Article Text
PDF
Abstract
Background The optimal number of lymph nodes to be dissected during lung cancer surgery to minimise the postoperative recurrence risk remains undetermined. This study aimed to elucidate the impact of the number of dissected lymph nodes on the risk of postoperative recurrence of non-small cell lung cancer (NSCLC) using machine learning algorithms and statistical analyses.
Methods We retrospectively analysed 650 patients with NSCLC who underwent complete resection. Five machine learning models were trained using clinicopathological variables to predict postoperative recurrence. The relationship between the number of dissected lymph nodes and postoperative recurrence was investigated in the best-performing model using Shapley additive explanations values and partial dependence plots. Multivariable Cox proportional hazard analysis was performed to estimate the HR for postoperative recurrence based on the number of dissected nodes.
Results The random forest model demonstrated superior predictive performance (area under the receiver operating characteristic curve: 0.92, accuracy: 0.83, F1 score: 0.64). The partial dependence plot of this model revealed a non-linear dependence of the number of dissected lymph nodes on recurrence prediction within the range of 0–20 nodes, with the weakest dependence at 10 nodes. A linear increase in the dependence was observed for ≥20 dissected nodes. A multivariable analysis revealed a significantly elevated risk of recurrence in the group with ≥20 dissected nodes in comparison to those with <20 nodes (adjusted HR, 1.45; 95% CI 1.003 to 2.087).
Conclusions The number of dissected lymph nodes was significantly associated with the risk of postoperative recurrence of NSCLC. The risk of recurrence is minimised when approximately 10 nodes are dissected but may increase when >20 nodes are removed. Limiting lymph node dissection to approximately 20 nodes may help to preserve a favourable antitumour immune environment. These findings provide novel insights into the optimisation of lymph node dissection during lung cancer surgery.
- Lung Cancer
- Non-Small Cell Lung Cancer
Data availability statement
Data are available upon reasonable request.
This is an open access article distributed in accordance with the Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited, appropriate credit is given, any changes made indicated, and the use is non-commercial. See: http://creativecommons.org/licenses/by-nc/4.0/.
Statistics from Altmetric.com
WHAT IS ALREADY KNOWN ON THIS TOPIC
Systematic lymph node dissection is the standard procedure for resection of lung cancer. However, the practicability of selective lymph node dissection, which curtails the degree of dissection, has been validated by the movement towards less invasive surgical approaches.
WHAT THIS STUDY ADDS
The postoperative recurrence risk of lung cancer is contingent on the number of dissected lymph nodes. The risk of recurrence diminishes when the number of dissected nodes is approximately 10 and increases linearly as the number of dissected nodes surpasses 20.
HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY
The effectiveness of immune checkpoint inhibitors in the context of neoadjuvant chemotherapy for lung cancer has been previously demonstrated. The significance of lymph nodes in antitumour immunity is becoming increasingly apparent. The findings of this study suggest that a suitable degree of lymph node dissection is crucial for sustaining the antitumour immune function of lymph nodes when considering immunotherapy.
Introduction
Lung cancer is the predominant cause of neoplastic mortality on a global scale. Surgical resection serves as the principal modality for management; however, recurrence is documented in 20%–26% of cases following surgery.1 2 Lymph node dissection is a standard surgical procedure that is typically performed concomitantly with lung cancer resection.3 Although systematic lymph node dissection is considered the gold standard for lymph node dissection during lung cancer resection worldwide, selective lymph node dissection, which involves a smaller dissection area than systematic lymph node dissection in clinical practice, is also employed. Some retrospective studies have demonstrated the non-inferiority of selective lymph node dissection to systematic lymph node dissection with regard to the prognosis.4 5 A meta-analysis showed that selective lymph node dissection was associated with comparable overall survival (HR), 0.83; 95% CI 0.67 to 1.02) and disease-free survival (HR, 0.82; 95% CI 0.65 to 1.04) to systematic lymph node dissection in patients with clinical N0-1 non-small cell lung cancer (NSCLC).4 Nevertheless, the specifics regarding the effect of the number of dissected lymph nodes on the prognosis of lung cancer following surgery remain unknown.
