Examined lymph node count and survival in resected non-small cell lung cancer: attenuation of benefit beyond 16 nodes is confined to node-negative disease

Abstract

Background

Adequate lymph node (LN) examination during resection of non-small cell lung cancer (NSCLC) underpins accurate staging, but the number of examined lymph nodes (ELNs) beyond which additional retrieval confers no further benefit remains uncertain, and it is not known whether any single threshold applies across nodal stages. This study aims to systematically examine the association between ELN count, staging migration, and long-term survival to identify an evidence-based optimal ELN threshold.

Methods

We analysed 61,994 patients undergoing resection for NSCLC in the program (SEER) 17 registry (2010–2019). The detection of nodal metastasis was assessed as the proportion of node-positive cases and the mean number of positive nodes across ELN categories. Non-linearity was examined with restricted cubic splines. A threshold of 16 nodes, proposed in earlier work, was pre-specified and tested using a two-piecewise Cox model rather than derived from these data; the analysis was repeated across candidate breakpoints from 10 to 30 and separately by pathological N stage and extent of resection. Internal validation used split-sample and bootstrap resampling.

Results

The median ELN count was 9 (interquartile range, 5–15). Regression analyses included the 59,961 patients with 1 to 30 nodes examined, among whom 18,849 deaths occurred. After full adjustment, each increment of 5 ELNs was associated with higher odds of node-positive disease [odds ratio (OR) 1.19; 95% confidence interval (CI): 1.17–1.22] and lower mortality [hazard ratio (HR) 0.94; 95% CI: 0.92–0.95]. Below 16 nodes, each additional node was associated with a 2.1% reduction in the hazard of death (HR 0.979; 95% CI: 0.976–0.983; P<0.001), whereas above 16 no association was detectable (HR 1.001; 95% CI: 0.994–1.007; P=0.88); the two slopes differed significantly (P<0.001). No association above the breakpoint was found at any candidate value from 12 nodes upwards. On stratified analysis the two-piecewise model improved fit only in node-negative disease (P<0.001) and only after lobectomy (P<0.001); among node-positive patients the association persisted across the full range without a plateau (P=0.32).

Conclusions

In node-negative NSCLC treated by anatomical resection, the survival benefit of additional nodal retrieval attenuates beyond approximately 16 examined nodes, and this count may serve as a practical benchmark for auditing the adequacy of nodal examination. No comparable ceiling was identified in node-positive disease, in which more extensive retrieval remained associated with lower mortality. As an observational registry analysis, these findings describe associations rather than causal effects.

Highlight box.

Key findings

• In 61,994 patients undergoing resection for non-small cell lung cancer (NSCLC), the survival benefit of additional nodal retrieval attenuated beyond approximately 16 examined nodes, but only in node-negative disease and only after lobectomy. Among node-positive patients the association with mortality persisted across the full range of nodal counts without a plateau.

What is known and what is new?

• Previous registry analyses have proposed minimum nodal counts of between 11 and 21 nodes for resected NSCLC. These analyses were confined to node-negative patients, and the numerical basis for any single threshold has rarely been reported.

• We tested a pre-specified threshold of 16 nodes rather than deriving one, reported the numerical output of the threshold model together with a sensitivity analysis across alternative breakpoints, and showed that the attenuation of benefit does not extend to node-positive disease.

What is the implication, and what should change now?

• A benchmark of 16 examined nodes is a reasonable audit standard for node-negative disease after anatomical resection, but should not be applied as a sufficient standard in node-positive patients, in whom more extensive retrieval remained associated with lower mortality. Numerical benchmarks complement, but do not replace, station-based criteria for adequate mediastinal dissection.

Introduction

Lung cancer remains the leading cause of cancer-related mortality worldwide (1-3). The disease’s stage and prognosis are directly related to lymph node (LN) metastasis (4-6). For non-small cell lung cancer (NSCLC) treatment purposes, accurate LN assessment is essential (7,8). For instance, NSCLC patients with any sign of LN metastasis are recommended to undergo adjuvant chemotherapy (9-13). The efficacy of adjuvant chemotherapy has been well established in patients with LN positivity. Additionally, adjuvant therapy for any undetected residual disease helps lower the risk of recurrence and improve survival rates (10,11,14-17).

Despite this, there is significant variation in the guidelines for appropriate resection. In other surgical fields, the appropriate number of LNs to be removed during tumor resection is clear. For instance, the American Society of Colon and Rectal Surgeons (ASCRS) recommends 12 LNs as a prognostic link for colon cancer (18-23). The American Society of Breast Surgeons (ASBS) recommends the removal of 13 sentinel LNs, as per the American College of Surgeons Oncology Group Z0011 randomized controlled trial, which has become an accepted practice (24-27). However, currently, no consensus or prospective randomized controlled trial exists to determine the minimum number of examined LNs (ELNs) for lung cancer to optimize clinically significant survival differences (28,29).

To achieve a high-quality surveillance standard resection, the American College of Surgeons Commission on Cancer (ACS-CoC) recommends the removal of a minimum of 10 LNs during surgical resection for early-stage lung cancer (30). However, there are no specific guidelines indicating which 10 LNs need to be sampled for any given lobe (24). The minimum number of LNs in a sample comes from three database studies that used monitoring, epidemiological, and final outcome databases (SEER) and the National Cancer Database (NCDB). These studies oversimplified the number of removed LNs and categorized them instead of analyzing them as a single integer variable. Several studies have shown a correlation between the number of ELNs and long-term survival, but the results have been contradictory (31-38). According to a study based on data from the SEER program in the United States, there was a positive correlation between the number of ELNs and overall survival (OS) improvement. However, a study based on an Asian cohort showed the opposite trend (38). Many of these studies have limitations because the use of univariate analysis is subject to bias from other significant confounding factors. Moreover, no study has analyzed the role of ELNs separately based on histology. The different biological behaviors of different histological types may significantly impact the role of ELNs in prognosis. Furthermore, the cutoff point identification method is not convincing enough.

