Dynamic changes in postoperative risk of recurrence of non-small cell lung cancer according to variations in PD-L1 expression levels

Dynamic changes in postoperative risk of recurrence of non-small cell lung cancer according to variations in PD-L1 expression levels | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Dynamic changes in postoperative risk of recurrence of non-small cell lung cancer according to variations in PD-L1 expression levels Kensuke Kojima, Hironobu Samejima, Takafumi Iguchi, Toshiteru Tokunaga, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4334704/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 10 You are reading this latest preprint version Abstract The accurate prediction of postoperative recurrence is important for optimizing treatment strategies for non-small cell lung cancer (NSCLC). Previous studies have identified the PD-L1 expression in NSCLC as a risk factor for postoperative recurrence. This study aimed to examine the contribution of the PD-L1 expression in predicting postoperative recurrence using machine learning. The clinical data of 647 NSCLC patients who underwent surgical resection were collected and stratified into training (80%), validation (10%), and testing (10%) datasets. Machine learning models were trained on the training data using clinical parameters including the PD-L1 expression. The top-performing model was assessed on the test data using a SHAP analysis and partial dependence plots to quantify the contribution of the PD-L1 expression. A multivariate Cox proportional hazards model was used to validate the association between the PD-L1 expression and postoperative recurrence. The random forest model demonstrated the highest predictive performance with the SHAP analysis highlighting the PD-L1 expression as an important feature, and the multivariate Cox analysis indicating a significant increase in the risk of postoperative recurrence with each increment in the PD-L1 expression. These findings suggest that variations in the PD-L1 expression may provide valuable information for clinical decision-making in lung cancer treatment strategies. Machine learning Random forest Multivariate Cox proportional hazard model Non-small cell lung cancer Recurrence-free survival Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Non-small cell lung cancer (NSCLC) remains a highly lethal malignancy worldwide 1 . While surgical resection is generally an effective therapeutic approach for NSCLC. Approximately 15–38% of patients who undergo NSCLC resection experience recurrence 2 . Predicting and preventing postoperative recurrence remains a major clinical challenge. Therefore, there is an urgent need to clinically assess the risk of postoperative recurrence of NSCLC and identify novel predictive factors. PD-L1 (programmed death ligand 1) and PD-1 (programmed death 1) are factors associated with the immune evasion mechanism of tumors 3 . PD-L1 expressed on tumor cells interacts with PD-1 expressed on immune cells, suppressing the activity of immune cells and promoting immune evasion by tumor cells 4 . Immune checkpoint inhibitors (ICIs) disrupt this immune evasion mechanism and exert immunological anti-tumor effects 5,6 . Clinical studies reported that in NSCLC, higher PD-L1 expression levels are correlated with increased efficacy of immune checkpoint inhibitors against tumors 7,8 . Paradoxically, this suggests that the increased expression of PD-L1 may lead to greater immune suppression, potentially increasing the risk of postoperative recurrence in NSCLC. Although this hypothesis is supported by several reports, including our previous studies 9,10 . conflicting reports suggest that the expression of PD-L1 is a favorable prognostic factor in NSCLC 11,12 . Thus, a consensus on the role of PD-L1 in the risk of postoperative recurrence in NSCLC remains elusive and further discussion is warranted. To address this gap, we conducted a study to elucidate the association between the risk of postoperative recurrence and the PD-L1 expression in NSCLC. Using machine learning, we constructed a postoperative recurrence prediction model based on the clinical and pathological features of patients who underwent NSCLC resection. By analyzing this prediction model, we evaluated the contribution of differences in the expression of PD-L1 to postoperative recurrence and explored the relationship between the increased PD-L1 expression and increased recurrence risk. The significance of our study is that it provides novel insights into the prediction of postoperative recurrence based on the expression of PD-L1 in NSCLC patients. Unexplored insights from a machine learning approach help improve the accuracy of the prediction of postoperative recurrence and may be useful for developing lung cancer treatment strategies tailored to PD-L1 expression levels. The results of our study could potentially provide new guidelines for recurrence prevention strategies in patients undergoing lung cancer resection. Materials and methods Patients The study population consisted of 647 patients who underwent lung cancer resection at the NHO Kinki Chuo Chest Medical Center (KCMC) between April 2017 and June 2022. Only patients with histologically confirmed pathologically complete resection (R0) were included. Cases with incomplete tumor removal (R1) were excluded. Patient background information was retrospectively collected from the electronic medical records. Patients whose electronic medical records were unavailable were also excluded. Histopathological diagnoses were performed by our institutional pathologists in accordance with the 2015 World Health Organization classification 13 . Eligible patients who provided their informed consent were treated with platinum-based adjuvant chemotherapy based on the guidelines of the Japanese Lung Cancer Association. Clinicopathological features, including age, sex, neutrophil-to-lymphocyte ratio (NLR) before surgery, pathological stage (American Joint Committee eighth edition), histological type, vascular invasion (v), lymphatic invasion (Ly), adjuvant chemotherapy, PD-L1 expression, and postoperative recurrence-free survival (RFS), were collected from the medical records as key features for the machine learning model and statistical analyses. Pathological information (i.e., pathological stage, v, Ly, and histological type) was collected based on pathology reports created by experienced pathologists. This study was approved by the KCMC Institutional Ethics Committee (Approval No. 2023-31), which granted retrospective exemption from obtaining informed consent from all study participants, considering the anonymous nature of the study. The KCMC website offers information on opting out of participation in the study. All research methods adhered to the applicable guidelines and regulations. The RFS The primary endpoint of this study was the assessment of postoperative recurrence through a clinical diagnosis. Recurrence-free survival (RFS) was defined as the period between lung cancer resection and clinical confirmation of recurrence. We considered the patients to be recurrence-free until it was definitively confirmed after surgery. Patients who received surgery underwent regular blood sampling and X-ray examinations every three–six months. In cases where any abnormal findings suggestive of disease recurrence were observed, additional diagnostic tests such as magnetic resonance imaging (MRI) of the head, contrast-enhanced computed tomography (CT), positron emission tomography (PET), and pathological examination of tissue biopsy samples were conducted. Recurrence was diagnosed through a comprehensive evaluation of these test results during joint conferences involving general thoracic surgeons, oncologists, pathologists, and radiologists. PD-L1 immunohistochemistry A pathologist examined all cancer cells detectable in the tissue samples extracted from the resected lung cancer specimens. The PD-L1 clone 22C3 pharmDx kit was used in conjunction with the Dako Automated Link 48 platform (Agilent Technologies, Santa, Clara, CA, USA) for immunohistochemical analysis to assess the expression of programmed death-ligand 1 (PD-L1). The tumor proportion score (TPS) for PD-L1 was calculated as the percentage of membranous staining, either complete or partial, observed in the tissue samples. The score ranges from 0–100%, and was calculated following the standard 22C3 assay protocol. The tumor region was visually segmented into four areas, and the proportion of PD-L1 positive cells in each area was quantified, resulting in an average value for the clinical TPS. Machine learning We used a machine learning approach to develop the postoperative recurrence prediction model for lung cancer. We chose machine learning algorithms with high predictive accuracy, including bagging and boosting techniques from the ensemble learning family, as well as random forest, gradient boosting, light gradient boosting, and Ada boosting 14,15 . We selected variables for the machine learning algorithm based on factors previously reported to correlate with postoperative recurrence, in addition to the expression of PD-L1. The following variables were selected: the expression of PD-L1 (TPS), pathological stage (stage I–III), invasion of blood vessels (v0–v1) 16 , lymphatic vessel invasion (Ly0–Ly1) 17 , histological classification of cancer (adenocarcinoma, squamous cell carcinoma, or others), neutrophil-to-lymphocyte ratio (NLR) 18 , adjuvant chemotherapy 19 , age, and sex. We assessed the correlations of individual variables within the machine learning model through a heat map analysis. To prevent overfitting, the entire dataset was randomly divided into three parts: a training set, validation set, and test set. The training set comprised 80% of the dataset, while the validation and test sets comprised 10% each. Python packages (RandomForestClassifier, lgb.LGBMClassifier, GradientBoostingClassifier, and AdaBoostClassifier) were used to construct the machine learning models. Training data were employed to train the model and explore hyperparameters. By employing optimal hyperparameters, the performance of the model was assessed using validation data. Finally, a conclusive evaluation of the model was conducted using the test data. We employed