, cases without radical surgery, non-adenocarcinoma cases, death cases within 1 month after operation, and cases with unknown lymph node information and AJCC TNM stage.

The study analyzed various factors such as age of diagnosis (<50 years, 50–69 years, >69 years), gender, race (white, black, other), AJCC T stage (T1–T4b), AJCC TNM stage (I–III), primary tumor location (stomach body, antrum/pylorus, cardia/fundus, greater gastric recurve, lesser gastric recurve, overlapping area, NOS), clinical features such as tumor size (≥5 cm, <5 cm, unknown), tumor grade (I–IV), chemotherapy, radiotherapy, number of lymph nodes retrieved and number of metastases, and lymph node positive rate. The populations of American Indian/Alaskan and Asian/Pacific Islander were classified as “other” due to small sample sizes. Tumor grade was also analyzed, with grades I–IV representing highly differentiated, moderately differentiated, poorly differentiated, and signed-ring cell carcinoma, respectively. Overall survival (OS) is the time from cancer diagnosis to death from any cause, while disease-specific survival (DSS) is the time from cancer diagnosis to death specifically due to the disease.

Statistical analysis

The chi-square test was used to compare differences between categorical variables, while the t-test was used for continuous variables. Univariate and multivariate Cox regression were used to examine the association between prognosis and the covariates. The Kaplan-Meier method is used to compare overall OS and DSS among groups. Hazard ratios (HR) for mortality were reported following adjusting for covariates including age, year of diagnosis, gender, race, tumor features (site, size, and grade).

The study used Mantel-Cox survival test to calculate chi-square values for 40 segments of LNR with a 0.025 interval, identified four peaks as cutoff points, and defined LNR-based N (rN) classification based on these cutoff points. Bootstraps with 1,000 resamples was conducted to internally validate the rN staging system. Patients included in the study were then redistributed according to the revised AJCC (rAJCC) staging system. We compared the performance of the rAJCC system with that of the 8th AJCC staging system in terms of discriminatory ability and the prognostic homogeneity. These comparisons were assessed using the area under the receiver operating characteristic curve (AUC), Akaike’s information criterion (AIC), and Bayesian Information Criterion (BIC). A lower AIC value or a higher AUC indicates a stronger discriminatory capacity of the staging system.

The proper threshold of ELN count analysis was conducted in two steps. COX regression firstly performed based on OS and DSS to examine the HR of each ELN count in patients with negative LNs. Patients with only one ELN were used as a reference, HR value for each ELN count group for both OS and DSS was analyzed using locally weighted scatterplot smoothing (LOWESS) scatter curve fitting. The minimum count of ELNs was determined by Chow test at which the slope of the curve changes significantly. Logistic regression analysis was then performed based on patients’ data with negative or only one LN metastasis, using node status as the outcome variable. Patients with only one ELN was defined as reference. LOWESS and Chow test help draw the odds ratio curve of each ELN count group and defined the ideal cutoff LN count for maximum detection efficacy of at least one involved LN.

The minimum ELN count and the rAJCC staging system merged to develop the integrated GC staging system. We compared the performance of the integrated GC staging system to that of the 8th AJCC staging system in terms of discriminatory ability and prognostic homogeneity. These aspects were evaluated using metrics such as AUC, AIC and BIC.

A significance level of p < 0.05 was used for all data analysis. Statistical analyses were conducted using IBM SPSS Statistics for Windows v. 20.0 (IBM Corp., Armonk, NY, USA) and R version 3.6.2 software (Bell Laboratories, Murray Hill, NJ, USA).

The current study is performed following the principles from the declaration of Helsinki. Patient from our center was informed before database enrollment individually, data from the TCGA database were publicly available and de-identified. The study was approval by the institutional review board of the Third Affiliated Hospital of Sun Yat-sen University for human data analysis as No. II2023-062-01.