The effectiveness of selective lymph node dissection compared with systematic lymph node dissection is currently under evaluation in randomised controlled trials. The Japan Clinical Oncology Group 1413 trial is currently investigating the non-inferiority of lobe-specific lymph node dissection to systematic lymph node dissection in patients with clinical stage I–II NSCLC.6 However, it will take a considerable amount of time for the outcomes to become clear. Retrospective investigations have examined the effect of the number of dissected lymph nodes on the postoperative prognosis of lung cancer; however, no consistent outcomes have been reported, and the number of dissected nodes that correlate with the prognosis continues to be undetermined.7–9 Recently, the effectiveness of immune checkpoint inhibitors in neoadjuvant and adjuvant chemotherapy for lung cancer management has garnered considerable attention. In a phase 2 trial, neoadjuvant chemotherapy plus nivolumab with or without ipilimumab resulted in major pathological response rates of 33% and 38%, respectively.10 It is highly probable that immunotherapy will play a crucial role in the surgical domain in the future.10 11 It has been increasingly reported that lymph nodes play a crucial immunological role in the fight against cancer within the antitumour immune environment.12 13 In addition, lymph node dissection has been linked to reduced therapeutic efficacy of immune checkpoint inhibitors.14 These observations suggest that lymph node removal may inhibit the immune response against tumours.
In recent years, progress in the field of machine learning has resulted in an increasing volume of research in the medical domain regarding forecasting prognosis using techniques grounded in machine learning algorithms.15–19 However, few studies have formulated prognostic models for the recurrence of lung cancer following surgery and have provided insight into the correlation between the number of dissected lymph nodes and the likelihood of recurrence.
We hypothesised that the risk of postoperative recurrence in NSCLC is directly correlated with the number of resected lymph nodes, considering the antitumour immune response of the lymph nodes. To validate our hypothesis, we developed a predictive model for postoperative recurrence using a machine learning algorithm and analysed the impact of the number of dissected lymph nodes on the prediction of postoperative recurrence. Furthermore, we employed a multivariable Cox proportional hazard analysis to assess the influence of the number of dissected lymph nodes on recurrence-free survival (RFS) following lung cancer surgery. Validation of our hypothesis through a multifaceted approach incorporating both machine learning algorithms and statistical methods may provide valuable insights into the significance of lymph node dissection during lung cancer resection in real-world clinical settings.
Methods
Patients
A total of 650 patients diagnosed with NSCLC who underwent surgical lung resection at Kinki-Chuo Chest Medical Center (KCMC) between April 2017 and April 2022 were enrolled in our retrospective study. Our inclusion criteria comprised patients who had undergone pathologically complete resection (R0) of lung cancer and patients with incomplete removal of the tumour (R1) were excluded from our study. Selective lymph node dissection was performed for right upper, right lower, left upper and left lower lobectomy, whereas systematic lymph node dissection was performed for right middle lobectomy, bilobectomy (right upper and middle lobectomy, right middle and lower lobectomy) and pneumonectomy. Lymph node sampling was performed for segmentectomy, and no lymph node dissection was performed for wedge resections.
The resected lung cancer tissue was histopathologically diagnosed by a pathologist in accordance with the 2015 WHO classification. Age, sex, preoperative blood sample data, histological type, pathological tumour-node-metastasis classification (American Joint Committee eighth edition), tumour diameter (invasive size), pleural involvement (pl), vascular invasion (v), lymphatic invasion (Ly), expression of programmed death-ligand 1 (PD-L1), number of dissected lymph nodes, clinical diagnosis of postoperative recurrence and RFS were retrospectively collected from medical records. Patients lacking these records were excluded from the study. The number of dissected lymph nodes was defined as the number of lymph nodes removed and examined by a pathologist.
Sample size estimation
We estimated the required sample size based on a previous study with a similar design.20 Assuming an expected success rate of lymph node dissection of 0.9, threshold success rate of 0.7, significance level of 0.05 (α=0.05) and power of 80% (β=0.2), a sample size of 144 cases would be required. The sample size of 650 cases in our study substantially exceeded this estimated requirement, leading us to conclude that the statistical power of our study was sufficient.
