The prognostic value of log odds of positive lymph nodes in patients with resectable locally advanced gastric cancer following neoadjuvant therapy: results from the SEER database

Abstract

Background

Lymph node status is a critical prognostic factor for locally advanced gastric cancer (LAGC), but the comparative value of different lymph node metrics after neoadjuvant chemotherapy (NACT) remains unclear. This study aims to evaluate and compare the prognostic utility of positive lymph nodes (pLNs), lymph node ratio (LNR), and the log odds of positive lymph nodes (LODDS) in patients with LAGC undergoing radical gastrectomy after NACT.

Methods

A total of 1,976 patients from the Surveillance, Epidemiology, and End Results (SEER) database and a propensity score-matched (PSM) validation cohort of 69 Chinese patients were analyzed. A NACT-only subgroup (n=658) was evaluated to exclude radiotherapy effects. Prognostic value was assessed via Cox models, Kaplan-Meier curves, concordance indices (C-indices), and time-dependent receiver operating characteristic (ROC) analyses.

Results

In the SEER cohort, LODDS >−2.02 and LNR >0.12 showed comparable hazard ratios (HRs) for overall survival (OS) and cancer-specific survival (CSS), both outperforming pLNs. LODDS demonstrated the highest prognostic accuracy (OS: C-index 0.671; CSS: 0.690). In the NACT-only subgroup, LODDS remained the strongest predictor (OS: HR =3.34; CSS: HR =3.69) with superior C-indices (OS: 0.718; CSS: 0.730). An independent Chinese validation cohort (n=69) preliminarily supported LODDS’ prognostic strength (multivariable OS: HR =20.61, P=0.008; CSS: HR =14.58, P=0.02).

Conclusions

LODDS provides modest but robust and supplementary prognostic value in LAGC patients treated with neoadjuvant therapy. Its integration into existing staging systems may enhance individualized risk stratification and guide clinical decision-making.

Highlight box.

Key findings

• Lymph node status is a key prognostic factor in gastric cancer (GC), with positive lymph nodes (pLNs), lymph node ratio (LNR), and log odds of positive lymph nodes (LODDS) previously studied in surgical populations, but their comparative value in neoadjuvant-treated locally advanced cases remains unclear.

What is known and what is new?

• While pLNs, LNR, and LODDS are established prognostic markers in GC after upfront surgery, their comparative accuracy in patients receiving neoadjuvant therapy—which alters lymph node yield—remains uncertain.

• This study identifies LODDS as the superior prognostic predictor over pLNs and LNR specifically for LAGC treated with neoadjuvant chemotherapy, demonstrating significantly higher accuracy in risk stratification despite treatment-induced nodal changes.

What is the implication, and what should change now?

• LODDS should be incorporated into clinical staging and risk assessment models for GC patients undergoing neoadjuvant therapy to improve prognostic accuracy and guide personalized treatment decisions.

Introduction

Gastric cancer (GC), among the predominant malignancies, ranks fifth in occurrence and death worldwide (1). Owing to the absence of specific clinical manifestations in its early stages, most are diagnosed at the locally advanced gastric cancer (LAGC) phase, which frequently involves invasion of the muscularis propria and regional lymph node metastasis, thereby leading to an unfavorable prognosis. Early diagnosis and timely intervention are essential for improving the clinical outcomes of GC patients (2,3). Radical surgical resection combined with neoadjuvant therapy, particularly neoadjuvant chemotherapy (NACT), can improve the prognosis of the LAGC population (4-8), and is recommended as the standard treatment in many guidelines (5,9-11). Lymph node (LN) is significant in the GC prognosis assessment (12,13). The 8th American Joint Committee on Cancer (AJCC) tumor-node-metastasis (TNM) staging system remains the most extensively adopted framework for prognostic evaluation in GC (14). However, its N classification based merely on the number of positive lymph nodes (pLNs), presents several notable limitations: (I) it fails to incorporate the total count of examined lymph nodes (ELNs), thereby lowering staging accuracy in limited nodal dissection (<15 LNs); (II) it relies on categorical stratification, which possibly obscure the continuous nature of metastatic burden; and (III) it lacks robustness against extreme values, often resulting in understaging when lymphadenectomy is inadequate (e.g., ≤5 ELNs) (15). Therefore, identifying more reliable prognostic indicators is imperative for better survival prediction for the LAGC population.

Recent studies have explored alternative LN-related metrics like the lymph node ratio (LNR, ratio of pLNs to ELNs) and absolute count of pLNs (16,17). A 2025 systematic review and meta-analysis further confirmed that an elevated LNR was strongly associated with poorer overall survival (OS) and disease-free survival (DFS) in patients with GC undergoing neoadjuvant chemotherapy, underscoring the robustness of nodal ratio-based indices in this clinical context (18). The log odds of positive lymph nodes (LODDS), which integrates pLNs and negative lymph nodes (nLNs), is an emerging prognosis factor in tumors like colorectal, esophageal, pancreatic, and lung cancers (19-23). Compared with LNR and pLNs, LODDS possibly offers superior prognostic accuracy in certain oncologic contexts (19,22). Although LODDS has been established as a prognostic marker in patients undergoing upfront surgery for GC, large-scale real-world evidence and independent external validation specifically in neoadjuvant-treated LAGC remain limited. Notably, a 2025 meta-analysis demonstrated the overall superiority of LODDS over pLN and LNR in prognostic discrimination for GC (24). Therefore, our study seeks to unravel the links of LNR, pLNs, and LODDS to OS in this patient population and to elucidate their respective roles in predicting clinical outcomes following NACT. We present this article in accordance with the STROBE and TRIPOD reporting checklists (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-0430/rc).

Methods

Study design and patients

A retrospective cohort study was performed on GC individuals who underwent radical gastrectomy after neoadjuvant therapy across 17 research centers and leveraged data from the Surveillance, Epidemiology, and End Results (SEER) database from January 2004 to December 2019.

Inclusion criteria were: (I) age ≥18 at diagnosis; (II) primary non-metastatic LAGC, defined as cT1N2–3M0 or cT2–4N0/+M0 as per pre-treatment computed tomography (CT) imaging or diagnostic laparoscopy and staged as per the 8th AJCC TNM Staging Manual; (III) radical gastrectomy for GC; (IV) NACT treatment; (V) histopathology codes based on the International Classification of Diseases for Oncology, 3rd Edition (ICD-O-3) (25,26), including 8144/3 (intestinal adenocarcinoma), 8140/3 [adenocarcinoma, not otherwise specified (NOS)], 8010/3 (carcinoma, NOS), and 8211/3 (tubular adenocarcinoma). The diffuse subtype encompassed codes 8145/3 (carcinoma, diffuse), 8490/3 (signet-ring cell carcinoma), as well as 8142/3 (linitis plastica).

Exclusion criteria were: (I) GC diagnosed via autopsy or death certificate reports; (II) received palliative surgery; (III) concurrent with other primary malignancies; (IV) did not undergo LN biopsy; (V) unknown number of pLNs; (VI) incomplete follow-up or clinicopathologic information.

The findings from the SEER cohort were confirmed via analysis of an independent retrospective cohort of LAGC cases undergoing radical gastrectomy at West China Hospital, Sichuan University between January 2017 and December 2024. All patients in this cohort received NACT consisting of fluorouracil- and platinum-based regimens. Eligibility criteria aligned with those applied in the SEER set. Given the ongoing controversy surrounding the role of perioperative radiotherapy in LAGC and the lack of its integration into standardized treatment protocols in China, this study applied strict exclusion criteria to eliminate patients who received preoperative or postoperative radiotherapy. Consequently, the final cohort was restricted to treatment-naïve LAGC patients who underwent NACT alone. After propensity score-matched (PSM) based on key clinical variables, 69 of 161 initially eligible patients were encompassed in the ultimate comparative analysis. This study was performed as per the protocol approved by the Medical Ethics Committee of West China Hospital, Sichuan University (No. 2024-1546). Since our analysis involved de-identified retrospective data, written informed consent was unnecessary. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Variables

In this study, GC patient data were gathered via SEER*STAT 8.4.0 (https://seer.cancer.gov/seerstat/). The number of pLNs and total ELNs were extracted directly from SEER. LNR is derived as pLNs/the number of ELNs. LODDS was calculated as log[(pLNs + 0.05)/(ELNs − pLNs + 0.05)] (27-29). Optimal cutoff values were determined using the maximally selected rank statistic (survminer, R 4.3.2). Patients were stratified into LODDS-based risk groups as follows: low-risk (≤−2.02, favorable prognosis) and high-risk (>−2.02, poor prognosis). Corresponding cutoffs were defined as LNR (≤0.12 vs. >0.12) and pLN count (≤1 vs. >1).