Bayesian hyperparameter optimization using the Python library Optuna, to select the hyperparameters for each machine learning model designed to predict postoperative recurrence. For each model, we conducted a series of 100 trials to determine the optimal hyperparameter configuration. Subsequently, we selected the hyperparameters linked to the most effective configuration as the optimal settings for each of the investigated models (Table S1 ). The predictive performance of each model was evaluated quantitatively using metrics such as accuracy, F1 score, Brier score, receiver operating characteristic area under the curve (ROC AUC), and precision-recall AUC (PR AUC). The model with the best predictive performance was used as the final model. The predictions were interpreted through the application of Shapley Additive Explanations (SHAP) values 20 . SHAP values are derived from Shapley values in coalition game theory, offering a robust and precise approach to quantifying the impact of individual variables on the predictions of a machine learning model. SHAP values were computed using the shap v0.28.5 Python library and visualizations were generated using Matplotlib v3.0.311. To investigate the influence of the PD-L1 expression on the model's predictions, we utilized the plot_partial_dependence function from the scikit-learn Python package and visualized the influence using partial dependence plots. Statistical Analysis To evaluate the contribution of the PD-L1 expression to postoperative recurrence, we used a multivariate Cox proportional hazards model. In this analysis, the outcome variable was postoperative recurrence, and the explanatory variable of interest was the PD-L1 expression in the resected lung cancer tissues. The PD-L1 expression was treated as a continuous variable that was independent of the threshold. Features, other than the PD-L1 expression used in the machine learning model were incorporated as confounders. It is important to note that in a multivariate Cox proportional hazards model, the number of variables that can be included is limited by the number of outcome events. In our study, the number of recurrence cases (151), allowed for up to 15 variables to be included in the model 21 . The primary variable in our study was the expression of PD-L1. We selected ten other confounding variables in our model. Specifically, we considered age (categorized as ≥ 65 and < 65 years, with < 65 years as the reference group), sex (male and female, with male as the reference), pathological stage (categorized as stage I–III, with stage I as the reference), histological type (categorized as adenocarcinoma, squamous cell carcinoma and others, with adenocarcinoma as the reference), invasion of lymphatic vessels (categorized as Ly0–Ly1, with Ly0 as the reference), invasion of blood vessels (categorized as v0–v1, with v0 as the reference), NLR (continuous variable), and adjuvant chemotherapy. We assessed the validity of the proportional hazard assumption in the Cox models by examining martingale residual plots. Multicollinearity among the variables in the multivariable Cox proportional hazards model was evaluated using a variance inflation factor (VIF) with a threshold of < 2 to determine its presence or absence. Statistical analyses were performed using Easy R (EZR) (Saitama Medical Center, Saitama, Japan), which is a graphical user interface of R (The R Foundation for Statistical Computing, Vienna, Austria) 22 . EZR is an improved version of the R commander with additional biostatistical functions. P values of < 0.05 were considered to indicate statistical significance. Results Patient characteristics The cohort consisted of 647 patients who were randomly stratified into training (517 patients, 80%), validation (65 patients, 10%) and test groups (65 patients, 10%). The incidence rate of postoperative recurrence was 23% and 151 patients were observed. The median PD-L1 expression in resected lung cancer tissue for the entire cohort was 5% (range, 0–39%). In the cohort, 73% were ≥ 65 years of age, and 59% were male. The pathological subtypes of lung cancer identified in our study included adenocarcinoma (74%, 478 cases), squamous cell carcinoma (18%, 116 cases), and other subtypes (8%, 53 cases). The pathological stages were as follows: stage I, 71%; stage II, 17%; and stage III, 12%. The median RFS was 729 days (range: 360–1157 days). No statistically significant differences were observed among the three subgroups for any of the variables, indicating a homogeneous distribution (Table 1 ). In the total cohort, patients were classified into three groups based on their PD-L1 expression levels: no expression (< 1%), low expression (1–49%), and high expression (50–100%). The probability of postoperative RFS in these three groups was compared using Kaplan-Meier curves. The analysis revealed that RFS was significantly shorter in the high-, low, and no-expression groups (Fig. 1 ). Table 1 Clinical characteristics and outcomes of 647 patients undergoing lung cancer resection Characteristic Total cohort (n = 647) Training cohort (n = 517) Validation cohort (n = 65) Test cohort (n = 65) P