We extracted data from the Surveillance, Epidemiology, and End Results Program (SEER) 17 registry study database, encompassing the most recent data available from 2010 to 2019. The SEER 17 database includes registries across various urban and rural regions of the United States. Cancer data collection initiates with the identification of cancer patients diagnosed or treated in hospitals, outpatient clinics, radiology departments, physician offices, laboratories, surgery centers, or other healthcare providers such as pharmacists. All 50 states mandate that newly diagnosed cancers be reported to a central registry. Cancer registries then review these reported cases to ascertain if the information is legally reportable. If deemed reportable, the registry extracts cancer information from the medical records as stipulated by the North American Association of Central Cancer Registries (NAACCR) Data Standards External Site Policy.

Initially, we identified patients aged 18 or older using center codes C34.0-C34.9, which correspond to lung and bronchus diagnoses. Subsequently, we refined our inclusion criteria to encompass individuals diagnosed with NSCLC between January 1, 2010, and December 31, 2019. We focused on patients who underwent primary site-specific surgery and presented with N1 or N2 disease, with histological confirmation of at least one lymph node. The histological tumor types were restricted to those specified in the International Classification of Diseases for Oncology, Third Edition, including squamous cell carcinoma, adenocarcinoma, and other variants. Inclusion criteria also mandated a minimum survival period of one month post-surgery and active follow-up. We extracted clinicopathological data such as age at diagnosis, N stage, histology, gender, race, surgical approach, grade, T stage and primary site.

Exclusion criteria encompassed the presence of multiple primary tumors, unexamined lymph nodes, unknown nPLN, distant metastases, and IIIB or IV stage. The study protocol was reviewed and approved by the Ethics Committee of the Second Hospital of Nanchang University, ensuring adherence to ethical standards and regulations.

OS and CSS were the primary endpoints of this study. Correlations among the LODDS, LNR, and the nPLN were examined through smoothed fitted curves. The Kaplan-Meier method was employed to estimate OS and CSS, with the log-rank test used for comparative analysis of the survival estimates. X-tile software facilitated the determination of optimal thresholds for LNR and nPLNs. The X-tile software analyzes different thresholds using Kaplan-Meier survival curves to determine the optimal cutoff values for LNR and nPLNs. The software selects thresholds that maximize the differentiation of patient prognosis based on survival rate comparisons, ensuring clinical relevance. Specifically, the output heatmap visualizes the impact of various thresholds on survival, aiding researchers in making informed decisions. For the analysis of the number of positive lymph nodes (nPLN), patients were stratified into two groups using x-tile software (Supplementary Figure S1): npLN ≤ 2 and nPLN > 2. Similarly, for the LNR analysis, patients were classified into LNR ≤ 0.5 and LNR > 0.5.

Cox proportional hazards models were utilized to evaluate the impact of demographic, pathological, and treatment variables in both univariate and multivariate contexts. Stratified analysis was used to evaluate the impact of LNR, nPLN, and LODDS on survival across different histologic types, T stages, and N stages. The goodness of fit for the models was assessed using the Akaike Information Criterion (AIC) and the Log Likelihood Ratio (LLR), with lower AIC or higher LLR values signifying superior model fit.

The LODDS were calculated as the logarithm of the ratio of the probability of a positive lymph node to that of a negative lymph node, according to the formula: LODDS=positivelymphnodes+0.5totallymphnodes−positivelymphnodes+0.5. The addition of 0.5 to both the numerator and the denominator was employed to avoid computational singularities. For LODDS analysis, patients were categorized into LODDS < 0.26 and LODDS ≥ 0.26. The values of LODDs were identified by recursive partitioning. Additionally, the number of lymph nodes examined was classified into <10 or ≥10, as the removal of 10 or more lymph nodes was associated with the longest median survival.

A predictive model for OS was constructed using the eXtreme Gradient Boosting (XGBoost) algorithm. The dataset was randomly divided into a training set (2,514 patients) and a validation set (618 patients). The primary endpoint was OS, with censoring applied to patients alive at the end of follow-up. Three lymph node-related features were selected based on their prognostic relevance in NSCLC: the LODDS, LNR, and nPLN. Model training was conducted on the training set, with hyperparameters optimized using five-fold cross-validation, and performance was evaluated on the independent validation set. SHAP (SHapley Additive exPlanations) values were calculated to quantify the contribution of each feature to the model's output, providing global rankings of feature importance as well as local interpretations for individual predictions. The predictive capacity of LODDS, LNR, and PLN was assessed by comparing their SHAP value distributions and mean SHAP importance scores. Statistical significance of differences in feature contributions was evaluated using non-parametric tests. All analyses were performed using Python (version 3.10), with key libraries including XGBoost, SHAP, and Scikit-learn. A p-value < 0.05 was considered statistically significant.

A total of 3,132 eligible patients meeting the criteria were identified from the SEER database. Patient demographics and tumor characteristics are detailed in Table 1. Of these patients, the majority were male (1,662, 53.07%), and the predominant histological types were adenocarcinoma (43.65%) and squamous cell carcinoma (29.76%). Regarding surgical procedures, approximately 87.71% underwent lobectomy, while 12.29% received total pneumonectomy. The mean number of lymph nodes resected was 13.89 (±SD: 9.75), with a median nPLN of 1. The median LNR was 0.14 (Q1-Q3: 0.08–0.27), and the median LODDS was −0.65 (Q1-Q3: −0.89 to −0.37) (Table 1).