For patients diagnosed between 2004 and 2009, the applied TNM staging system was the 6th edition (30). For those diagnosed between 2010 and 2017, the 7th edition was utilized (31). For individuals diagnosed between 2018 and 2019, staging was executed as per the 8th edition (32). In this study, tumor staging for all patients was ultimately re-staged as per the 8th edition for GC (14).

Other variables include marital status, sex, race, age, tumor grade, AJCC TNM stage, radiotherapy, tumor size, and follow-up time.

Outcomes

Our primary outcomes are OS and cancer-specific survival (CSS). The former is the interval from diagnosis to mortality due to all causes, whereas the latter is that from diagnosis to mortality attributable specifically to GC. Our follow-up spanned from diagnosis until death, loss to follow-up, or the study termination date (December 31, 2019).

Statistical analysis

The best cutoff values for LNR, pLNs, and LODDS were obtained via the surv_cutpoint function from the R survminer package 4.3.2. This method leverages the maximally selected rank statistic to identify thresholds that maximize survival differences between risk-stratified groups. Before categorical stratification, all three metrics were first validated as continuous variables in univariate and multivariate Cox models (all P<0.001). For each metric, the algorithm iteratively evaluates all possible cutoff points using OS as the endpoint. The significance of the optimal cutoff was validated by Kaplan-Meier (KM) analysis with log-rank testing (two-tailed ɑ=0.05). To mitigate overfitting, the stability of the selected cutoffs was confirmed through 1,000 bootstrap resampling iterations. They were split into high-risk (above the cutoff) and low-risk (below the cutoff) cohorts. To evaluate cutoff stability, sensitivity analyses were conducted using bootstrap resampling (1,000 iterations). All analyses followed the TRIPOD guidelines for prognostic model development (33). Continuous variables were analyzed in their original form without prior categorization to preserve the full informational content. Code reproducibility was ensured via R Markdown version control.

All the continuous variables included in this paper were non-normally distributed in the Shapiro-Wilk test (34). Continuous data were analyzed via the Wilcoxon rank-sum test, with results shown in median and interquartile range (IQR). Categorical data comparison was executed in the Pearson χ² test, with results shown in frequencies and percentages. Categorical data analyses were enabled through the chi-square or Fisher’s exact test. Survival rates were compared via the KM and log-rank test. The relations of pLNs, LNR, LODDS, and relevant clinicopathological data to OS, reporting hazard ratios (HRs) and 95% confidence intervals (CIs), were examined utilizing univariate and multivariate Cox proportional hazards models. Marital status, sex, race, age, tumor grade, AJCC TNM stage, radiotherapy, and tumor size were considered significant influencing factors. The links of pLNs, LNR, LODDS, and TNM to OS and CSS were explored via further subgroup analyses by the AJCC N stage. Additionally, the accuracy of LNR, pLNs, LODDS, and TNM for survival prediction was examined using the concordance index (C-index) or area under the curve (AUC). C-index measures the concordance between forecast and observed outcomes, serving as an indicator of the model’s predictive ability. 0.5 denotes no prediction ability, while 1 implies perfect concordance between forecast and actual results, signifying a strong prediction model.

Besides, a subgroup analysis was carried out specifically on people receiving only NACT, with exclusion of those who had undergone prior adjuvant or neoadjuvant radiotherapy.

Additionally, survival analysis was performed following PSM to balance clinical characteristics between the study groups. Variables included in the PSM process encompassed all clinical features except survival outcomes (OS and CSS), like marital status, sex, race, age, tumor grade, AJCC TNM, T, and N stages, receipt of radiotherapy, as well as tumor size.

Finally, within the PSM cohort derived from the West China Database, the prognostic performance of LODDS, LNR, number of pLNs, and the TNM staging system in survival prediction was rated utilizing C-index and AUC.

Our statistical analyses were enabled by R 4.2.1 (http://www.r-project.org/). R packages encompassed dplyr, compareGroups, plyr, MatchIt, ggplot2, survival, survminer, timeROC, car, gtsummary, and pROC. A two-sided P<0.05 signified statistical significance.

Results

Patient characteristics

A total of 1,976 LAGC individuals from the SEER database (development cohort) after neoadjuvant therapy were encompassed (Figure 1). The characteristics of patients stratified by AJCC N stage in the development cohort are presented in Table S1. Among the 1,976 patients included, 1,494 (75.6%) were male, and 1,639 (82.9%) were White. The median age was 65.0 (IQR, 57.0–71.0) years. 914 patients (46.3%) presented with grade III tumors, 1,078 (54.6%) were classified as AJCC stage II, and 1,178 (59.6%) as AJCC T3. Radiotherapy was administered to 1,318 patients (66.7%), among whom 1,220 (61.7%) received it before surgery. In 1,159 cases (58.7%), the tumor size was less than 2 cm. The median follow-up spanned 22.0 (IQR, 11.0–48.0) months. As our follow-up ended, 1,029 patients (52.1%) were alive, while 947 (47.9%) had died, including 805 (40.7%) deaths attributable to GC. The median numbers of ELNs and pLNs were 18.0 (IQR, 12.0–26.0) and 0.00 (IQR, 0.00–3.00). 1,371 patients (69.4%) had a LODDS ≤−2.02, 1,371 (69.4%) exhibited an LNR ≤0.12, and 1,284 (65.0%) had a pLN ≤1.

Figure 1.

Flow chart for screening of study patients. AJCC, American Joint Committee on Cancer; ICD-O-3, International Classification of Diseases for Oncology, 3rd Edition; NOS, not otherwise specified; SEER, the Surveillance, Epidemiology, and End Results database.

Marital status, sex, and age distributions were not significantly different across AJCC N stage groups (P>0.05). However, disparities were statistically significant in race, tumor grade, AJCC and AJCC T stages, receipt of radiotherapy, numbers of ELNs and pLNs, tumor size, LODDS, LNR, as well as pLNs (all P<0.05).

A subgroup analysis on the development cohort with LAGC who underwent only NACT was carried out. Baseline characteristics stratified by AJCC N stage are shown in Table S2. Among the analyzed 658 individuals, 425 (64.6%) were male and 472 (71.7%) were White. Their median age was 66.0 (IQR, 57.0–72.0) years. Poorly differentiated tumors (grade III) were observed in 335 patients (50.9%). AJCC stage II and III were reported in 282 (42.9%) and 264 (40.1%) patients. 332 patients (50.5%) were classified as AJCC T3, and 290 (44.1%) had tumors >5 cm in size. The median follow-up lasted 22.0 (IQR, 10.2–45.8) months. At the last follow-up, 379 (57.6%) remained alive, whereas 279 (42.4%) had died, including 239 (36.3%) deaths due to GC. The median numbers of ELNs and pLNs were 23.0 (IQR, 15.0–31.0) and 1.0 (IQR, 0.0–5.0). LODDS ≤−2.02 was observed in 416 patients (63.2%), LNR ≤0.12 in 421 (64.0%), and ≤1 pLN in 376 (57.1%). Baseline characteristics like marital status, sex, race, age, and tumor size were comparable among AJCC N stage subgroups (P>0.05). In contrast, significant differences were identified concerning tumor differentiation, AJCC stage, AJCC T stage, number of ELNs, LODDS, LNR, and pLNs (all P<0.05).

In addition, a validation cohort comprising 69 patients from a Chinese population was encompassed. Their characteristics stratified by AJCC N stage in this cohort are provided in Table S3.

All patients received NACT with fluorouracil and platinum-based regimens. The cohort encompassed 51 males (73.9%), with the median age being 60.0 (IQR, 53.0–68.0) years. Tumor grading revealed 36 cases (52.2%) as grade II. AJCC staging classified 25 patients (36.2%) as stage III, and 28 (40.6%) as T3. 63 patients (91.3%) completed three cycles of NACT. Surgical intervention included total gastrectomy in 40 patients (58.0%). Tumor size ranged from 2 to 5 cm in 50 cases (72.5%).

With a median follow-up lasting 67.4 (IQR, 63.4–74.2) months, 55 patients (79.7%) were alive as the follow-up ended, while 14 (20.3%) had died, including 11 (15.9%) deaths due to GC progression. The median number of ELNs per patient was 41.3±11.8, and that of pLNs was 1.00 (IQR, 0.00–7.00). LODDS ≤−2.02 was observed in 49 patients (71.0%), LNR ≤0.12 in 49 (71.0%), and ≤1 pLN in 37 (53.6%).

Sex, age, chemotherapy cycle, clinical evaluation, and surgical procedure were similar for these patients (P>0.05), while tumor grade, AJCC stage, AJCC T stage, positive nodes, Tumor size, LODDS, LNR, and pLNs were markedly different in the population distribution (P<0.05).