value Continuous variables, median (Q1, Q3) PD-L1 expression (TPS [%]) 5 (0–39) 5 (0–40) 5 (0–40) 3 (0–20) 0.83 NLR 1.8 (1.4–2.4) 1.8 (1.4–2.5) 1.8 (1.4–2.2) 1.9 (1.4–2.5) 0.79 Number of recurrent cases 151 ( 23 ) 129 ( 25 ) 10 ( 15 ) 12 ( 19 ) 0.14 RFS (day) 729 (360–1157) 730 (363–1172) 736 (336–1292) 647 (368–1105) 0.86 Categorical variables, n (%) Age (≥ 65) 472 (73) 380 (74) 50 (77) 42 (65) 0.24 Male sex 381 (59) 301 (58) 39 (60) 41 (63) 0.74 Histological types 0.35 Adenocarcinoma 478 (74) 383 (74) 43 (66) 52 (80) Squamous cell carcinoma 116 ( 18 ) 90 ( 17 ) 17 ( 26 ) 9 ( 14 ) Others † 53 ( 8 ) 44 ( 9 ) 5 ( 8 ) 4 ( 6 ) Pathological stage 0.98 Stage I 456 (71) 364 (70) 45 (69) 47 (72) Stage II 112 ( 17 ) 90 ( 17 ) 11 ( 17 ) 11 ( 17 ) Stage III 79 ( 12 ) 63 ( 12 ) 9 ( 14 ) 7 ( 11 ) Vascular invasion 0.89 v0 200 ( 31 ) 162 ( 31 ) 19 ( 29 ) 19 ( 29 ) v1 447 (69) 355 (69) 46 (71) 46 (71) Lympho-vessel invasion 0.35 Ly0 246 (38) 193 (37) 23 (35) 30 (46) Ly1 401 (62) 324 (63) 42 (65) 35 (54) Adjuvant chemotherapy 66 ( 10 ) 48 ( 9 ) 11 ( 17 ) 7 ( 11 ) 0.16 † Histological types excluding adenocarcinoma and squamous cell carcinoma in non-small-cell lung cancer were as follows: among 53 cases, 17 had pleomorphic carcinoma, 16 had large-cell neuroendocrine carcinoma, 12 had adenosquamous carcinoma, 5 had large-cell carcinoma, and 3 had carcinoid tumors. Evaluation of machine learning models The correlation between the features to be input into the machine learning model was examined using a heatmap (Fig. 2 ). It was observed that the correlation coefficients between features from different categories were all < 0.7, which suggests that there was no strong correlation. Machine learning models, including random forest, gradient boosting, light gradient boosting and Ada boosting algorithms, were trained using the training dataset for postoperative recurrence prediction. The performance of each model was then assessed using a validation dataset. Model performance was assessed by measuring the ROC AUC, PR AUC, accuracy, F1 score and Brier score. Among all prediction models, the random forest model showed the highest performance with an ROC AUC of 0.97 (95% CI 0.90–1.00), PR AUC of 0.86 (95% CI 0.72–1.00), Accuracy of 0.92 (95% CI 0.81–1.00), F1 score of 0.76 (95% CI 0.58–0.94) and Brier score of 0.08 (95% CI 0.01–0.16) (Table 2 ). Based on the overall performance, the random forest model was selected as the superior model. Finally, the performance of the random forest model was assessed on the test data and showed the following predictive accuracy: ROC AUC of 0.90 (95% CI 0.78–1.00), PR AUC of 0.78 (95% CI 0.61–0.94), accuracy of 0.92 (95% CI 0.81–1.00), F1 score of 0.74 (95% CI 0.55–0.92) and Brier score of 0.10 (95% CI 0.02–0.18) (Table 3 ). Table 2 Predictive performance of various machine learning models in identifying postoperative lung cancer recurrence in the validation cohort Machine learning model Assessment metrics (95% CI) ROC AUC PR AUC Accuracy F1 score Brier score Random forest 0.97 (0.90–1.00) 0.86 (0.72–1.00) 0.92 (0.81–1.00) 0.76 (0.58–0.94) 0.08 (0.01–0.16) Gradient boosting 0.92 (0.81–1.00) 0.73 (0.54–0.92) 0.89 (0.76–1.00) 0.67 (0.47–0.86) 0.08 (0.01–0.14) Light gradient boosting 0.93 (0.82–1.00) 0.62 (0.42–0.82) 0.89 (0.76–1.00) 0.67 (0.47–0.86) 0.07 (0.01–0.15) Ada boosting 0.93 (0.82–1.00) 0.61 (0.41–0.81) 0.89 (0.76–1.00) 0.64 (0.44–0.83) 0.21 (0.08–0.34) Abbreviations: CI, confidence interval; PR AUC, area under the precision-recall curve; ROC AUC, area under the receiver operating characteristic curve. Table 3 Predictive performance of a random forest model in identifying postoperative lung cancer recurrence in the test cohort Machine learning model Assessment metrics (95% CI) ROC AUC PR AUC Accuracy F1 score Brier score Random forest 0.90 (0.78–1.00) 0.78 (0.61–0.94) 0.92 (0.81–1.00) 0.74 (0.55–0.92) 0.10 (0.02–0.18) Abbreviations: CI, confidence interval; PR AUC, area under the precision-recall curve; ROC AUC, area under the receiver operating characteristic curve. Importance and dependence of features in machine learning predictions To assess the importance of each feature used in the prediction model, a SHAP analysis was performed using a random forest model. Among the features analyzed, the average SHAP values were highest for RFS, pathological stage I, pathological stage III, and vascular invasion, followed by PD-L1 expression (Fig. 3 A). The distribution of the SHAP scores was also analyzed for each feature (Fig. 3 B). With respect to the PD-L1 expression, the red dots were predominantly located in the positive SHAP value region, while the blue dots were located in the negative SHAP value region. Thus, an increase in the expression of PD-L1 was associated with an increase in its contribution to model predictions. To further explore the impact of the PD-L1 expression on the model predictions, partial dependence plots were generated (Fig. 4 ). The dependence on predictions greatly increased when PD-L1 was expressed in comparison to when it was not, and a subsequent linear increase in dependency was observed with increasing PD-L1 expression levels. Multivariate