Figure 1 illustrates the relationships among the nPLN, LNR, and LODDS. Notably, when the number of positive lymph nodes reaches 22, LODDS ceases to increase with the number of positive lymph nodes, whereas LODDS continues to rise in conjunction with LNR.

The univariate Cox regression analysis revealed that age, gender, histology, primary site, grade, histologic type, T stage, and surgical approach were significantly relatived with both OS and CSS (Table 2).

Following these findings, three distinct multivariate models were developed to assess the predictive capacity of npLN, LNR, and LODDS. In the fully adjusted model, using OS as the outcome measure for the entire cohort, LNR and LODDS as continuous variables demonstrated predictive value for both OS and CSS (P < 0.05), whereas npLN did not show predictive significance (P = 0.0924). In the fully adjusted model, each unit increase in LODDS was associated with a 44% higher risk of overall mortality (OS) (HR = 1.44, 95% CI: 1.29–1.60, p < 0.0001), while increases in LNR and PLN were linked to 4% (HR = 1.04, 95% CI: 1.01–1.07, p = 0.0180) and 1% (HR = 1.01, 95% CI: 1.00–1.01, p = 0.0924) increases in risk, respectively. For cancer-specific survival (CSS), each additional unit of LODDS corresponded to a 45% higher risk of cancer-specific mortality (HR = 1.45, 95% CI: 1.28–1.64, p < 0.0001), while changes in LNR and PLN were associated with modest increases of 2% (HR = 1.02, 95% CI: 0.98–1.07, p = 0.3352) and 1% (HR = 1.01, 95% CI: 1.00–1.01, p = 0.1310) in risk, respectively (Table 3). Significant survival differences were observed across different groups. For patients with PLN ≤ 2, the 1-year, 3-year, and 5-year survival rates were 72.19%, 40.79%, and 21.53%, respectively, with a median survival time of 27 months. In contrast, patients with PLN > 2 had reduced survival rates of 59.42% at 1 year and 11.76% at 5 years, with a shorter median survival time of 17 months. Similar trends were observed for LNR and LODDS groups, with patients in the LNR ≤ 0.5 and LODDS ≤ 0.26 categories showing higher survival rates and longer median survival times. Overall, lower PLN, LNR, and LODDS values were associated with better survival outcomes. When CSS was used as the outcome, LODDS as a continuous variable was a significant predictor of both OS and CSS (P < 0.0001), while LNR (P = 0.3352) and nPLN (P = 0.1310) did not demonstrate predictive capability. The respective HR and CI are detailed in Table 3.

In stratified analysis, the LODDS metric demonstrated a significant prognostic value for both lung cancer-specific mortality and all-cause mortality. Patients with LODDS > 0.26 had substantially higher lung cancer-specific mortality in SCC and ADC subgroups (SCC HR = 1.9, 95% CI: 1.1–3.3, p = 0.018; ADC HR = 1.7, 95% CI: 1.3–2.4, p < 0.001). Furthermore, among patients with N2 staging, LODDS > 0.26 was also associated with a markedly increased mortality risk (HR = 1.9, 95% CI: 1.1–3.4, p = 0.025). Consistent with these findings, all-cause mortality was similarly elevated in SCC and ADC patients with LODDS > 0.26 (SCC HR = 1.7, 95% CI: 1.1–2.7, p = 0.023; ADC HR = 1.7, 95% CI: 1.3–2.2, p < 0.001). Across various histologic subtypes and staging groups, LODDS > 0.26 was closely associated with poorer prognosis, underscoring its potential clinical value in risk stratification and outcome prediction (Table 4).

Among all patients, the LODDS model exhibited superior fit compared to both npLN and LNR (Table 3). Specifically, LODDS demonstrated the most accurate predictive capacity in cohorts with fewer than 10 ELN, characterized by a higher Log Likelihood Ratio (LLR) and a lower Akaike Information Criterion (AIC) which best predicted both OS and CSS. Similarly, in patients with 10 or more examined lymph nodes, LODDS continued to provide superior predictive value for OS and CSS in comparison to both LNR and npLN (Table 3). Higher scores on all three measures—nPLN, LNR, and LODDS—were substantially (P < 0.001) related with a worse prognosis, according to Kaplan-Meier analyses (Figures 2, 3).

The XGBoost model was developed to predict OS using a cohort of 2,514 patients in the training set and 618 patients in the validation set. The prognostic relevance of three lymph node-associated features—log odds of positive lymph nodes LODDS, LNR, and the number of positive lymph nodes PLN—was systematically evaluated.

In both the training and validation sets, LODDS emerged as the most critical feature for OS prediction, exhibiting the highest mean SHAP values (0.25), followed by LNR, while PLN demonstrated the least importance (Figures 4A,C). This hierarchy of feature importance was consistently observed across datasets (Figures 4B,D), underscoring the superior predictive capacity of LODDS.

SHAP summary plots further delineated the influence of individual features on model output (Figures 4A,C). Among the features, LODDS displayed the widest distribution of SHAP values, signifying its pronounced impact on survival prediction. Elevated LODDS values were strongly correlated with increased mortality risk, while LNR exhibited a moderate contribution. Conversely, PLN had a negligible effect on model predictions.

These results collectively establish LODDS as the most robust and reliable predictor of OS among the three evaluated features. Its consistent superiority across training and validation cohorts highlights its potential as a pivotal prognostic marker for NSCLC patients undergoing survival risk stratification.