The RFS
The primary endpoint of this study was the clinical diagnosis of postoperative recurrence. RFS was determined as the duration between surgery for lung cancer resection and the clinical diagnosis of recurrence. The period of being recurrence-free status was considered until the final confirmation of the patient’s recurrence-free status after surgery. Patients who underwent surgery underwent blood sampling and radiography at intervals of 3–6 months. Additional diagnostic tests, including MRI of the head, contrast CT, positron emission tomography and pathological examination of tissue biopsy samples, were performed when abnormal findings indicating possible disease recurrence were observed. The diagnosis of recurrence was based on a thorough clinical evaluation of these test results and was determined at a joint conference consisting of general thoracic surgeons, oncologists, pathologists and radiologists.
PD-L1 immunohistochemistry
Viable neoplastic cells present in the entire pathological specimen of the resected lung cancer tissue sample were assessed by a pathologist. The PD-L1 clone 22C3 pharmDx kit and Dako Automated Link 48 platform (Agilent Technologies, Dako, Carpinteria, California) immunohistochemistry assays were used to determine the level of PD-L1 expression. The PD-L1 tumour proportion score (TPS) was computed as the percentage (ranging from 0% to 100%) of complete or partial membranous staining observed in tissue samples.
Machine learning
The variables used in the machine learning algorithm were chosen based on factors that have been previously recognised as having a correlation with postoperative recurrence.21–24 The following variables were selected: tumour diameter (invasive size), presence of cancerous cells in lymph nodes (N0–N2), pathological stage (stage I–III), invasion of pleura (pl0–pl3), invasion of blood vessels (v0–v1), invasion of lymphatic vessels (Ly0–Ly1), histological classification of cancer (adenocarcinoma, squamous cell carcinoma or others), number of dissected lymph nodes, expression of PD-L1 (TPS), neutrophil-to-lymphocyte ratio (NLR), age and sex. Categorical variables were converted into binary variables using dummy coding. To prevent overfitting, the entire dataset was randomly divided into three parts: training, validation and test sets, comprising 70%, 15% and 15% of the dataset, respectively.
Python packages (Sklearn. ensemble. Random Forest Classifier, Sklearn. ensemble. Gradient boosting classifier, xgboost. XGBClassifier, Sklearn. ensemble. AdaBoostClassifier, and Sklearn. linear_model. LogisticRegression) were used to construct the machine learning models. The training set was used to instruct the machine learning models. A validation set was used to assess and compare the performance of each machine learning model. Receiver operating characteristic (ROC) and precision-recall (PR) curves were generated to qualitatively evaluate the performance of each model. In addition, the predictive capabilities of each model were quantitatively assessed using metrics such as accuracy, F1 score, Brier score, area under the ROC curve (ROC AUC) and PR curve (PR AUC). The final evaluation of the predictive performance of the selected model was conducted using a test set. Continuous variables were presented as medians (IQR), while categorical variables were presented as absolute numbers and percentages. Statistical significance was set at a threshold of p<0.05.
The prediction of postoperative recurrence was performed using machine learning algorithms employing five classifiers: random forest, gradient boosting, XGBoost, Adaboost and logistic regression. Gradient boosting, XGBoost and AdaBoost are decision tree-based methods similar to random forests, which have been widely used in recent times owing to their superior accuracy on complex data. In contrast, logistic regression is a non-decision tree-based method. Considering the difficulty of evaluating all machine learning models, we selected these five classifiers for our study because they are representative and possess distinct characteristics. Bayesian hyperparameter optimisation using Optuna, a Python library, was employed to choose the hyperparameters of each machine learning model to predict postoperative recurrence. For each model, we executed a set of 10 trials to identify the optimal hyperparameter configuration. Subsequently, we selected the hyperparameters associated with the most effective configuration as the optimal hyperparameters for each model under examination (online supplemental table S1). In our study, random forest was determined to be superior in predicting postoperative recurrence compared with the other four machine learning models; thus, we confirmed the significance of the variables incorporated within the random forest model.