Cox proportional hazards analysis We then examined the association between PD-L1 expression and postoperative recurrence using a statistical approach, specifically the multivariate Cox proportional hazards model. The results of the multivariate Cox proportional hazards analysis are presented in Table 4 . The multicollinearity of each explanatory variable was assessed using the variance inflation factor (VIF). The VIF was less than 2 for all variables, indicating the absence of multicollinearity (Table S2 ). A proportional hazards analysis was performed using a martingale residual plot. The generally horizontal trend of this curve confirmed the fulfilment of the proportional hazards assumption (Figure S1 ). An increase in the expression of PD-L1 was associated with a significantly increased risk of postoperative recurrence (adjusted hazard ratio, 1.006; 95% CI 1.000–1.011). Table 4 Results of a multivariate Cox proportional hazard analysis of RFS according to the expression of PD-L1. Unadjusted HR (95% CI), P † Adjusted HR (95% CI), P PD-L1 expression (TPS [%]) 1.013 (1.008–1.018), < 0.001 1.006 (1.000–1.011), 0.04 † In the multivariate analysis, the hazard ratio was adjusted for age, sex, histological type, pathological stage, vascular invasion, lymphovessel invasion, and adjuvant chemotherapy. Abbreviations: CI, confidence interval; HR, hazard ratio; PD-L1, programmed death ligand 1; RFS, recurrence-free survival. Discussion We conducted a single-center retrospective observational study of 647 postoperative NSCLC patients to investigate the association between the expression of PD-L1 and postoperative recurrence. In our cohort, the recurrence rate was 23.3%, consistent with the previously reported rates of 20–26% in larger cohorts studying lung cancer 23,24 . Our study showed that groups with higher PD-L1 expression levels had shorter RFS in the conventional classification based on PD-L1 expression levels (no expression [< 1%], low expression [1–49%], and high expression [50–100%]). In addition, using a machine learning model with a random forest algorithm and a multivariate Cox proportional hazards model with a statistical analysis, we investigated the impact of the PD-L1 expression on the postoperative recurrence of NSCLC. Our results showed a nonlinear increase in the risk of postoperative recurrence based on the PD-L1 expression level. Including our previous study and several conventional studies, there have been reports linking the PD-L1 expression in NSCLC to an increased risk of postoperative recurrence 10,25–28 . However, the results of our study may help to further develop this association. In previous studies, a statistical approach demonstrated an association between the expression of PD-L1 in NSCLC and postoperative recurrence. This association was similarly observed in the multivariate Cox proportional hazards analysis in our study, which is consistent with previous research. In contrast to previous studies, our study statistically demonstrated an increase in postoperative recurrence risk corresponding to each 1% increase in the PD-L1 expression level by conducting statistical analysis treating the PD-L1 expression as a continuous variable rather than as a categorical variable. Furthermore, we used a machine learning approach to evaluate the detailed effect of the PD-L1 expression on postoperative recurrence. Specifically, we revealed in detail the dynamic changes in postoperative recurrence depending on the PD-L1 expression level using a SHAP analysis and partial dependence plots. Our machine-learning approach has discovered a new finding that has not been reported in previous statistical analyses. This finding highlights the potential for dynamic variations in postoperative recurrence of NSCLC associated with PD-L1 expression levels. In conventional studies, the relationship between the expression of PD-L1 and the postoperative prognosis has been examined using categorical classifications such as no expression (< 1%), low expression (1–49%), and high expression (≥ 50%). However, our study, leveraging machine learning techniques, revealed a more nuanced, continuous association between PD-L1 expression levels and recurrence risk that could not be captured by these traditional categories. The nonlinear increase in the risk of recurrence with even minimal PD-L1 levels (as low as 1%) and the linear escalation of risk when the expression increased beyond 1% suggest the importance of considering the PD-L1 expression as a continuous variable rather than discrete category. This finding underscores the potential for dynamic variations in the postoperative recurrence risk across the spectrum of PD-L1 expression, providing a more granular understanding of the relationship between the expression of PD-L1 and the recurrence of NSCLC. The results of this study may contribute to its clinical application. It has been reported that adjuvant chemotherapy with immune checkpoint inhibitors during the perioperative period of lung cancer improves a patient’s prognosis, even with PD-L1 expression levels as low as 1%, and this effect becomes more pronounced when the