To address these knowledge gaps, we utilized the SEER database from the National Cancer Institute to systematically examine the association between ELN count, staging migration, and long-term survival in a large, contemporary cohort of surgically resected NSCLC patients, with the aim of identifying an evidence-based optimal ELN threshold. We present this article in accordance with the STROBE reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1437/rc).

Methods

Data source

We have obtained data from the latest SEER 17 registry study database, which was submitted in November 2021. This database comprises information from 2010 to 2019 and encompasses registries located in several regions, including San Francisco-Oakland, Connecticut, Hawaii, Iowa, New Mexico, Seattle (Puget Sound), Utah, Atlanta (Metropolitan), San Jose-Monterey, Los Angeles, Alaska Natives, Rural Georgia, California, Kentucky, Louisiana, New Jersey, and Greater Georgia. Collectively, the SEER 17 database covers approximately 26.5% of the entire United States population, as based on the 2010 Census. Because the study period spans three successive editions of tumour staging within SEER, stage was derived from the field applicable to each diagnosis year: the Derived AJCC 7th edition fields for 2010–2015, the Derived SEER Combined Stage fields for 2016–2017, and the Derived EOD 2018 fields for 2018–2019. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Patient selection

To generate the case list, we utilized SEER*Stat version 8.4.0 (available at SEER.cancer.gov/seerstat). Our selection criteria were limited to patients who had undergone surgical resection of first primary NSCLC and had at least one ELN. The number of examined and positive LNs was taken from the fields regional nodes examined (1988+) and regional nodes positive (1988+), and the extent of resection from RX Summ-Surg Prim Site (1998+). The case list contained variables such as sex, age, histological type, grade, stage, T-stage, N-stage, operation type, radiotherapy, and chemotherapy. Tumour grade was taken from the generic SEER differentiation fields (Grade thru 2017 and Grade Pathological 2018+), which record differentiation on a four-tier scale from well to undifferentiated; this coding is not equivalent to a histology-specific grading system such as the IASLC grading of invasive adenocarcinoma, and the small undifferentiated category should be interpreted accordingly. The recorded histologic types included squamous cell carcinoma (SCC), adenocarcinoma (ADC), adenosquamous carcinoma (ADSC), large-cell carcinoma, bronchoalveolar carcinoma (BAC), and other. Although the term bronchoalveolar carcinoma was superseded by the 2011 IASLC/ATS/ERS classification, the corresponding ICD-O-3 morphology codes (8250/3 and 8250/2) remained in use in SEER during part of the study period; the 188 tumours coded in this way were therefore retained as a separate category rather than reassigned. Operation type was classified as lobectomy, wedge resection, segmental resection, and pneumonectomy. Patients were excluded if they had not undergone curative-intent resection, if any of the study variables were missing or unknown, if the disease was stage IIIB or IV, or if the reported number of positive regional nodes was implausible under SEER coding conventions (codes 95, 97 and 99). The complete selection process is shown in Figure 1. Of 77,437 patients initially identified, 572 were excluded for absence of curative-intent resection, 14,798 for missing or unknown values, and 73 for implausible positive node counts, leaving a final eligible cohort of 61,994 patients. Because a complete-case approach was used, no missing values remained in the analytic dataset; the potential for selection bias introduced by this approach is addressed in the Discussion. As this was a retrospective study based on a national registry, the study size was determined by the total number of available cases in the SEER 17 database that met our strict eligibility criteria during the study period, without an a priori sample size calculation.

Figure 1.

Selection of the study population. ELN, examined lymph node; N, node; NSCLC, non-small cell lung cancer; SEER, Surveillance, Epidemiology, and End Results; T, tumor.

Statistical analysis

Descriptive statistics are based on the full eligible cohort (n=61,994). Because patients with more than 30 examined nodes were sparse (n=2,033; 3.3%) and yielded unstable estimates, all regression and threshold analyses were restricted to patients with 1 to 30 examined nodes (n=59,961, among whom 18,849 deaths occurred). We categorized the ELN into six groups ranging from 1–5, 6–10, 11–15, 16–20, 21–25, and 26–30; intervals of five nodes were used to preserve comparability with previous SEER-based analyses of this question. We utilized the proportion of node-positive cases (N1 or N2 disease) and the mean number of positive LNs at different ELNs to evaluate the detection of nodal metastasis. We used this term in preference to stage migration, since the present design cannot demonstrate reclassification of individual patients between stages and can only quantify how the detected burden of nodal disease varies with the extent of nodal examination. To compare the relationship between ELN and the detection of nodal metastasis as well as ELN and mortality, we conducted a linear trend test by including the median ELN in each group as a continuous variable in the model. Associations with the detection of nodal metastasis were estimated by logistic regression and expressed as odds ratios; associations with overall mortality were estimated by Cox proportional hazards regression and expressed as hazard ratios. In the subgroup analysis, we examined the relationship between ELN and detection of nodal metastasis and ELN and mortality in different histological types of the population. Non-linearity in the relationship between ELN and each outcome was examined using restricted cubic splines with five knots placed at the 5th, 27.5th, 50th, 72.5th and 95th percentiles of the ELN distribution, fitted with the rms package and adjusted for all covariates listed below. The threshold of 16 examined nodes proposed in a previous SEER-based analysis was pre-specified as the candidate breakpoint. We did not derive this value from the present data; rather, we tested formally whether the association between ELN and each outcome differed above and below it, using a two-piecewise regression model in which separate slopes were estimated on either side of the breakpoint. The two-piecewise model was compared with a single linear model by log-likelihood ratio test. To establish whether our conclusions depended on this particular value, the same analysis was repeated across all candidate breakpoints from 10 to 30. All fully adjusted models included age, sex, primary site, grade, histology, T stage, operation type, radiotherapy and chemotherapy; the mortality models additionally included pathological N stage. Because nodal upstaging is a plausible mediator of the association between ELN and survival, adjustment for N stage removes part of the pathway of interest and yields a conservative estimate; a model omitting N stage was therefore fitted as a sensitivity analysis. We plotted the smoothed fit curves of the relationship between ELN and the detection of nodal metastasis and mortality in different populations stratified based on covariates such as age, sex, histology, grade, primary site, stage, T-stage, N-stage, operation type, radiotherapy, and chemotherapy.