Survival analysis of the development cohort

Correlations of LODDS, LNR, and pLNs with OS and CSS in the overall development cohort of LAGC patients

Figure 2 displays the KM curves of OS and CSS influenced by LODDS in the overall development cohort of LAGC Patients. The median OS and CSS for the total development set were 40 (IQR, 37–45) and 44 (IQR, 39–50) months. Within the overall development cohort, LODDS >−2.02 is related to less favorable OS and CSS than those with LODDS ≤−2.02 (Figure 2). In patients with a LODDS >−2.02, the median OS was 19 (IQR, 18–21) months, whereas in those with a LODDS ≤−2.02, the median OS was 72 (IQR, 58–97) months. The median CSS for patients with a LODDS >−2.02 was 19 (IQR, 18–21) months.

Figure 2.

Kaplan-Meier survival curves for the impact of LODDS on OS (A) and CSS (B) of LAGC patients in the development cohort. CSS, cancer-specific survival; LAGC, locally advanced gastric cancer; LODDS, log odds of positive lymph nodes; OS, overall survival.

Table 1 presents the relations of LODDS and clinicopathological characteristics to OS and CSS in the overall development cohort of LAGC patients. Tables S4,S5 display the corresponding links of LNR, pLNs, and clinicopathological characteristics to OS and CSS.

Table 1. The correlations of LODDS with OS and CSS in the overall development cohort of patients with LAGC.

Variables	OS						CSS
	Univariable analysis			Multivariable analysis			Univariable analysis			Multivariable analysis
	HR (95% CI)	P		HR (95%CI)	P		HR (95% CI)	P		HR (95% CI)	P
LODDS
≤−2.02	Ref			Ref			Ref			Ref
>−2.02	2.98 (2.62–3.39)	<0.001		2.84 (2.48–3.25)	<0.001		3.40 (2.95–3.91)	<0.001		3.16 (2.73–3.65)	<0.001
Marital
Others	Ref			Ref			Ref			Ref
Married	0.77 (0.67–0.88)	<0.001		0.74 (0.64–0.85)	<0.001		0.79 (0.68–0.91)	0.001		0.75 (0.64–0.87)	<0.001
Sex
Male	Ref			Ref			Ref			Ref
Female	0.88 (0.75–1.03)	0.10		0.89 (0.75–1.04)	0.14		0.84 (0.71–0.99)	0.04		0.88 (0.73–1.05)	0.14
Race
White	Ref			Ref			Ref			Ref
Others	0.88 (0.74–1.05)	0.15		0.85 (0.71–1.02)	0.08		0.87 (0.72–1.05)	0.15		0.82 (0.68–1.00)	0.048
Age	1.01 (1.01–1.02)	<0.001		1.01 (1.01–1.02)	<0.001		1.01 (1.00–1.02)	0.004		1.01 (1.01–1.02)	<0.001
Grade
I	Ref			Ref			Ref			Ref
II	0.89 (0.61–1.30)	0.54		0.80 (0.55–1.18)	0.26		1.04 (0.65–1.64)	0.88		0.94 (0.59–1.50)	0.80
III	1.27 (0.88–1.84)	0.20		1.11 (0.76–1.61)	0.59		1.57 (1.00–2.45)	0.049		1.32 (0.84–2.08)	0.22
IV	0.84 (0.44–1.62)	0.61		0.65 (0.34–1.25)	0.20		1.11 (0.54–2.28)	0.77		0.88 (0.43–1.81)	0.73
Unknown	0.78 (0.52–1.18)	0.24		0.86 (0.57–1.31)	0.49		0.90 (0.55–1.47)	0.68		1.06 (0.65–1.75)	0.81
T
T1	Ref			Ref			Ref			Ref
T2	1.54 (1.14–2.08)	0.005		1.30 (0.96–1.76)	0.09		1.93 (1.34–2.78)	<0.001		1.57 (1.08–2.26)	0.02
T3	1.56 (1.18–2.06)	0.002		1.23 (0.93–1.63)	0.15		1.92 (1.36–2.70)	<0.001		1.48 (1.04–2.09)	0.03
T4	2.28 (1.67–3.12)	<0.001		1.62 (1.17–2.24)	0.003		3.02 (2.08–4.38)	<0.001		2.01 (1.37–2.94)	<0.001
Radiation
No				Ref						Ref
Yes	1.14 (0.99–1.31)	0.07		1.30 (1.12–1.51)	<0.001		1.13 (0.97–1.32)	0.10		1.33 (1.13–1.56)	<0.001
Tumor size, cm
<2	Ref			Ref			Ref			Ref
2–5	0.93 (0.50–1.74)	0.82		0.99 (0.53–1.86)	0.98		0.89 (0.45–1.79)	0.75		0.91 (0.45–1.85)	0.80
>5	0.67 (0.57–0.79)	<0.001		0.77 (0.64–0.92)	0.004		0.59 (0.49–0.70)	<0.001		0.68 (0.56–0.82)	<0.001

Multivariable Cox regression analysis adjusted for LODDS, marital, sex, race, age, grade, T, radiation, and tumor size. CI, confidence interval; CSS, cancer-specific survival; HR, hazard ratio; LAGC, locally advanced gastric cancer; LODDS, log odds of positive lymph nodes; OS, overall survival; T, tumor.

In univariable Cox regression analysis, LAGC patients with LODDS >−2.02 (HR =2.98; 95% CI: 2.62–3.39; P<0.001), LNR >0.12 (HR =2.98; 95% CI: 2.62–3.39; P<0.001), or pLNs >1 (HR =2.60; 95% CI: 2.29–2.96; P<0.001) demonstrated markedly worse OS. These associations remained statistically significant following adjustment for possible confounding variables: LODDS >−2.02 (HR =2.84; 95% CI: 2.48–3.25; P<0.001), LNR >0.12 (HR =2.84; 95% CI: 2.48–3.25; P<0.001), and pLNs >1 (HR =2.52; 95% CI: 2.20-2.88; P<0.001). After multivariable adjustment, statistically significant links were not observed between OS and sex, race, tumor grade, AJCC T category, or tumor size (P>0.05).

A similar pattern was found for CSS within the development set. In univariable analysis, elevated LODDS >−2.02 (HR =3.40; 95% CI: 2.95–3.91; P<0.001), LNR >0.12 (HR =3.40; 95% CI: 2.95–3.91; P<0.001), and pLNs >1 (HR =2.97; 95% CI: 2.58–3.41; P<0.001) strongly correlated with lowered CSS. These associations remained robust in multivariable models adjusting for confounders: LODDS >−2.02 (HR =3.16; 95% CI: 2.73–3.65; P<0.001), LNR>0.12 (HR =3.16; 95% CI: 2.73–3.65; P<0.001), and pLNs >1 (HR =2.78; 95% CI: 2.40–3.22; P<0.001). No marked links were identified between CSS and sex, race, tumor grade, AJCC T category, or tumor size after adjustment (P>0.05).

Correlations of LODDS, LNR, and pLNs with OS and CSS in the development cohort at different AJCC N stages

Figure 3 presents the KM survival curves displaying how LODDS influences OS and CSS among LAGC individuals in the development cohort, stratified by AJCC N stage. In the AJCC N1 (Figure 3A,3B) and N2 (Figure 3C,3D) subgroups, LODDS >−2.02 was markedly linked to less favorable OS and CSS than LODDS ≤−2.02. However, no statistically significant associations were observed in the AJCC N3 subgroup (OS: HR =2.00, P=0.14; CSS: HR =2.15, P=0.11; Figure 3E,3F), a finding that remained consistent across both the NACT-only cohort and the PSM cohort.

Figure 3.

Kaplan-Meier survival curves for the impact of LODDS on OS and CSS of LAGC patients in the development cohort with different N stages. (A) OS and (B) CSS of the N1 population in the development cohort; (C) OS and (D) CSS of the N2 population in the development cohort; (E) OS and (F) CSS of the N3 population in the development cohort. CSS, cancer-specific survival; LAGC, locally advanced gastric cancer; LODDS, log odds of positive lymph nodes; N, node; OS, overall survival.

Table 2 and Table S6 summarize the relations of LODDS and LNR to OS and CSS across AJCC N stages in the development cohort. In the OS analysis, following multivariable adjustment, LAGC patients with LODDS >−2.02 or LNR >0.12 demonstrated significantly poorer OS in the AJCC N1 (HR =2.50; 95% CI: 2.02–3.09; P<0.001) and AJCC N2 (HR =2.75; 95% CI: 1.96–3.86; P<0.001) groups. In contrast, no significant relation between LODDS or LNR and OS was observed in the AJCC N3 group (HR =2.00; 95% CI: 0.79–5.08; P=0.14).