Supplemental material
The predictions were interpreted using Shapley Additive exPlanations (SHAP) values.25 The SHAP values rely on Shapley values from coalition game theory, which provide a dependable and precise method for computing the contribution of each variable to a machine learning model’s prediction. The shap V.0.28.5 Python package was used to derive the SHAP values and matplotlib V.3.0.311 was used to construct the visualisations. The plot_partial_dependence function from the scikit-learn Python package was employed to examine the effect of variable changes on predictions in the random forest model.
Statistical analyses
Multivariable Cox proportional hazard analysis was conducted to estimate the HR for postoperative recurrence of NSCLC associated with the number of dissected lymph nodes. This analysis allows the examination of the covariate of the number of events, with outcomes divided by 10.26 In other words, a multivariable Cox proportional hazard analysis requires a minimum of 10 outcome events per variable (EPV). In the current study, this number amounted to 142 (ie, the number of cases of postoperative recurrence), divided by 10 (yielding a result of 14). The explanatory variable under consideration for the evaluation was the number of dissected lymph nodes. In addition, 13 variables were selected as confounding factors. As the variables of age, sex and blood vessel invasion (v0–v1) were deemed unimportant for prediction after crucial variable extraction through the SHAP values of the random forest model, these variables were excluded from the analysis, with the following chosen as confounding variables: tumour size, lymph nodes (N0–N2 (reference: N0)), pathological stage (stage I–III (reference: stage I)), histological type (squamous cell carcinoma and others (reference: adenocarcinoma)), invasion of the pleura (pl0–pl3 (reference: pl0)), invasion of the lymphatic vessels (Ly0–Ly1 (reference: Ly0)), TPS and NLR. However, it also suggests that the rule of using 10 EPV in the Cox model can be relaxed.25 That is, a simulation study suggested that a Cox model with 5–9 EPV yielded acceptable results. Therefore, we used a Cox proportional hazards model with 5 EPV. In the Cox proportional hazards model adhering to the 5 EPV rules, the maximum number of variables that could be included was 28. Consequently, the 5 EPV models incorporated additional variables relative to the 10 EPV models, including surgical procedure (right middle lobectomy, right lower lobectomy, left upper lobectomy, left lower lobectomy, segmentectomy, wedge resection, pneumonectomy, bilobectomy and lobectomy with combined resection (reference: right upper lobectomy)) as well as age, sex and blood vessel invasion. The validity of the proportional hazards assumption in the Cox models was assessed by examining martingale residual plots. To determine whether multicollinearity was present or absent in the variables of the multivariable Cox proportional hazards model, we used a variance inflation factor (VIF) with a threshold of <2.
Statistical analyses were performed using Easy R (EZR) (Saitama Medical Center, Saitama, Japan), a graphical user interface for R (The R Foundation for Statistical Computing, Vienna, Austria). EZR is an improved version of R commander with additional biostatistical functions.27
Study design
In summary, we employed a multi-faceted approach combining machine learning and statistical methods to comprehensively evaluate the association between the number of dissected lymph nodes and postoperative recurrence (online supplemental figure S1). First, 650 patients with resected NSCLC were randomly divided into a training set (455 cases, 70%), a validation set (97 cases, 15%) and a test set (98 cases, 15%). Next, we constructed postoperative recurrence prediction models using the training set and evaluated and selected the models using the validation set. The predictive performance of the final model was evaluated using a test set. Furthermore, to interpret the association between the number of dissected lymph nodes and postoperative recurrence, we evaluated the importance of each variable in the selected model using SHAP values and created a partial dependence plot of the number of dissected lymph nodes. As a statistical approach, we used a multivariable Cox proportional hazard model to evaluate the association between the number of dissected lymph nodes and RFS after adjusting for confounding factors. The evaluation variable was the number of dissected lymph nodes and the outcome was RFS.
Supplemental material
Patient and public involvement
None.
Results
Study cohort
Among the 650 patients, lobectomy was the most frequently performed surgical procedure (78%), with right upper lobectomy being the most prevalent (29%). The number of lymph nodes dissected during the surgical procedure was approximately 10–15 for lobectomy, 5 for reduction surgery and 20 for pneumonectomy or bilobectomy. The recurrence rate following right upper lobectomy was 17%, whereas the rates for other types of lobectomies (including right middle lobe, right lower lobe, left upper lobe, left lower lobe and bilobectomy) were approximately 25%. Pneumonectomy and lobectomy with combined resection had frequencies of 40% and 53%, respectively (online supplemental table S2).