To estimate the accuracy with which the number of positive nodes is reported, we developed a mathematical model of the number of examined nodes using hypergeometric distribution and Bayes’ theorem based on the method described by Iyer et al. (39). For a patient in whom m positive nodes are observed among n examined, the probability of harbouring M truly positive nodes was obtained as$r(M,n,m)=q(M)\times P(n,m,N,M)$ divided by the sum of $q(M)\times P(n,m,N,M)$ over M, where $P(n,m,N,M)$ is the hypergeometric probability and N denotes the assumed total number of regional nodes. Whereas the original model was developed for the axilla and assumed a total of 21 nodes, we assumed a total of 65 regional nodes to reflect the substantially larger number of intrathoracic nodal stations; this value was adopted as a working assumption and was not derived from a published anatomical series. The prior distribution q(M) was derived empirically from the observed distribution of positive node counts in the present cohort rather than assumed. We utilized the resulting model to estimate the probability of having at least one undetected positive LN in patients with node-negative disease with different observed sample sizes. To determine the effect of the number of ELNs on OS and to visualize survival curves, we utilized Cox proportional risk regression models, with the Breslow method for tied event times. Survival between patients with fewer than 16 and at least 16 examined nodes was compared by the log-rank test. Because the benefit of extended nodal dissection may differ according to nodal status and the extent of resection, the threshold analysis was repeated separately in patients with pN0, pN1 and pN2 disease, in patients with node-positive disease combined, and in patients undergoing lobectomy, sublobar resection and pneumonectomy; in these stratified models the stratifying variable was omitted from the covariate set. In the absence of an independent external cohort, internal validation was performed in two ways: the cohort was randomly divided into two halves and the analysis repeated in each, and the effect estimates on either side of the breakpoint were re-estimated in 200 bootstrap resamples of the full analysis population. All analyses were conducted using Empower(R) (www.enpowerstats.com, X&Y solutions, Inc. Boston, MA, United States) and R version 4.2.0 (http://www.R-project.org). A two-sided P value of less than 0.05 was considered statistically significant. Furthermore, the step-wise inclusion of covariates across the three models (Model I, II, and III) served as sensitivity analyses to test the robustness of our findings regarding the association between ELN and the detection of nodal metastasis or mortality.

Results

Baseline characteristics of the study population, by ELN category

This study included a total of 61,994 patients who met the eligibility criteria. The baseline characteristics of the study cohort are presented in Table 1. The median ELN count was 9 (interquartile range, 5–15). The majority of patients (approximately 81.96%) underwent lobectomy, while a smaller proportion (approximately 10.34%) underwent wedge resection. The percentage of patients with N0, N1, and N2 stages was 83.84%, 12.76%, and 3.40%, respectively. Because cases with missing essential clinical or pathological variables were excluded during the initial patient selection phase, there were no missing data for the variables of interest in the final analytical cohort. Baseline characteristics differed systematically across ELN categories (Table 1). With increasing ELN, patients were more often male (44.3% to 54.2%), more often diagnosed in the later period (45.0% to 61.8%), and more often had node-positive (9.1% to 30.6%) and T3-T4 (11.6% to 21.4%) disease. Pneumonectomy became progressively more frequent (1.0% to 8.8%), as did receipt of chemotherapy (13.8% to 32.3%). All comparisons were significant at P<0.001. Patients in whom more nodes were examined therefore represent a group with more advanced disease undergoing more extensive surgery, and this imbalance is considered further in the Discussion. Regression and threshold analyses were restricted to the 59,961 patients with 1 to 30 examined nodes (96.7% of the cohort), among whom 18,849 deaths were recorded over a mean follow-up of 43 months.

Table 1. Baseline characteristics of the study population according to examined lymph node category.