expression levels exceeds 50% 29 . The nonlinear and dynamic changes in the risk of recurrence based on PD-L1 expression levels may encourage the active introduction of perioperative immune checkpoint inhibitors for PD-L1-positive lung cancer patients as a clinical decision-making strategy. The results of this study can be explained from an immunological perspective. We showed that the contribution to recurrence increased non-linearly and sharply when PD-L1 was expressed, even at 1%, in comparison to PD-L1-negative cases. This trend suggests that, even the minimal expression of PD-L1 may have a significant impact on postoperative recurrence in NSCLC. Even when cancer cells express PD-L1 at 1%, subtle interactions between PD-L1-expressing cancer cells and the surrounding immune cells may lead to local immune suppression. This local suppression of immune cell activity may increase the immune resistance of cancer cells, activate local immune escape mechanisms and potentially increase the risk of postoperative recurrence. This possibility is supported by clinical studies showing the higher efficacy of immune checkpoint inhibitors in cases where PD-L1 is expressed, even at 1%, in comparison to cases with no expression 30,31 . In contrast, in cases with PD-L1 expression levels of ≥ 1%, a linear increase in the contribution to postoperative recurrence was observed with increasing expression levels. This mechanism suggests that, as the number of cancer cells expressing PD-L1 increases, the suppression of immune cell activity increases, resulting in more immune-resistant cancer cells and further activation of immune escape mechanisms. This possibility is supported by clinical trials that demonstrate the higher efficacy of immune checkpoint inhibitors in patients with high PD-L1 expression levels (≥ 50%) in comparison to those with low expression levels (1–49%) and PD-L1-negative cases (< 1%) 31–33 . In addition to direct interactions with immune cells, the involvement of PD-L1 in the formation of the tumor microenvironment could be a potential mechanism by which different PD-L1 expression levels exert a nonlinear influence on the risk of recurrence. PD-L1 is capable of nuclear translocation and has been reported to directly bind to DNA and regulate the transcriptional induction of genes involved in the tumor microenvironment, such as immune responses and inflammation. In other words, when PD-L1 is expressed, the higher its expression level, the more likely it is that signaling pathways associated with tumor immune evasion are activated, potentially contributing to the establishment of an immunosuppressive tumor microenvironment. The formation of the tumor microenvironment may be closely related to the survival and proliferation of residual cancer cells, possibly increasing the risk of postoperative recurrence 34 . The dynamic changes in the increased risk of recurrence as a function of PD-L1 expression levels revealed in our study suggest significant variations in the risk of recurrence, even within the PD-L1 expression categories of low (1–49%) and high (50–100%) that are commonly used in clinical practice. Our study is the first to reveal the machine-learning-based dynamic variation of postoperative recurrence risk dependent on the continuous range of PD-L1 expression levels from low to high expression. This study was associated with several limitations. First, as this was a retrospective observational study conducted at a single institution, caution should be exercised in generalizing the results. Given its retrospective nature, it is a possible that clinical factors other than the expression of PD-L1 were not considered. Future validation through additional studies, such as prospective investigations or multicenter collaborations, is essential to elucidate the impact of these factors. Second, the performance of machine learning models is strictly limited by predictive accuracy and requires careful consideration in clinical applications. The results of this study, derived from a limited dataset, require further validation in clinical research using other cohorts or larger datasets to determine whether similar trends can be observed in recurrence prediction models. Overall, these considerations highlight the need for cautious interpretation and for future research to improve the robustness and applicability of our findings. Conclusion In conclusion, our study using machine learning and statistical analysis revealed a significant, nonlinear association between the expression of PD-L1 and the risk of postoperative recurrence in NSCLC. We demonstrated that even minimal PD-L1 expression levels (as low as 1%) were associated with an increased risk of recurrence, suggesting the potential impact of subtle immune interactions. Furthermore, a continuous increase in the expression of PD-L1 beyond 1% corresponded to a linear increase in the risk of recurrence. These novel findings suggest a dynamic relationship between the expression of PD-L1 and postoperative recurrence, and provide valuable insights into personalized therapeutic strategies for NSCLC. Our findings highlight the importance of