Table 2. Association of LODDS with OS and CSS in the development cohort at different AJCC N stages.

Variable	N-staging	OS						CSS
		Univariable analysis			Multivariable analysis			Univariable analysis			Multivariable analysis
		HR (95% CI)	P		HR (95% CI)	P		HR (95% CI)	P		HR (95% CI)	P
LODDS (Ref.: ≤−2.02)
>−2.02 vs. ≤−2.02	N1	2.65 (2.16–3.24)	<0.001		2.50 (2.02–3.09)	<0.001		3.03 (2.43–3.77)	<0.001		2.80 (2.22–3.52)	<0.001
	N2	2.69 (1.95–3.73)	<0.001		2.75 (1.96–3.86)	<0.001		3.02 (2.13–4.29)	<0.001		3.08 (2.13–4.43)	<0.001
	N3	2.08 (0.85–5.11)	0.11		2.00 (0.79–5.08)	0.14		2.11 (0.86–5.18)	0.10		2.15 (0.84–5.52)	0.11
Marital (Ref.: others)
Married vs. others	N1	0.88 (0.71–1.09)	0.25		0.81 (0.65–1.01)	0.06		0.88 (0.70–1.11)	0.29		0.78 (0.61–0.99)	0.04
	N2	0.66 (0.50–0.89)	0.006		0.72 (0.53–0.98)	0.03		0.69 (0.50–0.95)	0.02		0.75 (0.54–1.05)	0.10
	N3	0.71 (0.49–1.01)	0.06		0.61 (0.41–0.91)	0.01		0.72 (0.50–1.05)	0.09		0.63 (0.42–0.95)	0.03
Sex (Ref.: male)
Female vs. male	N1	0.90 (0.70–1.15)	0.40		1.01 (0.78–1.32)	0.92		0.83 (0.63–1.10)	0.20		0.74 (0.52–1.05)	0.95
	N2	0.92 (0.50–0.89)	0.63		0.90 (0.63–1.28)	0.55		0.86 (0.59–1.25)	0.43		0.87 (0.59–1.29)	0.49
	N3	0.78 (0.53–1.15)	0.20		0.65 (0.42–1.00)	0.052		0.75 (0.50–1.13)	0.17		0.62 (0.39–0.97)	0.03
Race (Ref.: White)
Others vs. White	N1	0.74 (0.54–1.01)	0.054		0.77 (0.56–1.06)	0.10		0.72 (0.51–1.00)	0.053		0.74 (0.52–1.05)	0.09
	N2	0.72 (0.49–1.06)	0.09		0.73 (0.49–1.09)	0.12		0.73 (0.49–1.11)	0.14		0.74 (0.49–1.13)	0.16
	N3	0.98 (0.67–1.44)	0.93		0.87 (0.57–1.33)	0.52		0.98 (0.67–1.45)	0.92		0.88 (0.58–1.36)	0.57
Age	N1	1.02 (1.01–1.03)	0.001		1.02 (1.01–1.03)	0.001		1.01 (1.00–1.02)	0.01		1.01 (1.00–1.03)	0.009
	N2	1.00 (0.99–1.02)	0.71		1.00 (0.99–1.01)	0.76		1.00 (0.99–1.02)	0.86		1.00 (0.99–1.02)	0.69
	N3	1.01 (1.00–1.03)	0.08		1.01 (1.00–1.03)	0.03		1.01 (1.00–1.02)	0.13		1.01 (1.00–1.03)	0.052
Grade (Ref.: I)
II vs. I	N1	0.88 (0.52–1.49)	0.62		0.74 (0.43–1.27)	0.27		1.10 (0.57–2.12)	0.78		0.94 (0.49–1.84)	0.86
	N2	0.99 (0.24–4.09)	0.99		0.90 (0.21–3.77)	0.88		1.38 (0.19–10.05)	0.75		1.01 (0.14–7.54)	0.99
	N3	2.75 (0.36–20.8)	0.33		1.61 (0.20–13.18)	0.65		2.66 (0.35–20.33)	0.35		1.61 (0.19–13.53)	0.66
III vs. I	N1	1.18 (0.71–1.96)	0.53		0.98 (0.58–1.65)	0.94		1.57 (0.83–2.97)	0.16		1.28 (0.67–2.44)	0.45
	N2	1.24 (0.31–5.02)	0.76		1.27 (0.31–5.19)	0.74		1.67 (0.23–11.98)	0.61		1.38 (0.19–10.02)	0.75
	N3	3.85 (0.54–27.7)	0.18		2.67 (0.35–20.15)	0.34		3.71 (0.52–26.64)	0.19		2.72 (0.36–20.83)	0.34
IV vs. I	N1	0.74 (0.29–1,89)	0.53		0.55 (0.21–1.42)	0.21		1.01 (0.35–2.97)	0.98		0.82 (0.28–2.42)	0.71
	N2	1.55 (0.28–8.45)	0.61		1.04 (0.19–5.75)	0.97		2.21 (0.25–19.83)	0.48		1.10 (0.12–10.14)	0.93
	N3	1.91 (0.17–21.2)	0.60		2.73 (0.24–31.05)	0.42		1.87 (0.17–20.69)	0.61		3.00 (0.26–34.53)	0.38
Unknown vs. I	N1	0.68 (0.39–1.21)	0.19		0.71 (0.39–1.27)	0.24		0.88 (0.44–1.77)	0.72		0.96 (0.47–1.96)	0.91
	N2	1.25 (0.29–5.36)	0.77		1.28 (0.29–5.71)	0.75		1.66 (0.22–12.51)	0.62		1.49 (0.19–11.67)	0.71
	N3	2.51 (0.33–19.0)	0.37		2.45 (0.30–19.99)	0.40		2.67 (0.35–20.37)	0.34		3.65 (0.43–31.36)	0.24
T (Ref.: T1)
T2 vs. T1	N1	0.93 (0.59–1.47)	0.77		0.88 (0.56–1.40)	0.59		1.08 (0.63–1.84)	0.79		0.99 (0.58–1.70)	0.97
	N2	2.56 (0.77–8.53)	0.13		2.54 (0.75–8.62)	0.13		3.77 (0.89–16.05)	0.07		3.35 (0.77–14.57)	0.11
	N3	4.62 (0.59–36.4)	0.15		3.35 (0.38–29.78)	0.28		4.33 (0.54–34.48)	0.17		4.45 (0.47–41.65)	0.19
T3 vs. T1	N1	0.82 (0.54–1.26)	0.37		0.85 (0.55–1.31)	0.46		0.94 (0.56–1.56)	0.80		0.97 (0.58–1.62)	0.90
	N2	2.08 (0.66–6.54)	0.21		2.19 (0.68–7.00)	0.19		2.88 (0.71–11.64)	0.14		2.89 (0.70–11.96)	0.14
	N3	2.09 (0.29–15.07)	0.46		1.76 (0.22–14.05)	0.59		2.11 (0.29–15.23)	0.46		2.53 (0.30–21.09)	0.39
T4 vs. T1	N1	1.01 (0.609–1.678)	0.97		1.02 (0.61–1.71)	0.94		1.24 (0.69–2.20)	0.47		1.20 (0.67–2.17)	0.54
	N2	2.45 (0.75–8.04)	0.21		2.41 (0.73–7.98)	0.15		3.51 (0.84–14.73)	0.09		3.24 (0.77–13.71)	0.11
	N3	2.73 (0.38–19.81)	0.32		2.51 (0.31–20.48)	0.39		2.76 (0.38–20.08)	0.32		3.75 (0.44–31.67)	0.22
Radiation (Ref.: no)
Yes vs. no	N1	1.30 (1.01,1.66)	0.04		1.31 (1.01–1.71)	0.04		1.33 (1.02–1.75)	0.04		1.32 (0.99–1.75)	0.06
	N2	1.44 (1.05–1.96)	0.02		1.60 (1.15–2.24)	0.006		1.38 (1.00–1.92)	0.052		1.61 (1.13–2.29)	0.008
	N3	1.02 (0.72–1.44)	0.91		1.08 (0.75–1.56)	0.67		1.10 (0.76–1.57)	0.62		1.23 (0.83–1.81)	0.30
Tumor size (Ref.: <2 cm)
2–5 vs. <2 cm	N1	0.61 (0.20–1.92)	0.40		0.70 (0.22–2.21)	0.542		0.68 (0.22–2.13)	0.51		0.72 (0.23–2.28)	0.57
	N2	–	–		–	–		–	–		–	–
	N3	1.34 (0.55–3.30)	0.52		1.61 (0.62–4.18)	0.32		1.35 (0.55–3.32)	0.52		1.67 (0.64–4.36)	0.29
>5 vs. <2 cm	N1	0.65 (0.49–0.84)	0.001		0.81 (0.610–1.084)	0.16		0.58 (0.43–0.77)	<0.001		0.73 (0.53–1.00)	0.048
	N2	0.80 (0.57–1.12)	0.19		0.87 (0.59–1.27)	0.47		0.68 (0.47–0.98)	0.04		0.73 (0.48–1.11)	0.14
	N3	0.56 (0.36–0.86)	0.008		0.59 (0.36–0.98)	0.04		0.54 (0.35–0.84)	0.006		0.49 (0.28–0.87)	0.01

Multivariable Cox regression analysis adjusted for LODDS, marital, sex, race, age, grade, T, radiation, and tumor size. AJCC N, American Joint Committee on Cancer lymph node staging; CI, confidence interval; CSS, cancer-specific survival; HR, hazard ratio; LODDS, log odds of positive lymph nodes; OS, overall survival; T, tumor.