Characteristic	ELN 1–5	ELN 6–10	ELN 11–15	ELN 16–20	ELN 21–25	ELN 26–30	ELN >30	P
No. of patients	17,998	18,089	11,920	6,599	3,471	1,818	2,099
Age								<0.001
<65 years	5,662 (31.5)	6,137 (33.9)	4,024 (33.8)	2,150 (32.6)	1,068 (30.8)	540 (29.7)	615 (29.3)
≥65 years	12,336 (68.5)	11,952 (66.1)	7,896 (66.2)	4,449 (67.4)	2,403 (69.2)	1,278 (70.3)	1,484 (70.7)
Sex								<0.001
Male	7,979 (44.3)	8,124 (44.9)	5,491 (46.1)	3,048 (46.2)	1,736 (50.0)	915 (50.3)	1,138 (54.2)
Female	10,019 (55.7)	9,965 (55.1)	6,429 (53.9)	3,551 (53.8)	1,735 (50.0)	903 (49.7)	961 (45.8)
Year of diagnosis								<0.001
2010–2014	9,897 (55.0)	8,903 (49.2)	5,105 (42.8)	2,662 (40.3)	1,348 (38.8)	705 (38.8)	802 (38.2)
2015–2019	8,101 (45.0)	9,186 (50.8)	6,815 (57.2)	3,937 (59.7)	2,123 (61.2)	1,113 (61.2)	1,297 (61.8)
Primary site								<0.001
Upper lobe	9,904 (55.0)	10,500 (58.0)	7,015 (58.9)	3,852 (58.4)	2,009 (57.9)	1,048 (57.6)	1,195 (56.9)
Middle lobe	1,438 (8.0)	981 (5.4)	524 (4.4)	265 (4.0)	141 (4.1)	89 (4.9)	100 (4.8)
Lower lobe	6,263 (34.8)	6,124 (33.9)	4,042 (33.9)	2,229 (33.8)	1,174 (33.8)	606 (33.3)	703 (33.5)
Main bronchus	33 (0.2)	77 (0.4)	60 (0.5)	47 (0.7)	36 (1.0)	20 (1.1)	30 (1.4)
Other	360 (2.0)	407 (2.2)	279 (2.3)	206 (3.1)	111 (3.2)	55 (3.0)	71 (3.4)
Grade								<0.001
I	4,265 (23.7)	3,879 (21.4)	2,297 (19.3)	1,220 (18.5)	610 (17.6)	291 (16.0)	305 (14.5)
II	8,340 (46.3)	8,499 (47.0)	5,723 (48.0)	3,099 (47.0)	1,632 (47.0)	877 (48.2)	989 (47.1)
III	5,192 (28.8)	5,472 (30.3)	3,756 (31.5)	2,188 (33.2)	1,181 (34.0)	629 (34.6)	780 (37.2)
IV	201 (1.1)	239 (1.3)	144 (1.2)	92 (1.4)	48 (1.4)	21 (1.2)	25 (1.2)
Histology								<0.001
SCC	3,941 (21.9)	4,214 (23.3)	2,939 (24.7)	1,731 (26.2)	916 (26.4)	509 (28.0)	649 (30.9)
ADC	9,901 (55.0)	9,660 (53.4)	6,245 (52.4)	3,294 (49.9)	1,724 (49.7)	898 (49.4)	968 (46.1)
ADSC	396 (2.2)	423 (2.3)	250 (2.1)	143 (2.2)	75 (2.2)	40 (2.2)	44 (2.1)
Large-cell carcinoma	95 (0.5)	93 (0.5)	58 (0.5)	48 (0.7)	25 (0.7)	9 (0.5)	9 (0.4)
BAC	79 (0.4)	63 (0.3)	23 (0.2)	12 (0.2)	8 (0.2)	3 (0.2)	0 (0.0)
Other	3,586 (19.9)	3,636 (20.1)	2,405 (20.2)	1,371 (20.8)	723 (20.8)	359 (19.7)	429 (20.4)
T stage								<0.001
T1	10,213 (56.7)	9,306 (51.4)	5,711 (47.9)	3,001 (45.5)	1,520 (43.8)	712 (39.2)	789 (37.6)
T2	5,691 (31.6)	6,395 (35.4)	4,504 (37.8)	2,527 (38.3)	1,325 (38.2)	733 (40.3)	860 (41.0)
T3	1,604 (8.9)	1,895 (10.5)	1,360 (11.4)	834 (12.6)	486 (14.0)	282 (15.5)	339 (16.2)
T4	490 (2.7)	493 (2.7)	345 (2.9)	237 (3.6)	140 (4.0)	91 (5.0)	111 (5.3)
N stage								<0.001
N0	16,360 (90.9)	15,433 (85.3)	9,630 (80.8)	5,146 (78.0)	2,610 (75.2)	1,339 (73.7)	1,456 (69.4)
N1	1,282 (7.1)	2,113 (11.7)	1,807 (15.2)	1,149 (17.4)	684 (19.7)	384 (21.1)	492 (23.4)
N2	356 (2.0)	543 (3.0)	483 (4.1)	304 (4.6)	177 (5.1)	95 (5.2)	151 (7.2)
Operation type								<0.001
Lobectomy	12,361 (68.7)	15,574 (86.1)	10,571 (88.7)	5,827 (88.3)	3,069 (88.4)	1,576 (86.7)	1,833 (87.3)
Wedge resection	4,131 (23.0)	1,315 (7.3)	512 (4.3)	241 (3.7)	91 (2.6)	63 (3.5)	55 (2.6)
Segmental resection	1,324 (7.4)	771 (4.3)	400 (3.4)	177 (2.7)	66 (1.9)	36 (2.0)	27 (1.3)
Pneumonectomy	182 (1.0)	429 (2.4)	437 (3.7)	354 (5.4)	245 (7.1)	143 (7.9)	184 (8.8)
Radiotherapy								0.08
No	16,970 (94.3)	17,105 (94.6)	11,220 (94.1)	6,235 (94.5)	3,253 (93.7)	1,727 (95.0)	1,958 (93.3)
Yes	1,028 (5.7)	984 (5.4)	700 (5.9)	364 (5.5)	218 (6.3)	91 (5.0)	141 (6.7)
Chemotherapy								<0.001
No	15,515 (86.2)	14,684 (81.2)	9,207 (77.2)	4,924 (74.6)	2,498 (72.0)	1,282 (70.5)	1,422 (67.7)
Yes	2,483 (13.8)	3,405 (18.8)	2,713 (22.8)	1,675 (25.4)	973 (28.0)	536 (29.5)	677 (32.3)
Positive lymph nodes	0 (0–0)	0 (0–0)	0 (0–0)	0 (0–0)	0 (0–0)	0 (0–1)	0 (0–1)	<0.001

Data are presented as n (%) or median (interquartile range). ADC, adenocarcinoma; ADSC, adenosquamous carcinoma; BAC, bronchoalveolar carcinoma; ELN, examined lymph node; N, node; SCC, squamous cell carcinoma; T, tumor.

This study included a total of 61,994 patients who met the eligibility criteria. The baseline characteristics of the study cohort are presented in Table 1. The median ELN count was 9 (interquartile range, 5–15; Table 1). The majority of patients (approximately 81.96%) underwent lobectomy, while a smaller proportion (approximately 10.34%) underwent wedge resection. The percentage of patients with N0, N1, and N2 stages was 83.84%, 12.76%, and 3.40%, respectively. Because cases with missing essential clinical or pathological variables were excluded during the initial patient selection phase, there were no missing data for the variables of interest in the final analytical cohort.