The entire cohort of 650 patients was randomly stratified into a training set of 455 (70%), validation set of 97 (15%) and test set of 98 (15%). The incidence of postoperative recurrence was 142 (22% N=142) cases. The median (range) number of dissected lymph nodes for the entire cohort was 13 (range, 7–19). The histological subtypes of lung cancer identified in our study were adenocarcinoma in 478 cases (74%), squamous cell carcinoma in 116 cases (18%) and other subtypes in 56 cases (8%). Pathological staging was stage I, stage II and stage III in 459 (71%), 112 (17%) and 79 (12%) patients, respectively. The median (range) RFS was 729 days (range, 360–1150 days).
Statistical analysis did not indicate any statistically significant differences among the three subgroups with regard to each variable, showing a relatively homogeneous distribution (table 1).
Clinical characteristics and outcomes of 650 patients undergoing lung cancer resection
Performance evaluation of machine learning models
The random forest, gradient boosting, XGBoost, AdaBoost and logistic regression machine learning models were trained using the training set. Subsequently, the performance of each model was assessed by using a validation set. The model performance was evaluated by measuring various metrics such as the ROC AUC, PR AUC, accuracy, F1 score and Brier score. Table 2 shows the predictive performance of each machine learning model in identifying postoperative recurrence in the validation cohort. The random forest model achieved the highest accuracy of 0.88 (95% CI 0.78 to 0.97), F1 score of 0.68 (95% CI 0.55 to 0.82) and lowest Brier score of 0.11 (95% CI 0.05 to 0.18) among all models. The AdaBoost model showed the highest ROC AUC of 0.93 (95% CI 0.85 to 1.00) and PR AUC of 0.75 (95% CI 0.62 to 0.87), followed by the random forest model with an ROC AUC of 0.89 (95% CI 0.80 to 0.98) and PR AUC of 0.73 (95% CI 0.61 to 0.86) (online supplemental figure S2). Relative to the other models, the gradient boosting model had the lowest accuracy of 0.77 (95% CI 0.65 to 0.89) and F1 score of 0.35 (95% CI 0.23 to 0.48). The logistic regression model showed similar performance to the random forest model in terms of accuracy, F1 score and Brier score. Based on the overall performance across the different evaluation metrics, the random forest model was selected as the best model for predicting postoperative recurrence. Although the AdaBoost model showed slightly higher ROC AUC and PR AUC values, the random forest model achieved the best balance of performance across all metrics, including the highest accuracy, F1 score and lowest Brier score. Therefore, we chose the random forest model as the final model to conduct further analyses and interpretation. The random forest model’s performance was assessed on the test set (online supplemental table S3) and yielded the following high levels of predictive accuracy: ROC AUC of 0.92 (0.84 to 0.99), PR AUC of 0.79 (0.67 to 0.90), accuracy of 0.83 (0.72 to 0.93), F1 score of 0.64 (0.51 to 0.77) and Brier score of 0.10 (0.04 to 0.17).
Supplemental material
Predictive performance of various machine learning models in identifying postoperative lung cancer recurrence in the validation cohort
Importance and dependence of the variables in the prediction
The SHAP analysis was conducted using a random forest model to assess the significance of the variables. Of all the variables analysed, the number of dissected lymph nodes had the fourth highest mean SHAP value, preceded by RFS, pathological stage I and pathological N0 (figure 1A). In addition, the distribution of the SHAP values for each feature was investigated (figure 1B). Regarding the number of dissected lymph nodes, red dots were located in areas with positive SHAP values, whereas blue dots were located in areas with negative SHAP values. Hence, an increased number of dissected lymph nodes corresponded to a greater contribution to the model output. To further explore the effect of the number of dissected lymph nodes on prediction, a partial dependency graph was generated (figure 2). Although the impact on prediction was highly non-linear from 0 to 20 and lowest at approximately 10, the dependency on prediction increased linearly with the number of nodes dissected exceeding >20.