Patient characteristics

The study analyzed data from 8,137 patients in the SEER database, 277 patients in the TCGA database, and 719 patients from the authors’ center. The SEER patients were divided into training (2010–2015) and validation (2016–2017) sets based on years of diagnosis. The selection process is shown in Fig. 1. The datasets utilized in our analysis were designated as follows: the training set as dataset 1, the validation set as dataset 2, the TCGA set as dataset 3, and the data from the author’s institution as dataset 4.

The majority of patients in the training set, validation set, TCGA set, and data from author’s center were over 50 years old (90.35%, 90.15%, 92.42% and 79.28%) and had moderate to poor differentiation (85.72%, 84.15%, 97.12% and 90.41%). In the datasets derived from the SEER database, namely dataset 1 and dataset 2, the majority of patients identified as White, comprising 65.67% and 62.88%, respectively. This demographic distribution was mirrored in the TCGA dataset 3, where the White ethnicity constituted 68.23% of the patient population. In contrast, all patients from our single-center study were of Chinese ethnicity, with no representation from other ethnic groups.

Regarding gastric tumor size, approximately half of the patients in the SEER and TCGA datasets had tumors measuring less than 5 cm, with percentages of 56.54%, 56.66%, and 48.38%, respectively. In our single-center study, this proportion was higher, at 63.00%. The primary tumor sites included the body, antrum/pylorus, cardia/fundus, greater curvature, lesser curvature, overlapping regions, and stomach without detailed specification. Notably, the cardia/fundus was the most common primary tumor site in the SEER and TCGA datasets, whereas in our single-center study, the antrum/pylorus was the predominant site. Most patients were diagnosed with T3 or T4 gastric cancer, accounting for 59.49%, 61.09%, 69.32%, and 67.32% of cases. Furthermore, patients found to have node-positive disease according to the 8th AJCC staging criteria, account for 56.53%, 57.39%, 69.31%, and 60.08% in each group. The median ELN counts were higher in the data from our center (35 [IQR26-45]) than in the other sets (16 [IQR10-24], 18 [IQR12-27], 17[IQR10-31]), while the PLN counts were similar across all sets (1 [IQR 0, 4], 0 [IQR 0, 3], 2 [IQR 0, 7] and 2 [IQR 0, 7]). In terms of treatment modalities, across dataset 1, dataset 2, and dataset 3, 56.95%, 61.42%, 35.02% and 59.25% of patients, respectively, received postoperative chemotherapy. Radiotherapy was administered to 37.77%, 33.26%, and 10.83% of patients in the respective datasets, none underwent radiotherapy in our single center. The median follow-up time was longest in the training set (50 months, [19, 77]) and data from author’s center (45 months, [20, 73]). Detailed clinicopathologic characteristics can be found in Table S1.

Univariate and multivariate analysis

Prognosis data and related variables from the training set were analyzed. As shown in Table 1, univariate analysis revealed that male patients (1.13 [1.05–1.21], p = 0.001), age over 69 years old (1.54, [1.36–1.75], p < 0.001), non-white/black patients (p < 0.001), degree of differentiation (p < 0.001), tumor size ≥5 cm (1.68, [1.57–1.80], p < 0.001), tumor locate at cardia/fundus (1.26, [1.11–1.44], p < 0.001), depth of tumor invasion (p < 0.001), N stage (p < 0.001), AJCC pathological classification (p < 0.001), chemotherapy (1.28, [1.20–1.37], p < 0.001), radiotherapy (1.28, [1.19–1.37], p < 0.001), ELN count (0.99, [0.99–1.00], p < 0.001), PLN count (1.06, [1.06–1.06], p < 0.001), and LNR (7.84, [7.05–8.72], p < 0.001) were significantly associated with overall survival of patients diagnosed with gastric cancer patients in dataset 1.