Association between ELN and the detection of nodal metastasis

In a model fully adjusted for covariates, we found a significant positive association between increasing ELN counts and the proportion of node-positive patients (P for trend <0.001). Using the subgroup with ELN of 1-5 as a reference, the odds of node-positive disease were 1.41 [95% confidence interval (CI): 1.30, 1.52] in the subgroup with ELN of 6–10, 1.75 (95% CI: 1.61, 1.89) for ELN 11–15, 1.96 (95% CI: 1.79, 2.15) for ELN 16–20, 2.11 (95% CI: 1.89, 2.35) for ELN 21–25, and 2.31 (95% CI: 2.01, 2.65) for ELN 26–30 (all P<0.001). For each 5 increase in ELN, the odds of node-positive disease increased by a factor of 1.19 (95% CI: 1.17, 1.22). The 3 models adjusted for different covariates showed a robust positive association of ELN with the detection of nodal metastasis. The number of patients and events in each category and the specific adjusted variables are detailed in Table 2.

Table 2. Association between examined lymph node count, detection of nodal metastasis, and overall mortality.

ELN	N	Events	Model I, HR (95% CI)	Model II, HR (95% CI)	Model III, HR (95% CI)
Detection of nodal metastasis
1–5	18,013	1,648	Reference	Reference	Reference
6–10	18,112	2,672	1.72 (1.61, 1.83)	1.71 (1.60, 1.82)	1.41 (1.30, 1.52)
11–15	11,929	2,297	2.37 (2.21, 2.53)	2.35 (2.19, 2.52)	1.75 (1.61, 1.89)
16–20	6,609	1,460	2.82 (2.61, 3.04)	2.81 (2.60, 3.03)	1.96 (1.79, 2.15)
21–25	3,473	863	3.28 (2.99, 3.60)	3.26 (2.97, 3.57)	2.11 (1.89, 2.35)
26–30	1,825	486	3.60 (3.21, 4.05)	3.59 (3.20, 4.03)	2.31 (2.01, 2.65)
Per 5-node increase			1.32 (1.30, 1.34)	1.32 (1.29, 1.34)	1.19 (1.17, 1.22)
P for trend			<0.001	<0.001	<0.001
Overall mortality
1–5	18,013	6,406	Reference	Reference	Reference
6–10	18,112	5,684	0.92 (0.88, 0.95)	0.92 (0.89, 0.95)	0.89 (0.86, 0.92)
11–15	11,929	3,414	0.89 (0.85, 0.93)	0.88 (0.85, 0.92)	0.81 (0.78, 0.85)
16–20	6,609	1,863	0.91 (0.87, 0.96)	0.91 (0.86, 0.96)	0.81 (0.77, 0.85)
21–25	3,473	952	0.88 (0.83, 0.95)	0.85 (0.80, 0.91)	0.75 (0.70, 0.81)
26–30	1,825	530	0.96 (0.88, 1.05)	0.91 (0.84, 1.00)	0.76 (0.70, 0.84)
Per 5-node increase			0.98 (0.96, 0.99)	0.97 (0.96, 0.98)	0.94 (0.92, 0.95)
P for trend			<0.001	<0.001	<0.001

Model I adjusted for: None. Model II adjusted for: age, sex. Model III adjusted for: age, sex, primary site, grade, histology, T stage, operation type, radiotherapy, chemotherapy. CI, confidence interval; ELN, examined lymph node; HR, hazard ratio; T, tumor.

Association between ELN and overall mortality

A significant inverse association between increasing ELN and total mortality was present in all three models (P for trend <0.001 throughout). Using the subgroup with ELN 1–5 as a reference, total mortality was 0.89 (95% CI: 0.86, 0.92) in the subgroup with ELN 6–10, 0.81 (95% CI: 0.78, 0.85) in the subgroup with ELN 11–15, 0.81 (95% CI: 0.77, 0.85) in the subgroup with ELN 16–20, 0.75 (95% CI: 0.70, 0.81) in the subgroup with ELN 21–25, and 0.76 (95% CI: 0.70, 0.84) for the ELN 26–30 group (all P<0.001). For every 5 increase in ELN, the adjusted hazard of death was reduced by 6% (HR 0.94, 95% CI: 0.92, 0.95). Adjusted mortality was lowest in the 21–25 category rather than declining monotonically, and the estimate for the 26–30 category, based on 1,825 patients, was slightly higher; these categories are the smallest in the cohort and the corresponding estimates are correspondingly less precise. The specific covariates are detailed in Table 2.

Histological subgroup analyses

The results showed that a higher ELN was associated with an increased likelihood of node-positive disease in both squamous and adenocarcinoma (Table S1). For each 5 increase in ELN, the odds of node-positive disease increased by a mean of 1.28 (95% CI: 1.23, 1.33) in squamous carcinoma and 1.15 (95% CI: 1.12, 1.19) in adenocarcinoma. Additionally, ELN showed a significant inverse association with total mortality in both squamous and adenocarcinoma histological subgroups (Table S2). However, when ELN reached 26–30, total mortality was slightly higher than in the preceding category. These findings indicate that the extent of nodal examination is associated with both the detected burden of nodal disease and subsequent mortality across the two principal histological subtypes.

Association between ELN and the detected burden of nodal disease

The study conducted a smoothed curve fit between ELN and the proportion of node-positive cases, revealing a non-linear relationship in which the proportion rose steeply up to approximately 16 examined nodes and increased only gradually thereafter (Figure 2A). This relationship was consistent across different strata based on age, sex, primary site, histology, grade, stage, T-stage, and operation type mode for eight covariates (Figure S1).

Figure 2.