Consistent results were observed in the CSS analysis. Multivariable analysis indicated that LODDS >−2.02 or LNR >0.12 was notably linked to poorer CSS in the AJCC N1 (HR =2.80; 95% CI: 2.22–3.52; P<0.001) and AJCC N2 (HR =3.08; 95% CI: 2.13–4.43; P<0.001) subgroups, while no significant relation was noted in the AJCC N3 subgroup (HR =2.15; 95% CI: 0.84–5.52; P=0.11).

Table S7 details the prognostic value of the count of pLNs on OS and CSS across AJCC N stages in the development cohort. In the OS analysis, following possible confounder adjustment, pLNs >1 markedly correlated with reduced OS in the AJCC N1 (HR =1.96; 95% CI: 1.59–2.42; P<0.001) and AJCC N2 (HR =3.38; 95% CI: 2.22–5.13; P<0.001) groups. Nevertheless, the relation was insignificant in the AJCC N3 cohort (HR =2.00; 95% CI: 0.79–5.08; P=0.14).

Similar trends were observed for CSS. Multivariable analysis demonstrated that pLNs >1 was significantly related to decreased CSS in the AJCC N1 (HR =2.17; 95% CI: 1.73–2.72; P<0.001) and AJCC N2 (HR =3.83; 95% CI: 2.42–6.05; P<0.001) cohorts. No significant link was identified within the AJCC N3 subgroup (HR =2.15; 95% CI: 0.84–5.52; P=0.11).

Prediction effect of pLNs, LNR, LODDS, and TNM staging system on survival in the development cohort

Table 3 summarizes the performance of LODDS, pLNs, LNR, and TNM staging in predicting OS and CSS in the entire LAGC cohort, as well as in subgroups stratified by N stage. Compared with LNR [OS: C-index =0.653 (95% CI: 0.635–0.671); CSS: C-index =0.671 (95% CI: 0.652–0.691)], pLNs [OS: C-index =0.643 (95% CI: 0.624–0.661); CSS: C-index =0.659 (95% CI: 0.640–0.679)], and TNM staging [OS: C-index =0.579 (95% CI: 0.561–0.597); CSS: C-index =0.589 (95% CI: 0.570–0.608)], LODDS exhibited a superior discriminative ability for OS and CSS in the overall population [OS: C-index =0.671 (95% CI: 0.652–0.690); CSS: C-index =0.690 (95% CI: 0.670–0.710)].

Table 3. C-index of LODDS, LNR, pLNs, and TNM staging system on OS and CSS in the development cohort.

Outcomes	Populations	Variables	C-index (95% CI)	P
OS	Overall	LODDS	0.671 (0.652–0.690)	Ref.
		LNR	0.653 (0.635–0.671)	<0.001
		pLNs	0.643 (0.624–0.661)	<0.001
		TNM	0.579 (0.561–0.597)	<0.001
	N1 stage	LODDS	0.640 (0.609–0.671)	Ref.
		LNR	0.629 (0.599–0.658)	0.003
		pLNs	0.614 (0.584–0.643)	<0.001
		TNM	0.515 (0.495–0.535)	<0.001
	N2 stage	LODDS	0.674 (0.630–0.719)	Ref.
		LNR	0.673 (0.630–0.716)	0.14
		pLNs	0.620 (0.577–0.663)	0.056
		TNM	0.502 (0.476–0.528)	<0.001
	N3 stage	LODDS	0.634 (0.578–0.690)	Ref.
		LNR	0.634 (0.578–0.689)	0.53
		pLNs	0.573 (0.519–0.627)	0.02
		TNM	0.502 (0.485–0.518)	<0.001
CSS	Overall	LODDS	0.690 (0.670–0.710)	Ref.
		LNR	0.671 (0.652–0.691)	<0.001
		pLNs	0.659 (0.640–0.679)	<0.001
		TNM	0.589 (0.570–0.608)	<0.001
	N1 stage	LODDS	0.660 (0.627–0.692)	Ref.
		LNR	0.648 (0.616–0.679)	<0.001
		pLNs	0.632 (0.601–0.663)	<0.001
		TNM	0.514 (0.493–0.536)	<0.001
	N2 stage	LODDS	0.694 (0.650–0.738)	Ref.
		LNR	0.691 (0.647–0.735)	0.03
		pLNs	0.644 (0.601–0.688)	0.02
		TNM	0.502 (0.474–0.530)	<0.001
	N3 stage	LODDS	0.644 (0.589–0.699)	Ref
		LNR	0.644 (0.589–0.699)	0.35
		pLNs	0.565 (0.509–0.621)	0.02
		TNM	0.501 (0.483–0.520)	<0.001

AJCC, American Joint Committee on Cancer; C-index, concordance index; CI, confidence interval; CSS, cancer-specific survival; LNR, lymph node ratio; LODDS, log odds of positive lymph nodes; OS, overall survival; pLNs, positive lymph nodes; TNM staging system, the 8th Edition AJCC tumor-node-metastasis staging system.

In the N1, N2, and N3 subgroups, the C-index of LODDS for OS prediction was 0.640 (95% CI: 0.609–0.671), 0.674 (95% CI: 0.630–0.719), and 0.634 (95% CI: 0.578–0.690). For CSS, the values were 0.660 (95% CI: 0.627–0.692), 0.694 (95% CI: 0.650–0.738), and 0.644 (95% CI: 0.589–0.699).

In survival analyses, although LODDS demonstrated superior prognostic accuracy compared with pLNs, LNR, and TNM staging, the absolute improvement in discriminative performance was modest, and its clinical relevance should therefore be interpreted with caution. Therefore, LODDS should be regarded as a complementary risk stratification tool rather than a replacement for conventional staging systems. Within the N2 subgroup, LODDS provided significantly better OS prediction performance than LNR and TNM staging (P<0.05), and for CSS than pLNs, LNR, and TNM staging (P<0.05). In the N3 subgroup, although LODDS yielded a significantly higher C-index than pLNs and TNM staging (P<0.05), none of the three nodal metrics (LODDS, LNR, pLNs) showed statistically significant associations with OS or CSS in multivariable Cox regression (all P>0.05), indicating limited clinical utility for risk stratification in this population.

Figure 4 illustrates the temporal trends in the AUC for OS and CSS forecasts by LNR, pLNs, LODDS, as well as TNM staging within the development cohort. Figure 5 depicts the temporal variation in AUC for OS and CSS forecasts by these same indicators across different N-stage subgroups.

Figure 4.

The changes in the AUC over time for OS (A) and CSS (B) predicted by LNR, pLNs, LODDS, and TNM in the development cohort. AUC, area under the curve; CSS, cancer-specific survival; LNR, lymph node ratio; LODDS, log odds of positive lymph nodes; OS, overall survival; pLNs, positive lymph nodes; TNM, tumor-node-metastasis.

Figure 5.

The changes in AUC over time for OS and CSS predicted by LNR, pLNs, LODDS, and TNM with different N stages in the development cohort. (A,B) Changes in AUC for predicting (A) OS and (B) CSS in the stage N1 populations in the development cohort; (C,D) changes in AUC for predicting (C) OS and (D) CSS in the stage N2 populations in the development cohort; (E,F) changes in AUC for predicting (E) OS and (F) CSS in the stage N3 populations in the development cohort. AUC, area under the curve; CSS, cancer-specific survival; LNR, lymph node ratio; LODDS, log odds of positive lymph nodes; OS, overall survival; pLNs, positive lymph nodes; TNM, tumor-node-metastasis.

