Log odds of positive lymph nodes compared to positive lymph node ratio and number of positive lymph nodes in prognostic modeling for patients with NSCLC undergoing lobectomy or total pneumonectomy: a population-based study using Cox regression and XGBoost with SHAP analysis.

Abstract

Background: Methods such as the number of positive lymph nodes (nPLN), lymph node ratio (LNR), and log odds of positive lymph nodes (LODDS) are used to predict prognosis in patients with non-small cell lung cancer (NSCLC). We hypothesized that LODDS could be a superior independent predictor of prognosis and aimed to compare its effectiveness with nPLN and LNR in predicting survival outcomes in stage I-IIIA NSCLC patients.

Methods: We utilized data from the Surveillance, Epidemiology, and End Results (SEER) 17 registry (2010-2019) to study NSCLC patients, focusing on those who underwent surgery with confirmed lymph node involvement (N1 or N2 disease). We aimed to compare overall survival (OS) and cancer-specific survival (CSS) based on nPLN, LNR, and LODDS. Kaplan-Meier and Cox regression analyses were employed to evaluate survival, with thresholds determined using X-tile software. An XGBoost model was constructed to predict overall survival in patients using three features: LODDS, LNR, and PLN. SHapley Additive exPlanations (SHAP) analysis was applied to assess feature importance and provide interpretable insights into the model's predictions.

Results: The study analyzed 3,132 eligible NSCLC patients from the SEER database, predominantly male (53.07%) with adenocarcinoma (43.65%) or squamous cell carcinoma (29.76%). Survival outcomes were assessed using nPLN, LNR, and LODDS. LODDS showed superior predictive value for both OS and CSS compared to nPLN and LNR, as indicated by a larger Log Likelihood Ratio (LLR) and smaller Akaike Information Criterion (AIC). Higher scores on npLN, LNR, and LODDS were strongly related with a poorer prognosis, according to Kaplan-Meier analyses (P < 0.001). The SHAP (SHapley Additive exPlanations) analysis of the XGBoost model demonstrated that the LODDS exhibited the highest SHAP values (0.25) for predicting overall survival in patients, consistently outperforming the LNR and the number of nPLN across both training and validation datasets.

Conclusions: Compared to the nPLN and LNR staging systems, LODDS demonstrates superior prognostic power for patients with stage I-IIIA NSCLC undergoing lobectomy or pneumonectomy. By integrating both positive and negative lymph node information, LODDS offers a refined risk stratification that is particularly valuable in cases with high lymph node heterogeneity.