Adjusted associations between examined lymph node count and nodal and survival outcomes. (A) Proportion of node-positive (N1 or N2) cases. (B) Mean number of positive lymph nodes. (C) Probability of harbouring undetected positive lymph nodes, estimated from a hypergeometric Bayesian model. (D) Log hazard ratio for overall mortality. Curves were fitted with restricted cubic splines using five knots; dashed lines denote 95% confidence intervals. The vertical line marks the pre-specified breakpoint of 16 nodes. Adjusted for age, sex, primary site, grade, histology, T stage, N stage, operation type, radiotherapy and chemotherapy. Analysis population: patients with 1 to 30 examined nodes (n=59,961). ELN, examined lymph node.

The study also analyzed the relationship between ELN and the mean number of positive LNs, which increased progressively across the examined range (Figure 2B). This pattern was consistent across different strata based on age, sex, primary site, histology, grading, staging, T-staging, and surgical modality for eight covariates (Figure S2). Lastly, the study analyzed the probability of having more positive nodes than observed and found a negative correlation between ELNs and probabilities. As ELNs increase, the probability of having more positive nodes than observed decreases from 0.99 at a single node examined to 0.79 at 16 nodes and 0.61 at 30 nodes (Figure 2C), indicating that more ELNs were associated with a lower probability of occult nodal metastasis.

Attenuation of the mortality association beyond the pre-specified breakpoint

To demonstrate the correlation between ELN and mortality, we utilized various analytical techniques, including smoothed curve fitting. The adjusted smoothed fit curve displayed a significant decrease in mortality with increasing ELN when the value was less than 16. However, when the ELN exceeded 16, mortality no longer displayed any significant increase or decrease and the curve reached a plateau (Figure 2D). This pattern was confirmed in the formal threshold analysis (Table 3). Below 16 examined nodes each additional node was associated with a 2.1% reduction in the hazard of death (HR 0.979, 95% CI: 0.976–0.983, P<0.001), whereas above 16 no association was detectable (HR 1.001, 95% CI: 0.994–1.007, P=0.88). The two slopes differed significantly (ratio 1.021, 95% CI: 1.012–1.031, P<0.001), and the two-piecewise model fitted significantly better than a single linear model (P<0.001). Because 16 was pre-specified rather than derived from these data, the analysis was repeated across candidate breakpoints from 10 to 30 (Table S3, Figure 3): a significant association above the breakpoint persisted at 10 and 11 nodes but was absent at every value from 12 upwards, indicating that the attenuation of marginal benefit is a feature of the data rather than an artefact of the particular value chosen. Findings were unchanged when pathological N stage was omitted from the model (Table S4). Internal validation supported these estimates (Table S5). In two random halves of the cohort the association below the breakpoint remained significant (HR 0.981 and 0.978, both P<0.001) with no association above it, and the two-piecewise model fitted better than the linear model in both (P=0.001). Across 200 bootstrap resamples the hazard ratio below the breakpoint was below unity in every resample, whereas above the breakpoint it exceeded unity in 51.5%, consistent with a null association. For a more nuanced analysis, we generated smoothed fitted curves based on several variables, including age, sex, histology, grade, T-stage, N-stage, operation type, radiotherapy, and chemotherapy, among the different strata of the 9 covariates (Figure S3). In most populations, the decreasing trend of mortality slowed down and plateaued when the ELN reached approximately 16. However, in populations with T1, T4, undergoing lung wedge resection, receiving radiotherapy, and receiving chemotherapy, mortality continued to decrease with increasing ELN, and there was no plateauing period (Figure S3).

Table 3. Threshold effect analysis of examined lymph node count on overall mortality.

Exposure: ELN (per 1-node increase)	HR (95% CI)	P value	P for LRT
Model I—single linear model
Across full ELN range	0.986 (0.983, 0.988)	<0.001
Model II—two-piecewise model (pre-specified K=16)
ELN <16	0.979 (0.976, 0.983)	<0.001
ELN ≥16	1.001 (0.994, 1.007)	0.88
Difference (above vs. below)	1.021 (1.012, 1.031)	<0.001	<0.001

Analysis population: patients with ELN 1–30 (n=59,961; 18,849 deaths). Adjusted for age, sex, histology, T stage, N stage, operation type, grade, radiotherapy, chemotherapy, and primary site. The breakpoint of 16 examined lymph nodes was pre-specified on the basis of the threshold previously proposed by Liang et al. (14) rather than derived from the present data. Model II was formally compared with Model I using a log-likelihood ratio test. HR, hazard ratio; CI, confidence interval; ELN, examined lymph node; LRT, log-likelihood ratio test; N, node; T, tumor.

Figure 3.

Sensitivity of the threshold analysis to the choice of breakpoint. Each point shows the hazard ratio per additional examined node above the candidate breakpoint, with 95% confidence intervals. Red denotes P<0.05 and blue P≥0.05. No significant association above the breakpoint was detected at any value from 12 nodes upwards. Adjusted for age, sex, primary site, grade, histology, T stage, N stage, operation type, radiotherapy and chemotherapy. Analysis population: patients with 1 to 30 examined nodes (n=59,961). CI, confidence interval; ELN, examined lymph node; HR, hazard ratio; N, node; T, tumor.

To examine the breakpoint in the relationship between ELN and mortality, we divided the population into two groups based on the breakpoint, the ELN <16 group and the ELN ≥16 group. In the overall cohort, OS did not differ significantly between the two groups (log-rank P=0.89; Figure 4A). Within nodal strata, however, survival was significantly better in the ELN ≥16 group, both in pN0 disease (5-year survival 69.3% versus 67.2%; log-rank P<0.001) and in pN1 disease (51.5% versus 46.5%; log-rank P=0.002; Table S6) (Figure 4B,4C). The discrepancy between the pooled and stratified comparisons is consistent with a shift in case mix: because patients with more nodes examined were more frequently classified as node-positive, the pooled comparison combines groups of differing nodal composition. Patients with pN2 disease were not included in this comparison, since 99% of that subgroup was diagnosed in 2015-2019 with a median follow-up of 23 months and only 1.2% followed beyond 48 months, leaving the stratum underpowered for survival comparison.