Survival and subgroup analyses after PSM in the development set

Following clinical variable adjustment of marital status, sex, race, age, tumor grade, AJCC T and N stages, radiotherapy, as well as tumor size, the baseline characteristics between groups were well balanced. PSM analysis was subsequently performed to evaluate OS and CSS. Figure S1 illustrates the KM survival curves depicting how LODDS impacts OS and CSS in the LAGC population after PSM in the development cohort. Figure S2 presents KM survival curves assessing LODDS’ prognosis value for OS and CSS across different AJCC N stages in the same cohort after PSM. The results of the PSM analysis demonstrated that, after adjustment for clinical covariates, patients with LODDS ≤−2.02 displayed markedly superior OS and CSS than individuals having LODDS >−2.02 (P<0.001; Figure S1). In subgroup analyses by AJCC N stage, LAGC patients with LODDS ≤−2.02 in the N1 and N2 categories showed markedly improved OS and CSS in comparison to their counterparts with LODDS >−2.02 (P<0.001; Figure S2A,S2B,S2D,S2E). In the N3 subgroup, although a significant OS benefit was noted for those with LODDS ≤−2.02 (P<0.001; Figure S2C), statistically significant disparity in CSS was not detected (P>0.05; Figure S2F).

Table S8 summarizes the clinical characteristics of GC individuals across AJCC N stages in the development cohort after PSM.

Survival analysis of the NACT-only cohort

Correlations of LODDS, LNR, and pLNs with OS and CSS in the NACT-only cohort

Figure S3 illustrates the KM curves for OS and CSS in the LAGC group receiving NACT alone, stratified by LODDS. Among these patients, the median OS and CSS were 42 (IQR: 37–56) and 49 (IQR: 40–97) months. Within this NACT-only cohort, a LODDS value >−2.02 was related to notably poorer OS and CSS than individuals with LODDS ≤−2.02, as depicted in Figure S3.

Tables S9-S11 provide a summary of the relations of LODDS, LNR, number of pLNs, and clinicopathological characteristics to OS and CSS in LAGC people receiving merely NACT. In univariable Cox regression analysis, LAGC patients with LODDS >−2.02 (HR =3.75; 95% CI: 2.95–4.78; P<0.001), LNR >0.12 (HR =3.58; 95% CI: 2.81–4.55; P<0.001), or pLNs >1 (HR =3.08; 95% CI: 2.41–3.92; P<0.001) demonstrated significantly reduced OS. These relations remained statistically significant after adjustment for possible confounders: LODDS >−2.02 (HR =3.34; 95% CI: 2.58–4.31; P<0.001), LNR >0.12 (HR =3.16; 95% CI: 2.45–4.07; P<0.001), and pLNs >1 (HR =2.78; 95% CI: 2.15–3.59; P<0.001). The relations to OS were insignificant for sex, race, tumor grade, AJCC T category, or tumor size following multivariable adjustment (P>0.05).

A pattern was similar for CSS in the NACT-only cohort. Univariable analysis revealed that LODDS >−2.02 (HR =4.40; 95% CI: 3.37–5.73; P<0.001), LNR >0.12 (HR =4.20; 95% CI: 3.23–5.46; P<0.001), and pLNs >1 (HR =3.58; 95% CI: 2.74–4.67; P<0.001) were strongly related to reduced CSS. These associations remained robust following multivariable adjustment: LODDS >−2.02 (HR =3.69; 95% CI: 2.80–4.87; P<0.001), LNR >0.12 (HR =3.51; 95% CI: 2.67–4.63; P<0.001), as well as pLNs >1 (HR =3.03; 95% CI: 2.29–4.01; P<0.001). As with OS, no significant associations with CSS were found for sex, race, tumor grade, or tumor size after adjustment (P>0.05).

Correlations of LODDS, LNR, and pLNs with OS and CSS in the NACT-only cohort at different AJCC N stages

Figure S4 presents the KM survival curves illustrating how LODDS influences OS and CSS among the LAGC population who underwent NACT alone, stratified by AJCC N stage. In the AJCC N1 (Figure S4A,S4B) and N2 (Figure S4C,S4D) subgroups, a LODDS value >−2.02 was significantly related to inferior OS and CSS in comparison to LODDS ≤−2.02. Consistent with the overall development cohort, none of the three nodal metrics (LODDS, LNR, pLNs) showed a statistically significant association with OS or CSS in the AJCC N3 subgroup (all P≥0.05; Figure S4E,S4F).

Tables S12-S14 summarize the links of LODDS, LNR, and the number of pLNs to OS and CSS across AJCC N stages in patients treated exclusively with NACT. In multivariable-adjusted OS analysis, patients in the AJCC N1 subgroup with LODDS >−2.02 (HR =3.40; 95% CI: 2.06–5.62; P<0.001), LNR >0.12 (HR =2.95; 95% CI: 1.77–4.92; P<0.001), and pLNs >1 (HR =1.77; 95% CI: 1.10–2.83; P=0.02) exhibited significantly poorer survival outcomes. In the AJCC N2 subgroup, LODDS >−2.02 (HR =3.54; 95% CI: 1.57–8.01; P=0.002) and LNR >0.12 (HR =3.56; 95% CI: 2.06–5.62; P=0.002) were notably linked to decreased OS. However, the link of LODDS, LNR, or pLNs to OS was not notable in the AJCC N3 subgroup (P>0.05).

Comparable results were derived in our CSS analysis. Multivariate analysis revealed that, in the AJCC N1 subgroup, LODDS >−2.02 (HR =4.19; 95% CI: 2.41–7.27; P<0.001), LNR >0.12 (HR =3.78; 95% CI: 2.17–6.59; P<0.001), and pLNs >1 (HR =2.01; 95% CI: 1.20–3.36; P=0.008) markedly correlated with poorer CSS. In the AJCC N2 subgroup, LODDS >−2.02 (HR =4.10; 95% CI: 1.71–9.83; P=0.002) and LNR >0.12 (HR =4.11; 95% CI: 1.72–9.86; P=0.002) similarly demonstrated significant associations with inferior CSS. No significant correlations between LODDS, LNR, or pLNs and CSS were found in the AJCC N3 subgroup (P>0.05).

Prediction effect of pLNs, LNR, LODDS, and TNM staging system on survival in the NACT-only cohort

Table S15 displays the performance of LODDS, pLNs, LNR, and the TNM staging system for OS and CSS prediction among LAGC patients treated with only NACT, including stratification across AJCC N stages. LODDS demonstrated superior discriminative ability for OS (C-index =0.718, 95% CI: 0.687–0.750) and CSS (C-index =0.730, 95% CI: 0.697–0.764), in comparison to LNR (OS: 0.703, 95% CI: 0.671–0.735; CSS: 0.719, 95% CI: 0.686–0.753), pLNs (OS: 0.690, 95% CI: 0.589–0.652; CSS: 0.703, 95% CI: 0.670–0.737), and TNM staging (OS: 0.620, 95% CI: 0.561–0.597; CSS: 0.632, 95% CI: 0.598–0.666).

In subgroup analyses stratified by N stage, the C-index values of LODDS in predicting OS were 0.678 (95% CI: 0.611–0.745), 0.697 (95% CI: 0.617–0.778), and 0.655 (95% CI: 0.582–0.728) for N1, N2, and N3. Regarding CSS, they were 0.700 (95% CI: 0.626–0.774), 0.705 (95% CI: 0.623–0.788), and 0.657 (95% CI: 0.585–0.729), respectively.

Time-dependent AUC curves comparing the accuracy of LODDS, LNR, pLNs, and TNM staging in forecasting OS and CSS among NACT-only patients are illustrated in Figure S5. Stratified time-dependent AUC curves by N stage are presented in Figure S6, further delineating the prognostic discrimination of these metrics in N1, N2, and N3 subgroups.

Survival and subgroup analysis after PSM in NACT-only patients

Following the adjustments of marital status, sex, race, age, tumor grade, AJCC T and N stages, radiotherapy, and tumor size, adequate covariate balance was achieved in the matched cohort. Survival outcomes (OS and CSS) were analyzed in this matched population.

KM survival curves illustrating the impact of LODDS on OS and CSS after PSM are displayed in Figure S7. Those with LODDS ≤−2.02 exhibited significantly improved OS and CSS in comparison to those with LODDS >−2.02 (P<0.001).

Figure S8 displays KM survival curves for OS and CSS by N stage in the PSM cohort. In N1 and N2 subgroups, patients with LODDS ≤−2.02 had markedly better OS and CSS than those with LODDS >−2.02 (all P<0.001; Figure S8A,S8B,S8D,S8E). However, within the N3 subgroup, differences in OS or CSS were statistically insignificant across risk strata defined by any of the three nodal metrics (LODDS, LNR, or pLNs) (all P>0.05; Figure S8C,S8F).

