XGBoost and SHAP based lymph node ratio thresholds for predicting overall survival in stage I to IIIA NSCLC

The lymph node ratio (LNR) may emerge as a more reliable prognostic indicator for non-small cell lung cancer (NSCLC). This study utilized data collected between 2010 and 2019 from the Surveillance, Epidemiology, and End Results (SEER) 17 registry research database. The cohort comprised patients diagnosed with stage I-IIIA non-small cell lung cancer (NSCLC) who had undergone surgical resection and had at least one lymph node examined. Overall survival was predicted using XGBoost, a robust tree-based ensemble algorithm, based on LNR, and prognostic LNR thresholds were determined using SHAP (SHapley Additive exPlanations) to interpret the model. Smooth-fit curves were also employed to determine the cutoff values for the lymph node ratio (LNR) in relation to mortality. Trend tests and Cox proportional hazards regression models were applied to evaluate the association between LNR and overall survival (OS). This investigation encompassed a total of 61,990 patients who met the criteria for inclusion. The cohort was stratified into three categories based on smooth-fitting curves and SHAP dependence plot: low (<0.1), medium (0.1≤,<0.4), and high (≥ 0.4) LNR groups. Kaplan-Meier curves illustrated that patients in the low LNR group exhibited superior overall survival compared to those in the medium and high LNR groups (P < 0.001). This trend was evident in all subgroup analyses. The Kaplan-Meier curves stratified by LNR groups demonstrate superior discriminatory power compared to those stratified by N-stage grouping. In the population of patients with NSCLC, an elevated LNR is linked to diminished OS. This relationship holds true across all stratified cohorts. Evaluation of LNR serves as a dependable prognostic marker for OS in patients with stage I-IIIA NSCLC undergoing radical surgery.