Figure 4.

Kaplan-Meier overall survival by examined lymph node count. (A) Full eligible cohort. (B) Node-negative (pN0) disease. (C) pN1 disease. Numbers at risk are shown beneath each panel and P values were obtained by log-rank test. Patients with pN2 disease are not shown, since 99% of that subgroup was diagnosed in 2015–2019 with a median follow-up of 23 months, leaving the stratum underpowered for long-term comparison. Analysis population: full eligible cohort (n=61,994). ELN, examined lymph node.

Discussion

In this population-based cohort of 61,994 patients undergoing resection for NSCLC, a more extensive nodal examination was associated with a higher proportion of node-positive cases and with lower overall mortality. Both associations were attenuated beyond the pre-specified threshold of 16 examined nodes, and this attenuation was robust across candidate breakpoints from 12 to 30 and across split-sample and bootstrap validation. The principal new observation, however, concerns where this threshold applies. When the analysis was repeated within nodal strata, a distinct breakpoint was demonstrable only in node-negative disease: among pN0 patients the two-piecewise model fitted significantly better than a single linear model (P<0.001), whereas among node-positive patients it did not (P=0.32; Table 4), the association with mortality persisting across the full range of nodal counts without a plateau. A parallel pattern was seen by extent of resection, the threshold being evident after lobectomy (P<0.001) but not after sublobar resection (P=0.22; Table 5). The benchmark of 16 nodes therefore appears to describe node-negative patients undergoing standard anatomical resection, and should not be applied uniformly across all patients. The correlation between ELN and OS was particularly evident in the histological populations of squamous and adenocarcinoma but not consistent in other histological types, which may be due to the small number of cases. Several previous studies have proposed minimum nodal counts for resected NSCLC. Osarogiagbon et al. reported maximal reduction in long-term mortality risk at 18 to 21 nodes in pathologically node-negative disease, and Liang et al., in a combined analysis of SEER and a Chinese multi-institutional registry, proposed a threshold of 16 nodes; earlier work by Ou et al. suggested benefit up to 11 to 15 nodes. We did not set out to derive yet another candidate value, and we regard the reproduction of a threshold close to those already reported as confirmatory rather than novel. The present analysis extends this literature in three respects. First, the earlier threshold analyses were restricted to node-negative patients; by including node-positive disease we are able to show that the plateau does not generalise, and that in pN1 and pN2 disease the association between nodal count and mortality continues without evidence of attenuation. Second, we report the numerical output of the threshold model rather than relying on the visual appearance of a smoothed curve, together with a sensitivity analysis across alternative breakpoints, which allows the reader to judge how firmly any single value is supported. Third, by examining the extent of resection separately we show that the threshold is a property of standard anatomical resection rather than of nodal examination in general.

Table 4. Threshold effect analysis stratified by pathological N stage.

Stratum	N	Deaths	HR below 16 (95% CI)	HR above 16 (95% CI)	P for LRT
pN0	50,535	14,829	0.978 (0.975, 0.982)*	1.005 (0.997, 1.013)	<0.0001
pN1	7,455	3,466	0.981 (0.972, 0.989)*	0.994 (0.981, 1.007)	0.16
pN2	1,971	554	1.007 (0.984, 1.029)	0.981 (0.949, 1.014)	0.30
pN-positive (N1+N2)	9,426	4,020	0.984 (0.976, 0.991)*	0.992 (0.980, 1.005)	0.32

*, P<0.0001. Hazard ratios represent the change in mortality risk per one-node increase in examined lymph nodes below and above the pre-specified breakpoint of 16. Within each stratum, N stage was omitted from the covariate set; all other covariates were as in Table 3. A significant improvement in model fit with the two-piecewise specification was observed only among pN0 patients, indicating that the attenuation of marginal benefit beyond 16 nodes applies to node-negative disease. Among pN-positive patients the association with mortality persisted across the full range of examined lymph node counts without evidence of a plateau. CI, confidence interval; ELN, examined lymph node; HR, hazard ratio; LRT, log-likelihood ratio test; N, node.

Table 5. Threshold effect analysis stratified by extent of surgical resection.

Surgical procedure	N	Deaths	HR below 16 (95% CI)	HR above 16 (95% CI)	P for LRT
Lobectomy	49,031	15,016	0.979 (0.975, 0.983)*	1.001 (0.994, 1.009)	<0.001
Sublobar resection†	9,131	2,902	0.978 (0.969, 0.988)*	0.999 (0.972, 1.028)	0.22
Pneumonectomy	1,799	931	0.986 (0.968, 1.004)	0.997 (0.976, 1.018)	0.52

*, P<0.0001. ^†, wedge and segmental resection combined. Within each stratum, operation type was omitted from the covariate set; all other covariates were as in Table 3. Analysis population: ELN 1–30. Significant improvement in fit with the two-piecewise specification was confined to patients undergoing lobectomy, the procedure accounting for 82% of the cohort. Among patients undergoing sublobar resection the association with mortality was significant across the full range without a detectable plateau, and no association was observed after pneumonectomy. CI, confidence interval; ELN, examined lymph node; HR, hazard ratio; LRT, log-likelihood ratio test.