The following factors were independently correlated with poorer OS using multivariate analysis, male patients (HR = 1.14, p = 0.001); patients between 50 to 69 years old (HR = 1.18, p = 0.009) and over 69 years old (HR = 1.88, p < 0.001); non-white/black patients (HR = 0.78, p < 0.001); tumor diameter ≥5 cm (HR = 1.10, p = 0.014); tumors locate at cardia/fundus area (HR = 1.39, p < 0.001); with advancing of pT and pTNM classification (p < 0.05). Furthermore, chemotherapy could help improve the overall prognosis (HR = 0.76, p < 0.001). Higher ELN count was significantly associated with better prognosis of GC (HR = 0.98, p < 0.001), while increasing PLN count and LNR were adverse prognostic factors (HR = 1.03, p < 0.001 and HR = 2.51, p < 0.001). The detailed data was shown in Table 1.

LNR classification and revised GC staging evaluation

LNR was then divided into 40 segments using a 0.025 interval, and chi-square values were calculated through Mantel-Cox survival test between adjacent segments in patients from the training set. Four peaks of the chi-square value were identified as cut-off points at 0.025, 0.175, 0.45, and 0.6 (p < 0.05), as shown in Table S2. The rN classification was used to distinguish five groups: rN0 (0 ≤ LNR < 0.025), rN1 (0.025 ≤ LNR < 0.175), rN2 (0.175 ≤ LNR < 0.45), rN3a (0.45 ≤ LNR < 0.6), and rN3b (0.6 ≤ LNR ≤ 1) based on LNR values. An 8th AJCC N classification was replaced with the corresponding rN classification to develop a rAJCC staging system with rI, rII, and rIII stages.

According to the 8th AJCC, N classification was divided into N0, N1, N2, N3a, and N3b with percentages of 2,761 (43.47%), 1,558 (24.53%), 1,044 (16.44%), 717 (11.29%) and 271(4.27%), respectively. In the revised classification system, the rN categories were distributed as 3,192 (50.26%) for N0, 1,204 (18.96%) for N1, 1,023 (16.11%) for N2, 328 (5.16%) for N3a, and 604 (9.51%) for N3b.

Patients were initially classified into AJCC stage categories of I, II, and III with percentages of 28.99%, 28.61%, and 42.40% respectively. The patients were then regrouped using rAJCC staging system resulting in 32.96% classified as rI, 35.18% as rII, and 31.87% as rIII. atients from SEER validation set, TCGA validation set, and our single center were regrouped subsequently according to the modified N classification.

The revised LNR-based rAJCC system, as illustrated in Fig. 2, it has validated the clinical relevance and effectiveness. The rAJCC system effectively stratifies patient prognoses across all evaluated datasets based on the LNR-based N classification (panels a, c, e, g), exhibiting statistically significant differences (p < 0.001). When patients were regrouped according to the rAJCC staging system, survival plots (panels b, d, f, h) remained distinct and showed significant statistical differences (p < 0.001). The detailed OS data and comparison results were shown in Table S3.

When evaluating the performance of two models, a lower AIC value suggests a model that fits the data well while balancing model complexity, whereas a lower BIC value implies a more stringent control over complexity, thus reducing the risk of overfitting. In our analysis, the rAJCC staging system demonstrated superior statistical performance over the 8th AJCC staging system, with both AIC (57,042.6 vs. 57,278.3) and BIC (57,054.9 vs. 57,290.7) values being lower in the primary training dataset. In the current study, we chose to depict the AUC across all time intervals using box plots, rather than focusing solely on the 3-year and 5-year AUCs. This methodological choice provides a more nuanced view of the rAJCC staging system’s predictive capabilities over time, as opposed to the 8th edition AJCC staging system for gastric cancer, across multiple datasets. The rAJCC staging system showed a significantly higher AUC (73%, CI [71.7–74.2%]) compared to the 8th AJCC staging system (71.7%, CI [70.4–72.9%]) in primary training dataset 1 indicating greater discriminatory power, as illustrated in Fig. 2 and Table S4.