Baseline characteristics of NACT-only patients stratified by AJCC N stage after PSM are presented in Table S16.

Survival analysis of the validation cohort

Correlations of LODDS, LNR, and pLNs with OS and CSS in the validation cohort after PSM

After tumor grade and clinical evaluation were adjusted, PSM achieved balanced baseline characteristics across groups. Survival analyses for OS and CSS were subsequently carried out.

KM curves for OS and CSS according to LODDS classification in the validation cohort are shown in Figure S9. Markedly better OS and CSS were observed in people having LODDS ≤−2.02 in comparison to those with LODDS >−2.02 (P<0.05).

Table S17 presents baseline characteristics of GC individuals across AJCC N stages after PSM. Tables S18-S20 provide a detailed summary of the links of LODDS, LNR, pLNs, as well as clinicopathological variables to OS and CSS in the validation cohort following PSM.

In univariable Cox regression, patients with LODDS >−2.02 (HR =7.20; 95% CI: 1.56–33.12; P=0.01) or LNR >0.12 (HR =7.20; 95% CI: 1.56–33.12; P=0.01) had significantly poorer OS. The difference in OS was not marked across those with pLNs >1 and ≤1 (HR =7.13; 95% CI: 0.92–55.44; P=0.06). Following multivariable adjustment, these associations remained statistically significant for LODDS >−2.02 (HR =20.61; 95% CI: 2.21–191.90; P=0.008), LNR >0.12 (HR =20.61; 95% CI: 2.21–191.90; P=0.008), and pLNs >1 (HR =10.98; 95% CI: 1.10–109.40; P=0.04). Other variables, including sex, age, tumor grade, AJCC T category, tumor size, chemotherapy cycle, clinical evaluation, and surgical procedure, showed no statistically significant associations with OS in multivariable analysis (P>0.05).

Similarly, in univariable analysis, LODDS >−2.02 (HR =6.17; 95% CI: 1.30–29.27; P=0.02) as well as LNR >0.12 (HR =6.17; 95% CI: 1.30–29.27; P=0.02) were notably linked to reduced CSS. The difference in CSS was statistically insignificant across those with pLNs >1 and ≤1 (HR =6.08; 95% CI: 0.77–48.21; P=0.09). After adjusting for confounding variables, these associations remained robust for LODDS >−2.02 (HR =14.58; 95% CI: 1.58–134.50; P=0.02) and LNR >0.12 (HR =14.58; 95% CI: 1.58–134.50; P=0.02). However, the association between pLNs >1 and CSS was statistically insignificant in our multivariable analysis (HR =8.98; 95% CI: 0.91–88.81; P=0.06). Variables including sex, age, tumor grade, AJCC T category, tumor size, chemotherapy cycle, clinical evaluation, and surgical procedure were not significantly related to CSS (P>0.05).

Prediction effect of pLNs, LNR, LODDS, and TNM staging system on survival in the validation cohort after PSM

Table S21 presents the prognosis performance of LODDS, pLNs, LNR, and TNM staging in OS and CSS forecasting within the validation cohort following PSM. The LODDS metric demonstrated predictive performance comparable to that of other nodal indicators, with C-indices of 0.799 (95% CI: 0.682–0.915) and 0.770 (95% CI: 0.655–0.886) for OS and CSS. These values did not differ significantly from those related to LNR [OS: 0.800 (95% CI: 0.685–0.916); CSS: 0.773 (95% CI: 0.658–0.887)], pLNs [OS: 0.816 (95% CI: 0.712–0.921); CSS: 0.797 (95% CI: 0.692–0.902)], or TNM staging [OS: 0.783 (95% CI: 0.701–0.864); CSS: 0.793 (95% CI: 0.710–0.876)] (all P>0.05).

Figure S10 illustrates the temporal variation in the AUC for OS and CSS forecasts by LNR, pLNs, LODDS, and TNM within the validation set after PSM. The analysis demonstrated that the AUC values for pLNs exhibited more pronounced temporal fluctuations in prognostic accuracy in comparison to LNR, LODDS, and TNM staging. Notably, despite this temporal variability in pLNs’ performance, no statistically significant differences in time-dependent AUC were noted between LODDS and LNR, pLNs, or TNM staging (P>0.05 for all comparisons).

Discussion

Our study retrospectively evaluated survival data from 1,976 SEER patients and 69 patients from a Chinese validation set, all diagnosed with LAGC staged as cT1N2–3M0 or cT2–4N0/+M0 who underwent neoadjuvant therapy. A recent 2025 study further demonstrated that a LODDS-based nomogram achieved superior predictive performance for survival compared with pLN count in advanced GC, thereby supporting the rationale for incorporating LODDS into risk stratification (35). The prognostic significance of pLNs, LNR, as well as LODDS, was assessed in the context of curative gastrectomy following neoadjuvant treatment. Jiang et al.’s study in 2022 showed that among 148 LAGC patients receiving NACT, those with low LNR had markedly ameliorated OS and PFS compared with those having high LNR (16). Zhou et al.’s study in 2024 proved LNR as an independent OS risk factor in the LAGC population receiving neoadjuvant immunochemotherapy (17). These findings are further reinforced by a 2025 meta-analysis including over 10,000 patients, which identified LNR as a consistent and independent prognostic factor across diverse neoadjuvant treatment regimens (36). However, studies focused solely on LNR and did not perform direct comparisons with other LN indicators, like the number of pLNs or LODDS. The present analysis expands upon prior findings by systematically comparing the prognostic performance of all three nodal parameters, pLNs, LNR, and LODDS, across a large-scale development cohort and an independent external validation cohort.

In 2023, Que et al. proposed a fresh staging system termed ypTLoddsS (ypTLM) based on LODDS (37), which stratified patients into five stages: LODDS 1: <−4, LODDS 2: −4 to −2.5, LODDS 3: −2.5 to 0, and LODDS 4: >0. They concluded that this system outperformed the 8th AJCC TNM staging system regarding prediction accuracy (C-index: 0.663 vs. 0.633, P<0.001). Nonetheless, that study compared the new system solely with TNM staging and did not evaluate the prognostic utility of individual LN-related indicators. Moreover, the complexity of the LODDS-based staging system possibly impedes its clinical applicability. Notably, most patients (69.5%) in that cohort had not received NACT, thereby limiting generalizability to NACT-treated populations (38). A multicenter study conducted by Pengfei Gu in 2020, involving 7,620 Chinese GC patients who underwent gastrectomy, produced findings consistent with those of Que et al. In this study, LODDS cutoff values were categorized as follows: <−1.5, −1.5 to −1.0, −1.0 to −0.5, −0.5 to 0, and >0.

The present findings confirmed LODDS as an independent prognostic factor in LAGC. Within the SEER set, a LODDS value >−2.02 was markedly linked to less favorable OS (HR =2.84) and CSS (HR =3.16). In line with earlier reports, LODDS yielded the highest predictive accuracy among all nodal indicators, with C-indices of 0.671 and 0.690 for OS and CSS, outperforming LNR, pLNs, and conventional TNM staging. Furthermore, following PSM adjustment, patients categorized into the high-LODDS group exhibited markedly worse survival outcomes (P<0.001). Notably, in people who received only NACT, the discriminatory ability of LODDS improved further (OS: C-index =0.718; CSS: C-index =0.730), suggesting a greater capacity of LODDS to capture biological heterogeneity and treatment response in the absence of radiotherapy. Although LNR and pLNs retained prognostic relevance, their predictive performances consistently lagged behind that of LODDS across multiple statistical methodologies. In the external validation cohort from West China Hospital, LODDS supported high prognostic accuracy (OS: C-index =0.799; CSS: C-index =0.770), although it did not significantly outperform LNR or pLNs. The statistically insignificant disparities possibly arise from the limited samples and the high quality of lymphadenectomy within this cohort (mean number of ELNs >40), which possibly have mitigated the inherent limitations of the TNM N stage and pLN-based indicators. Nevertheless, multivariate Cox regression analysis identified LODDS >−2.02 as an independent adverse survival predictor, with particularly elevated HRs for OS and CSS.

LODDS achieves an optimal balance between the number of ELNs and pLNs, and possibly offers superior prognostic accuracy for patients with GC compared to other LN-related staging systems (39). In 2024, Liu et al. demonstrated that in 1,130 resectable gastroesophageal cancer individuals receiving NACT, the LODDS-based system exhibited enhanced prognostic performance [Akaike information criterion (AIC): 7,356.862] relative to the conventional TNM staging system (AIC: 7,385.357) (40). Similarly, in the context of curative gastrectomy, LODDS demonstrated superior predictive efficacy for survival in GC patients (C-index: 0.636, AIC: 4,304.0) compared to the LNR system (C-index: 0.631, AIC: 4225.8) and the node score system (C-index: 0.632, AIC: 4,308.4) (41). Our findings aligned with these observations.