(A) The most impactful variables on the prediction of postoperative recurrence (ranked from most to least important). (B) Distribution of the impacts of each variable on the model output. The colours represent the numeric features of variables: red for larger values and blue for smaller values. A negative SHAP value (extending to the left) indicates a decreased probability of postoperative recurrence, whereas a positive value (extending to the right) indicates an increased probability. SHAP, Shapley additive explanation.
Partial dependency graph for exploring the impact of the number of dissected lymph nodes on the prediction of postoperative recurrence. The x-axis shows the number of dissected lymph nodes, and the y-axis shows the contribution (partial dependence) to the prediction of postoperative recurrence.
Multivariable Cox proportional hazard analyses
A multivariable model was used to investigate the correlation between the number of dissected lymph nodes and postoperative recurrence. The number of dissected lymph nodes was divided into two categories as binary variables with a threshold of 20. Ideally, all variables used in a machine learning model should be considered confounding factors. However, the number of variables that could be incorporated into the model as confounding factors was limited by the number of recurrences. In this study, 13 variables were eligible for inclusion as confounders in the multivariable model. Considering that age, sex and vascular invasion had little impact on the machine learning model, as demonstrated by the variable importance in figure 1A, variables other than these three were included as confounding factors in the multivariable model.
Table 3 presents the results of the multivariable Cox proportional hazard analysis of the 10 EPV. The multicollinearity of each explanatory variable was assessed using VIF. For all variables, VIF<2 was verified, indicating the absence of multicollinearity (online supplemental table S4). The proportional hazard was confirmed by generally horizontal martingale residual plots (online supplemental figure S3). The risk of recurrence was significantly higher in the group with ≥20 dissected lymph nodes than in the group with <20 nodes removed (adjusted HR 1.45, 95% CI 1.003 to 2.087). Online supplemental table S5 presents the results of the multivariable Cox proportional hazard analysis with the 5 EPV model. The 5 EPV model was subjected to an analysis by incorporating the surgical procedure, sex, age and vascular invasion into the variables employed in the 10 EPV model. Furthermore, T-size was transformed into a T-factor for categorical variables. The risk of recurrence was significantly higher in the group with ≥20 dissected lymph nodes than in those with <20 dissected nodes removed (adjusted HR 1.52; 95% CI 1.04 to 2.23). For all variables, VIF<2 was verified, indicating the absence of multicollinearity. Proportional hazards were also confirmed by generally horizontal martingale residual plots (online supplemental figure S4).
Supplemental material
Supplemental material
A multivariable Cox proportional hazard analysis of the RFS according to the number of dissected LNs
Discussion
The primary aim of our study was to investigate the relationship between the number of dissected lymph nodes and the risk of postoperative recurrence in patients with NSCLC using a multifaceted approach employing both machine learning algorithms and statistical analyses. By leveraging these complementary methods, we aimed to provide a comprehensive and robust assessment of the association between the number of dissected lymph nodes and the risk of recurrence. Our analysis revealed two key outcomes. First, the association between the number of dissected lymph nodes and the risk of postoperative recurrence is non-linear when the number of dissected nodes is less than 20. Second, the risk of recurrence may increase linearly when the number of dissected lymph nodes exceeds 20. These two discoveries are unprecedented and have not been previously reported. These findings, consistently observed across both machine learning and statistical approaches, suggest that there may be an optimal range for the number of lymph nodes that should be dissected during lung cancer surgery to minimise the risk of postoperative recurrence. Specifically, our results indicate that dissecting approximately 10–20 lymph nodes may be associated with a lower risk of recurrence, while removing more than 20 lymph nodes may potentially increase this risk. Taken together, our findings suggest that the risk of postoperative recurrence may increase when the number of dissected lymph nodes exceeds a certain threshold, highlighting the importance of striking a balance between adequate lymph node dissection for accurate staging and minimising the potential negative impact of extensive dissection on patient outcomes.