Further validation of the model using validation set from SEER (2016–2017), data from The Cancer Genome Atlas (TCGA), and a single-center dataset consistently demonstrated the rAJCC staging system’s enhanced discriminatory power. This finding is corroborated by the visual representation in Fig. 3 and the numerical data presented in Table 2.

Ideal number of ELNs analysis based on LN involvement status and survival

The ELN count was assumed to be similar between radical total and distal gastric resection due to insufficient data on the surgical procedures, despite the theoretical difference in LN numbers harvested from the different lymphadenectomy ranges. The optimal threshold for ELN count analysis was established through a two-step approach. COX regression was initially applied, based on OS and DSS, to assess the HR for each ELN count in patients with negative LNs, using patients with a single ELN as the reference. The HR values for each ELN count group were analyzed with LOWESS scatter curve fitting. The Chow test identified the minimum ELN count where the curve’s slope indicated a significant change. Logistic regression was then conducted on data from patients with negative or single LN metastasis, defining node status as the outcome variable and patients with a single ELN as the reference. The LOWESS and Chow test were utilized to plot the odds ratio curve for each ELN count group, identifying the ideal cutoff LN count for maximizing the detection efficacy of at least one involved LN.

Patients with negative LN was analyzed first, multivariate analysis have proved a greater number of LN was examined, the risk of potentially positive LN decreases thus improved the prognosis. HR of each ELN count compared with only one ELN (as reference) was analyzed using Cox proportional hazards regression model to determine the effect of ELN number on OS/DSS, after adjusting for other significant prognostic factors. LOWESS analysis was utilized to visualize the survival curves. Structural break ELN count was determined by Chow test as significant curvature change (p < 0.05) in the HR curve, which was minimum ELN count to obtain survival benefit, as shown in Figs. 4A, 4C. In the current study, curve of OS revealed 30 ELNs at least to guarantee survival benefit while which was 30 in DSS curve in Figs. 4B, 4D.

With regard to the correlation between ELN count and positive LNs identified, a greater ELN count was associated with a greater number of PLNs. Patients from SEER (2010–2015) cohort was then redistributed into node-negative and only one positive LN group. LN status was then assessed by correlating the ELN number and the proportion of each node stage category (node negative vs. one node positive) by using a binary logistic regression model after adjusting for other potential confounders. The curves of odds ratios (ORs; LN involvement) were fitted by using a LOWESS smoother. The minimum ELN count in detecting at least one involved LN was clarified using Chow test (p < 0.05) as 29 lymph nodes. Both approaches indicated a minimum ELN of 30 was necessary. All curved and Chow test results were shown in Fig. 4.

Integration and validation of both ideal ELN count and LNR GC staging system

Patients with an ELN count of 30 or more were identified from each dataset, were categorized as Internal validation set-1, Internal validation set-2, External validation set-1, and External validation set-2, and subsequently reassessed using the rAJCC staging system. All validation datasets, both internal and external, showed a significantly different prognosis (p < 0.001) and improved HR performance compared to the 8th edition AJCC system in patients with ELN > 30, as illustrated in Fig. 5 (panels a, b, c, d) and detailed in Table S5. The AUC of the novel classification significantly exceeded that of the 8th AJCC classification, with a clear advantage in terms of AIC and BIC, as depicted in Fig. 5 (panels e, f, g, h), and summarized in Table 3. Furthermore, for patients with ELN < 30, the rAJCC staging system still showed superiority than the 8^th AJCC system, as shown in Fig. S1.

Furthermore, the integrated model, which combines the ELN count with the rAJCC staging system, was compared with the original rAJCC model without ELN count integration. Given the inability to evaluate long-term survival rates for patients registered between 2016 and 2017 in the SEER database, the comparative analysis was restricted to the TCGA database and our single-center dataset. The findings indicated that the integrated model achieved a higher AUC, outperforming the rAJCC system in terms of predictive accuracy, as demonstrated in Fig. S2.