considering the continuum of PD-L1 expression levels in real-world clinical practice, as opposed to the traditional classification of PD-L1 expression levels into low versus high. This study contributes to a better understanding of the immunological factors influencing the recurrence of NSCLC, paving the way for tailored treatment interventions based on the expression of PD-L1. Declarations Author contributions K.K. conceived and designed the study. K.K. and H.S. collected the patient data. K.K. performed the data analysis. K.K., H.S., T.I., T.T., K.O. and H.Y. interpreted the analyzed data. K.K. wrote the manuscript. All authors reviewed and approved the final manuscript. Data availability The database used in this study is not available to the public. Participants in our study did not agree for their data to be publicly shared. The Python code used in this study is available on request from the corresponding author, [email protected] , upon reasonable request. Funding Kensuke Kojima received a specific grant from Grant-in-Aid for Clinical Research from the Osaka Foundation for the Prevention of Cancer and Cardiovascular Disease, a public funding agency. Conflict of interest Authors have no conflict of interest to declare. Ethical approval statement This study was approved by the Institutional Review Board (IRB) of the National Hospital Organization Kinki Chuo Chest Medical Center (KCMC) (approval number: 2023-31) 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. Patient consent statement Due to the retrospective nature of this study, informed consent from the patients was not required. References Sung, H. et al . Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin . 71 , 209–249 (2021). Subotic, D., Van, Schil, P. & Grigoriu, B. Optimising treatment for post-operative lung cancer recurrence. Eur Respir J . 47 , 374–378 (2016). Reck, M. et al . Pembrolizumab versus chemotherapy for PD-L1-positive non-small-cell lung cancer. N Engl J Med . 375 , 1823–1833 (2016). Han, Y., Liu, D. & Li, L. PD-1/PD-L1 pathway: current researches in cancer. Am J Cancer Res . 10 , 727–742 (2020). Daud, AI. et al . Programmed death-ligand 1 expression and response to the anti-programmed death 1 antibody pembrolizumab in melanoma. J Clin Oncol . 3 , 4102–4109 (2016). Garon, EB. et al . Pembrolizumab for the treatment of non-small-cell lung cancer. N Engl J Med . 372 , 2018–2028 (2015). Rosenberg, JE. et al . Atezolizumab in patients with locally advanced and metastatic urothelial carcinoma who progressed following treatment with platinum-based chemotherapy: a single arm, phase 2 trial. Lancet . 387 , 1909–1920 (2016). Brahmer, J. et al . Nivolumab versus docetaxel in advanced squamous-cell non-small-cell lung cancer. N Engl J Med . 373 , 123–135 (2015). Sun, JM. et al . Prognostic Significance of PD-L1 in Patients with Non-Small-Cell Lung Cancer: A Large Cohort Study of Surgically Resected Cases. J Thorac Oncol . 11 , 1003–1011 (2016). Kojima, K., Sakamoto, T., Kasai, T., Kagawa, T., Yoon, H. & Atagi, S. PD-L1 expression as a predictor of postoperative recurrence and the association between the PD-L1 expression and EGFR mutations in NSCLC. Sci Rep . 11 , 17522 (2021). Cooper, WA. et al . PD-L1 expression is a favorable prognostic factor in early stage non-small cell carcinoma. Lung Cancer . 89 , 181–188 (2015). Motono, N., Mizoguchi, T., Ishikawa, M., Iwai, S., Iijima, Y. & Uramoto, H. PD-L1 expression in not a predictive factor for recurrence in resected non-small cell lung cancer. Lung . 201 , 95–101 (2023). Travis, WD. et al . The 2015 world health organization classification of lung tumors: Impact of genetic, clinical and radiologic advances since the 2004 classification. J Thorac Oncol . 10 , 1243–1260 (2015). Rahman, S., Irfan, M., Raza, M., Ghori, K. M., Yaqoob, S. & Awais, M. Performance analysis of boosting classifiers in recognizing activities of daily living. Int J Environ Res Public Health . 17 , 1082 (2020). Patel, H., Vock, D. M., Marai, G. E., Fuller, C. D., Mohamed, A. S. R. & Canahuate, G. Oropharyngeal cancer patient stratification using random forest based-learning over high-dimensional radiomic features. Sci Rep . 11 , 14057 (2021). Vieira, T. et al . Blood vessel invasion is a major feature and a factor of poor prognosis in sarcomatoid carcinoma of the lung. Lung Cancer . 85 , 276–281 (2014). Al-Alao, B. S., Gately, K., Nicholson, S., McGovern, E., Young, V. K. & O’Byrne, K. J. Prognostic impact of vascular and lymphovascular invasion in early lung cancer. Asian Cardiovasc Thorac Ann . 22 , 55–64 (2014). Mizuguchi, S., Izumi, N., Tsukioka, T., Komatsu, H. & Nishiyama, N. Neutrophil-lymphocyte ratio predicts recurrence in patients with resected stage 1 non-small cell lung cancer. J Cardiothorac Surg . 13 , 78 (2018). Pirker, R. Adjuvant chemotherapy in patients with completely resected non-small cell lung cancer. Trans Lung Cancer Res . 3 , 305–310 (2014). Lundberg, S. M. & Lee, S-I. A unified approach to interpreting model predictions. Proceedings of the 31st international conference on neural information processing systems; 2017. Peduzzi, P., Concato, J., Feinstein, A. R. & Holford, T. R. Importance of events per independent variable in proportional hazards regression analysis. Ⅱ. Accuracy and precision of regression estimates. J Clin Epidemiol . 