The postoperative recurrence rate in our cohort was 22%, which is consistent with previous reports.2 Nevertheless, our cohort comprised of more advanced cases than those in previous reports. If the cohort was comparable to those in prior studies, the recurrence rate might have been even lower. Some studies have reported a correlation between the number of dissected lymph nodes and postoperative prognosis. A study in a significant cohort of 16 800 patients demonstrated that individuals who underwent dissection of 13–16 lymph nodes, according to the reference of 1–4 lymph nodes, exhibited the most favourable prognosis. Conversely, those who had ≥17 nodes dissected had a relatively poorer outcome than those who had 13–16 nodes dissected.28 A study involving 7627 patients reported a positive prognosis for a subgroup of patients with ≥4 dissected lymph nodes (with the reference being set at <4).29 An analysis of 549 patients indicated that dissection of ≥12 lymph nodes (as defined by the reference of 1–2) may be associated with a favourable prognosis.30 A report on 974 patients indicated that an optimal prognosis is linked to the dissection of 31–40 nodes (criteria: 0–10).20 Nonetheless, the outcomes of prior research have been mixed, and information on the correlation between the number of dissected lymph nodes and postoperative prognosis, as well as knowledge regarding the ideal number of nodes to dissect, has yet to be clearly established. In previous studies, the number of dissected lymph nodes was set as a reference within a non-linear spectrum of association with postoperative recurrence, which may explain the lack of uniform results. Therefore, the relationship between the number of dissected lymph nodes, postoperative prognosis and optimal number of lymph nodes to be dissected remains unclear.
To address this issue, we constructed a postoperative recurrence prediction model for NSCLC based on machine learning algorithms and thoroughly investigated the association between the number of dissected lymph nodes and postoperative recurrence. Our results demonstrated that the random forest model exhibited the best predictive performance. A previous study on machine learning reported that the random forest model exhibits superior predictive performance.15 31 32 In this study, we found that the number of dissected lymph nodes is one of the most important factors used to predict postoperative recurrence in the random forest model, and that the dependence of the number of dissected lymph nodes on the prediction of postoperative recurrence in the random forest model is non-linear. Previous research using conventional statistical analyses in a large cohort reported that dissecting 11–16 lymph nodes was associated with the best outcomes in terms of mortality and survival duration.28 The results of the analysis using our machine learning algorithm showed that the dependence of the number of dissected lymph nodes within the range of 0–20 nodes on the prediction of postoperative recurrence was non-linear. In addition, there was a tendency for the prediction of postoperative recurrence to exhibit its lowest dependence on approximately 10 dissected lymph nodes. In contrast, the association between the number of dissected lymph nodes and the prediction of recurrence became linearly stronger when the number of dissected nodes exceeded 20. Given these findings, the optimal number of dissected lymph nodes with a low risk of postoperative NSCLC recurrence may be in the range of approximately 10–20 nodes. We showed that the HR for postoperative recurrence was higher in the group with ≥20 lymph nodes dissected than in the group with <20 nodes dissected, even in a conventional multivariable analysis. Restricting lymph node dissection to approximately 20 nodes may thus be a prudent approach, considering the potential for increased recurrence risk associated with the dissection of numerous lymph nodes.
A biological mechanism may account for the non-linear relationship between the number of dissected lymph nodes and the risk of postoperative recurrence. In recent years, it has become evident that tumour-draining lymph nodes (TDLNs) are crucial components of the immune response against tumours. Dendritic cells that recognise tumour antigens migrate to TDLNs and trigger T-cell activation. The route by which activated T cells are transported to the systemic circulation via efferent lymph vessels and then released to micrometastatic foci is believed to be the primary mechanism by which anticancer T cells become activated.33 In a study utilising a mouse model of carcinoma, it was reported that the quantity of activated T cells was more abundant in TDLNs than in non-TDLNs. In addition, mice that underwent TDLN resection exhibited a greater inclination towards tumour growth than the non-resection group.14 Expanded dissection of TDLNs may lead to the extraction of numerous lymph nodes that act as crucial sites for cancer antigen presentation and T-cell activation, thereby possibly elevating the risk of postoperative recurrence owing to compromised antitumour immunity. Supporting this concept, a recent retrospective study demonstrated that among patients with postoperative recurrent NSCLC, those who underwent systematic lymph node dissection experienced significantly inferior progression-free survival after immune checkpoint inhibitor therapy in comparison to those who received selective dissection (HR 3.709, p=0.034), even after adjusting for potential confounding factors.34 This report provides clinical evidence to support that extensive lymph node removal may attenuate the efficacy of immunotherapy, consistent with our hypothesis.
