Correspondence on “Identifying the minimum number of retrieved lymph nodes in node-negative duodenal adenocarcinoma: a retrospective analysis of real-world data and the SEER database”

Dear Editor, I am writing to express my perspectives on the article entitled “Identifying the minimum number of retrieved lymph nodes in node-negative duodenal adenocarcinoma: a retrospective analysis of real-world data and the SEER database” by Zhenxing Jiang et al[1]. This study addresses a critical clinical challenge in the management of duodenal adenocarcinoma (DA), aiming to define the optimal number of retrieved lymph nodes (rLNs) for accurate N-stage assessment. By leveraging dual-cohort data (real-world data and SEER database), the authors established stratified rLN thresholds for node-negative (N0) and node-positive (N1) DA patients, providing valuable insights for optimizing surgical and postoperative therapeutic strategies[1]. Nevertheless, after a careful review of the manuscript, we identified several critical points that require further clarification and improvement to enhance the validity and generalizability of the conclusions. First, the retrospective design inherently carries selection bias, and temporal confounding factors may have compromised the consistency of the results. This study relies on historical data with a 10-year follow-up period, during which significant advancements have been made in DA diagnostic techniques (e.g., enhanced imaging modalities) and treatment protocols (e.g., refined surgical approaches and adjuvant therapy regimens). Such dynamic changes in the clinical management system are likely to reduce the homogeneity of research data across different periods, thereby undermining the stability of the study conclusions. Notably, Figure 3B of the external validation cohort clearly demonstrates an upward trend in rLN counts over time. This observation directly corroborates the interference of time-related confounding factors, suggesting that variations in diagnosis and treatment practices across the study period may have biased the core research indicator (rLN count). Second, the lack of detailed information on surgical approaches precludes clarification of the causal relationship between rLNs and prognosis. The manuscript fails to specify the surgical procedures adopted for the included patients. Extant literature has well established that standard surgical approaches for DA are heterogeneous, including distal gastrectomy, pancreaticoduodenectomy (PD), and segmental/local excision[2]. These surgical approaches not only exert a direct and significant impact on patient prognosis but also indirectly affect prognostic analyses by altering rLN counts. For instance, Sakamoto et al confirmed that patients undergoing PD have significantly higher rLN counts than those receiving limited resection[2]. Consequently, the current study cannot distinguish between two potential causal pathways: whether rLN count itself is a prognostic factor, or whether different surgical approaches indirectly influence prognosis by modifying rLN counts through a mediating effect. We recommend the authors supplement detailed surgical data and perform mediation effect analysis to clarify the true relationship between these core variables. Third, the prognostic association pattern of rLNs remains unclear, and the methods used to determine cut-off values have inherent flaws. On one hand, existing studies have reported an association between increased rLN counts and decreased survival rates in certain tumor types[3]. However, this study did not explore whether such a specific association exists in DA, leading to an incomplete understanding of the rLN-prognosis relationship. On the other hand, the authors employed the Contal and O’Quigley method and X-tile method to determine optimal rLN cut-off values, which have three major limitations: (1) forced stratification of the continuous rLN variable (e.g., binary or ternary classification) results in loss of continuous distribution information; (2) inability to capture potential non-linear associations between rLNs and prognosis; and (3) data-driven segmentation is susceptible to outliers, leading to high overfitting risk and poor result stability. If cut-off value analysis is retained, strict statistical validation (e.g., cross-validation, multiple comparison correction, and external cohort validation) is essential to reduce bias and ensure result reliability and generalizability. A more robust alternative is to adopt restricted cubic splines or the K-Adaptive Partitioning for Survival Data method, which preserve variable continuity, accurately characterize non-linear relationships, and enhance result scientificity[4]. In conclusion, Jiang et al’s study explores an important clinical question, but the aforementioned limitations may affect the validity of its conclusions. Addressing these issues will significantly improve the study’s academic value and clinical implications.