The N category of the AJCC staging system is determined by the count of metastatic regional LNs (14). Accordingly, the relation of LODDS to survival outcomes was further analyzed among LAGC patients stratified by N stage. Stratified analyses demonstrated that LODDS was a significant predictor of survival in N1-N2 patients, outperforming pLNs and LNR; however, this prognostic advantage was not observed in N3 patients, who exhibited uniformly poor outcomes across all nodal metrics. LODDS showed optimal performance in the NACT-only cohort (OS: C-index =0.718; CSS: C-index =0.730), with findings remaining robust following PSM. Notably, this limitation is not specific to LODDS but rather reflects a broader constraint inherent to conventional lymph node-based prognostic indices in the context of N3 disease. N3 stage, defined by ≥7 metastatic lymph nodes, represents extremely high systemic tumor burden and inherently aggressive biology, with uniformly poor prognosis regardless of minor variations in nodal burden. Additionally, rapid recurrence and limited physiological reserve possibly reduce the benefit derived from postoperative treatments in this subgroup. These findings underscore the need for alternative systemic strategies, including immunotherapeutic approaches or intensified chemotherapy regimens, particularly for the high-risk population. Limited N3 individuals in this cohort (n=194, 9.82%) possibly have limited the statistical power of subgroup analyses, warranting further validation in larger, independent datasets.

The forecasting performance of pLNs, LNR, and LODDS was also evaluated specifically in LAGC patients undergoing NACT. A 2023 review by Li et al. concluded that LODDS is not only related to prognosis in GC but also demonstrates superior predictive ability for survival compared to pLNs and LNR (42). Our findings demonstrated the superior performance of LODDS for OS and CSS forecasting in this clinical context, aligning with prior literature. Nevertheless, considering the high risk and mortality related to GC, additional studies remain necessary to further refine survival prediction models.

Moreover, multiple investigations have reported that neoadjuvant therapy reduces LN yield in colorectal cancer, which possibly similarly compromises the prognostic utility of nodal metrics in GC, including LODDS, LNR, and pLNs (43-45). This effect is thought to result from mechanisms like cytotoxic lymphocyte depletion, radiation-induced stromal fibrosis, and adipocyte replacement, which collectively diminish LN detectability and obscure histopathological margins (46,47). These changes are particularly marked following radiotherapy, whereas chemotherapy exerts a more modest effect (48). Future research should emphasize the integration of molecular profiling technologies (e.g., circulating tumor DNA analysis) with advanced imaging modalities (e.g., dual-energy CT radiomics) to develop therapy-specific assessment protocols and refine postoperative adjuvant treatment strategies.

In summary, these findings highlight the necessity of revising existing LN staging frameworks in the context of neoadjuvant therapy. Static LN counts possibly no longer adequately reflect the underlying biological complexity or treatment response. By incorporating quantitative and proportional LN information, LODDS facilitates a more nuanced and reliable assessment of nodal burden, particularly among patients treated with NACT.

This study presented several important innovations. First, it represents the first systematic comparison of three LN assessment metrics, pLNs, LNR, and LODDS, in LAGC patients receiving neoadjuvant therapy, utilizing a large SEER development cohort and a rigorously curated Chinese validation cohort. Second, the analysis focused exclusively on patients who underwent NACT alone, thereby minimizing confounding effects from radiotherapy and more accurately capturing chemotherapy’s impact on nodal assessment. Third, a comprehensive statistical methodology was employed, incorporating multivariate Cox regression modeling, time-dependent receiver operating characteristic (ROC) curve analysis, C-index calculation, and PSM to guarantee the robustness of our conclusions. Fourth, stratified analysis by AJCC N stage demonstrated that LODDS maintained robust prognostic discrimination in N1-N2 disease, where traditional nodal metrics exhibited reduced effectiveness. Fifth, LODDS exhibited consistent prognostic performance irrespective of lymphadenectomy extent, affirming its clinical applicability in diverse surgical contexts. Sixth, the dichotomization of patients based on a LODDS threshold (≤−2.02 vs. >−2.02) simplified its clinical implementation. Finally, the inclusion of an external validation cohort from a high-volume tertiary center in Asia increases the generalizability of our findings and supports the incorporation of LODDS into future GC LN staging systems.

Limitations

Despite the strengths conferred by the multicenter design and rigorous statistical methodology, our limitations merit acknowledgment. First, although the SEER database provides comprehensive population-based data, it has inherent limitations in capturing granular clinical details. Specifically, SEER does not systematically record information on neoadjuvant chemotherapy regimens, treatment cycles, pathological response, surgical techniques, perioperative management, or postoperative complications. The absence of these variables may introduce residual clinical heterogeneity within the cohort and could modestly limit the generalizability of our findings to settings with standardized treatment protocols. Importantly, this limitation does not materially compromise our primary conclusion that LODDS serves as a superior prognostic marker in this population, as the results remained robust following propensity score matching and independent external validation. Second, essential prognostic factors like dietary habits and Helicobacter pylori infection status were not captured in the dataset, precluding their inclusion in the analysis. Third, the North American predominance of the SEER cohort may limit generalizability. While our Chinese validation cohort preliminarily supports these findings, larger prospective international multicenter studies are needed to validate the −2.02 cutoff for widespread clinical use. Fourth, for patients diagnosed in 2019 (representing 11.13% of the SEER set), the follow-up lasted less than 12 months, with a median duration of 22.0 (IQR, 11.0–48.0) months, possibly compromising the robustness of long-term survival estimations. Fifth, although LODDS demonstrated strong prognostic performance, the magnitude of improvement in the C-index was modest; therefore, its clinical relevance should be interpreted with caution. Moreover, the optimal cutoff value was statistically derived and may not be directly generalizable across heterogeneous clinical settings. Finally, the relatively small sample size of the validation cohort resulted in wide 95% CIs and limited statistical power to detect meaningful differences in prognostic performance among the nodal indicators. Seventh, as a retrospective investigation, the possibility of selection bias and residual confounding cannot be entirely excluded, despite the application of PSM.

Therefore, prospective studies with standardized data collection procedures are necessary to verify the foregoing findings and develop nodal assessment frameworks tailored to the neoadjuvant treatment context in LAGC.

Clinical implications

First, our findings support LODDS as a valuable adjunct to the AJCC TNM staging system, particularly for patients with N1–2 stage LAGC following neoadjuvant therapy; however, its clinical utility appears limited in N3 disease, likely due to the uniformly poor prognosis in this subgroup. From a practical perspective, LODDS can be calculated postoperatively, with a cutoff of −2.02 used for risk stratification, and patients in the high-risk category may warrant consideration for intensified adjuvant therapy and closer surveillance. Second, by incorporating both positive and nLNs, LODDS enhances risk discrimination in settings of suboptimal lymph node retrieval (<15 ELNs), which remains a common challenge in real-world clinical practice. Third, the incorporation of LODDS into postoperative decision-making processes possibly assists in identifying the high-risk population who benefits from intensified surveillance protocols or adjuvant treatment. Finally, the integration of LODDS into clinical staging guidelines has the possible to support more individualized and biologically informed therapeutic approaches in the management of GC.

Conclusions

This multicenter retrospective study demonstrates that LODDS is a robust supplementary prognostic indicator for patients with resectable LAGC undergoing neoadjuvant therapy followed by curative gastrectomy. In the SEER development cohort, LODDS outperformed pLNs, LNR, and AJCC TNM N staging, with preliminary validation provided by an independent Chinese cohort. Importantly, LODDS does not supplant the TNM staging system but offers incremental prognostic value, particularly in N1–2 disease and among patients receiving NACT alone. By integrating both positive and nLNs into a single metric, LODDS enables a more comprehensive and adaptable assessment of nodal status, especially in clinical contexts characterized by variable lymph node yield. These findings highlight the potential utility of incorporating LODDS into future staging frameworks and postoperative risk stratification models in GC. Nevertheless, further prospective, multicenter investigations are warranted to validate its clinical applicability and to clarify its role within standardized treatment algorithms.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was performed as per the protocol approved by the Medical Ethics Committee of West China Hospital, Sichuan University (No. 2024-1546). Since our analysis involved de-identified retrospective data, written informed consent was unnecessary. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Data Sharing Statement

Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-0430/dss

tcr-15-06-468-dss.pdf ^{(127.1KB, pdf)}

DOI: 10.21037/tcr-2026-0430