Several mechanisms may underlie the observed association, and the present design cannot distinguish between them. A greater number of ELNs increases the probability of detecting positive nodes; comprehensive nodal assessment, coupled with appropriate adjuvant therapy, may reduce the likelihood of undetected residual disease (40). A more thorough dissection may also remove micrometastatic deposits that would otherwise remain in situ. Equally, however, the number of nodes retrieved reflects the conduct of the operation, the diligence of pathological processing, and institutional practice, and may therefore act in part as a marker of overall quality of care rather than as an intervention in its own right. Table 1 shows that patients in whom more nodes were examined differed systematically from those in whom fewer were examined, and residual confounding by unmeasured factors such as surgeon experience, institutional volume and intraoperative judgement cannot be excluded. One observation bears on this question. When pathological N stage was omitted from the model, the association below the breakpoint was slightly weaker rather than stronger (HR 0.983 versus 0.979 per node; Table S4). Had the survival advantage arisen solely through improved staging and the consequent administration of adjuvant therapy, removing N stage from the model would have been expected to strengthen the estimate. That it did not suggests that the association is not attributable to nodal upstaging alone, although it does not establish a direct therapeutic effect of dissection itself.

A numerical benchmark of this kind must be interpreted alongside the anatomical basis of nodal staging. Determining the number of LNs that should be removed during therapeutic surgery for lung cancer has been a challenging task. The current NCCN guidelines for NSCLC do not offer recommendations regarding ELN count, and the latest 8th edition of this classification only modified the T and M descriptors, while the N descriptor has remained unchanged since the 6th edition (41). The nodal grouping’s benefit lies in its simple and uncomplicated qualification system, making it easily manageable. Nevertheless, this system overlooks the quantitative aspects of LN metastasis and the significance of other phenomena that contribute significantly to LN subgroup heterogeneity (42). Previous studies have proposed differing benchmarks, ranging from 11 to 21 nodes (14). Conversely, a count alone disregards anatomy. The IASLC nodal map defines the stations that constitute an adequate dissection, and current surgical recommendations are framed in terms of station coverage, sampling of the hilar station together with a specified number of mediastinal stations, rather than in terms of a total count. Twenty nodes retrieved from a single station do not constitute the same operation as twelve nodes drawn from several stations, yet the two are indistinguishable in registry data. A comparable observation has been made in other solid tumours, where both the number of nodes removed and the anatomical template of the dissection were independently associated with cancer-specific survival (32,33,43). A numerical benchmark should therefore be regarded as a complement to station-based criteria and not as a substitute for them; it is best suited to auditing the adequacy of nodal examination at a population level, where station-level detail is rarely recorded. Within these limits, our data indicate that the survival benefit was concentrated from 1 to 16 ELNs, with little additional benefit detectable beyond that point in node-negative disease. Identifying additional positive LNs does not alter treatment strategies since adjuvant chemotherapy is routinely administered to LN-positive patients who can tolerate it (14). This may partly explain why the association attenuates in node-negative patients while persisting in those already known to be node-positive, in whom further nodal retrieval may continue to remove residual disease.

Taken together, these findings suggest that 16 examined nodes may serve as a practical benchmark for auditing the adequacy of nodal examination and for postoperative prognostic stratification in patients with node-negative disease undergoing anatomical resection. In node-positive disease our data provide no basis for a comparable ceiling, and a benchmark framed as a minimum standard rather than as a sufficient one would be more appropriate for such patients. This has the potential to support greater consensus between surgeons and pathologists regarding the degree of completeness of resection (44). Strict and objective inclusion criteria were applied, and fully adjusted multivariable models incorporating demographic, pathological and treatment covariates were used throughout; these measures reduce but do not eliminate confounding in an observational registry analysis.

Despite the strengths of our study, several limitations must be acknowledged. As a retrospective study, our results are subject to potential biases. The most important constraint is that SEER records only the total number of nodes examined; the surgical technique employed is likewise unrecorded (40). The anatomical stations from which nodes were retrieved, whether a systematic dissection or sampling was performed, and whether specimens were fragmented during handling are all unrecorded (45). Fragmentation in particular inflates the count without any corresponding increase in the tissue examined. A nodal count is therefore an imperfect proxy for the completeness of dissection, and this is a limitation of the measure itself rather than a technical detail of the present analysis. Second, approximately 19% of initially identified patients were excluded because one or more study variables were missing. Registry completeness is unlikely to be independent of institutional characteristics, and the analysed cohort may over-represent centres with more complete reporting. Third, the study period spans three staging systems, and the criteria assigning N2 disease to stage IIIA are not identical between them. The consequence is visible in our data: 99% of pN2 patients were diagnosed in 2015-2019, with a median follow-up of 23 months compared with 39 and 30 months in pN0 and pN1 disease. Long-term survival estimates for the pN2 stratum are correspondingly unreliable, and the absence of a detectable association in that subgroup should be read as insufficient follow-up rather than as evidence of no benefit. Fourth, the probabilistic model assumes a fixed total of 65 regional nodes and treats all nodes as equally likely to harbour metastasis. Neither assumption holds exactly, since nodal counts vary between patients and metastasis follows anatomical drainage pathways; the resulting probabilities should be read as approximations. Fifth, no external validation cohort was available. Internal validation by split-sample and bootstrap methods was consistent with the primary analysis, but confirmation in an independent population, ideally one in which nodal stations are recorded, remains necessary. Additionally, important clinical information such as performance status, quality of life, driver gene status, and comorbidities were not available in the SEER database. Furthermore, we were unable to explore the impact of additional chemotherapy or targeted therapy on the survival of patients. The lack of such information may limit the generalizability of our findings and should be taken into consideration when interpreting our results (46).

Conclusions

In this population-based analysis, a more extensive nodal examination was associated with a higher detected burden of nodal disease and with lower overall mortality, with the association attenuating beyond approximately 16 examined nodes. The attenuation was confined to node-negative disease and to patients undergoing lobectomy; among node-positive patients the association persisted across the full range examined. A count of 16 nodes may therefore serve as a practical benchmark for auditing the adequacy of nodal examination in node-negative disease after anatomical resection, but should not be regarded as a sufficient standard in node-positive disease. As an observational analysis of registry data, these findings describe associations rather than causal effects, and confirmation in cohorts recording nodal stations is required.

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. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.