Exposure to cancer antigens may be crucial for eliciting antitumour immunity by activating diverse T cells.35 Hence, the elimination of many TDLNs during lymph node dissection can affect the diversity of the activated T cells. If the number of dissected lymph nodes is within a certain range, residual TDLNs may support T cell activation, thus preserving the effectiveness of antitumour immunity. However, when the number of dissected lymph nodes exceeds a certain limit, T cell activation by TDLNs can become dysfunctional, thereby increasing the risk of postoperative recurrence by reducing the influx of diverse activated T cells into micrometastatic tissues. Our research suggests that the impairment of the lymph node function in antitumour immunity may begin when the number of lymph nodes exceeds 20.
Several limitations of the present study warrant mention. First, it was a single-centre study. Whether comparable trends can be observed in a multicentre cohort should be explored in a future studies. Second, the number of patients was small. The small size of the study cohort may have affected the robustness of the results. We employed a different analytical approach to compensate for the robustness of our results. We showed that the more lymph nodes dissected, the higher the SHAP value and the higher the risk of recurrence (figure 1). Multivariable analysis also showed that the risk was higher in the group with ≥20 dissected nodes than in those with fewer dissected nodes (table 3). These different analytical approaches demonstrate the robustness of the results. However, future analyses in larger cohorts should examine the association between the number of dissected lymph nodes and the risk of postoperative recurrence. Third, the molecular mechanisms underlying the association between the number of dissected lymph nodes and postoperative recurrence are unknown. We considered a possible mechanism related to the TDLNs, but further studies are needed to validate this.
In conclusion, using a machine learning algorithm, we demonstrated that a non-linear association exists between the number of dissected lymph nodes and the risk of postoperative recurrence of NSCLC. Our findings also indicate that the risk of postoperative recurrence remains relatively low when the number of dissected lymph nodes is approximately 10; however, it may increase beyond 20. Lymph node dissection is an essential component of lung cancer resection for accurate staging. Nevertheless, limiting the number of dissected lymph nodes to approximately 20 may aid in maintaining a favourable antitumour immune environment, thus reducing the risk of recurrence.
Supplemental material
Data availability statement
Data are available upon reasonable request.
Ethics statements
Patient consent for publication
Ethics approval
This study was approved by the Institutional Review Board (IRB) of the National Hospital Organization Kinki-Chuo Chest Medical Center (KCMC) (approval number: 2023-1) and was carried out in accordance with the Declaration of Helsinki. The IRB of KCMC waived the requirement for informed consent from all research participants because of the retrospective and anonymous nature of the study. Information about opting out of this study is provided on the KCMC homepage.
References
Supplementary materials
Supplementary Data
This web only file has been produced by the BMJ Publishing Group from an electronic file supplied by the author(s) and has not been edited for content.
Footnotes
Contributors Study design and concept: KK. Data collection: KK and HS. Machine learning and statistical analysis: KK. Manuscript drafting: KK. Review of the manuscript: HY, HS, TT, KO, SA. Study supervision: SA. KK was responsible for the overall content as the corresponding author. KK is the guarantor.
Funding KK received a specific grant from a Grant-in-Aid for Clinical Research from the National Hospital Organization, a public funding agency, for this study.
Competing interests None declared.
Patient and public involvement Patients and/or the public were not involved in the design, or conduct, or reporting, or dissemination plans of this research.
Provenance and peer review Not commissioned; externally peer-reviewed.
Supplemental material This content has been supplied by the author(s). It has not been vetted by BMJ Publishing Group Limited (BMJ) and may not have been peer-reviewed. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by BMJ. BMJ disclaims all liability and responsibility arising from any reliance placed on the content. Where the content includes any translated material, BMJ does not warrant the accuracy and reliability of the translations (including but not limited to local regulations, clinical guidelines, terminology, drug names and drug dosages), and is not responsible for any error and/or omissions arising from translation and adaptation or otherwise.