48 , 1503­–1510 (1995). Kanda, Y. Investigation of the freely available easy-to-use software ‘EZR’ for medical statistics. Bone Marrow Transplant . 48 , 452–458 (2013). Demicheli, R. et al . Recurrence dynamics for non-small-cell lung cancer: effect of surgery on the development of metastases. J Thorac Oncol . 7 , 723–730 (2012). Lou, F., Huang, J., Sima, C. S., Dycoco, J., Rusch, V. & Bach, P. B. Patterns of recurrence and second primary lung cancer in early-stage lung cancer survivors followed with routine computed tomography surveillance. J Thorac Cardiovasc Surg . 145 , 75–81 (2013). Samejima, H., Kojima, K., Fujiwara, A., Tokunaga, T., Okishio, K. & Yoon, H. The combination of PD-L1 expression and the neutrophil-to-lymphocyte ratio as a prognostic factor of postoperative recurrence in non-small-cell lung cancer: a retrospective cohort study. BMC Cancer . 23 , 1107 (2023). Miyazawa, T. et al . PD-L1 expression in non-small-cell lung cancer including various adenocarcinoma subtypes. Ann Thorac Cardiovasc Surg . 25 , 1–9 (2019). Cha, Y. J., Kim, H. R., Lee, C. Y., Cho, B. C. & Shim, H. S. Clinicopathological and prognostic significance of programmed cell death ligand-1 expression in lung adenocarcinoma and its relationship with p53 status. Lung Cancer . 97 , 73–80 (2016). Handa, Y. et al . Prognostic impact of programmed death-ligand 1 and surrounding immune status on stage I lung cancer. Clin Lung Cancer . 21 , e302–e314 (2020). Forde, P. M. et al . Neoadjuvant nivolumab plus chemotherapy in resectable lung cancer. N Engl J Med . 386 , 1973–1985 (2022). Herbst, R. S. et al . Pembrolizumab versus docetaxel for previously treated, PD-L1-positive, advanced non-small-cell lung cancer (KEYNOTE-010): a randomized controlled trial. Lancet . 387 , 1540–1550 (2016). Mok, T. S. K. et al . Pembrolizumab versus chemotherapy for previously untreated, PD-L1-expressing, locally advanced or metastatic non-small-cell lung cancer (KEYNOTE-042): a randomized, open-label, controlled, Phase 3 trial. Lancet . 393 , 1819–1830 (2019). Jotte, R. et al . Atezolizumab in combination with carboplatin and nab-paclitaxel in advanced squamous NSCLC (IMpower131): Results from a randomized phase III trial. J Thorac Oncol . 15 , 1351–1360 (2020). Herbst, R. S. et al . Atezolizumab for first-line treatment of PD-L1-selected patients with NSCLC. N Engl J Med . 383 , 1328–1339 (2020). Gao, Y. et al . Acetylation-dependent regulation of PD-L1 nuclear translocation dictates the efficacy of anti-PD-1 immunotherapy. Nat Cell Biol . 22 , 1064–1075 (2020). Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 03 Jun, 2024 Reviews received at journal 20 May, 2024 Reviewers agreed at journal 10 May, 2024 Reviewers agreed at journal 10 May, 2024 Reviewers agreed at journal 10 May, 2024 Reviewers invited by journal 10 May, 2024 Editor assigned by journal 10 May, 2024 Editor invited by journal 02 May, 2024 Submission checks completed at journal 30 Apr, 2024 First submitted to journal 27 Apr, 2024 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-4334704","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":298892640,"identity":"98285d41-a587-4b6f-be0e-6ca1ed19b5eb","order_by":0,"name":"Kensuke Kojima","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA1UlEQVRIie3PMQrCMBTG8VeEZIm4phT1Ci2BLgpepUGoq2PHiBDHroqX8ArysC4eoOKiFJwLjnZQxE2J7eaQ//KW9xs+AJvtD+P96+v6QGcKzgCO+kkgfhO2VRA1I/z5X4u4KpPFtELRcYtZGSXYVRR3GxPxHI1ipTF013LOowMKxeI4N5Fei2qvrXDon6QGqVEqzkIzIXRxZxUOR8dtTeIxkrUYwdDnTk3iLsnYa+uJ4IfXlonQv7bwnAQ3Vg2CdLG/lGUy6KYUMyP5jDR7t9lsNtu3Hi+9StwETLU8AAAAAElFTkSuQmCC","orcid":"","institution":"NHO Kinki Chuo Chest Medical Center","correspondingAuthor":true,"prefix":"","firstName":"Kensuke","middleName":"","lastName":"Kojima","suffix":""},{"id":298892642,"identity":"be602556-589e-485b-86b9-20ee50f56ef4","order_by":1,"name":"Hironobu Samejima","email":"","orcid":"","institution":"NHO Kinki Chuo Chest Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Hironobu","middleName":"","lastName":"Samejima","suffix":""},{"id":298892644,"identity":"b738db98-6fa0-49b4-a6f3-8ac73a7d78f7","order_by":2,"name":"Takafumi Iguchi","email":"","orcid":"","institution":"NHO Kinki Chuo Chest Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Takafumi","middleName":"","lastName":"Iguchi","suffix":""},{"id":298892646,"identity":"8ce82dea-9227-47b8-9dc1-8a32b8108c09","order_by":3,"name":"Toshiteru Tokunaga","email":"","orcid":"","institution":"NHO Kinki Chuo Chest Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Toshiteru","middleName":"","lastName":"Tokunaga","suffix":""},{"id":298892648,"identity":"a93fdafa-bf93-48d3-8c11-4be93d58f121","order_by":4,"name":"Kyoichi Okishio","email":"","orcid":"","institution":"NHO Kinki Chuo Chest Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Kyoichi","middleName":"","lastName":"Okishio","suffix":""},{"id":298892650,"identity":"6b0dbff5-4ef8-4c15-8486-d0e7a82e75b8","order_by":5,"name":"Hyungeun Yoon","email":"","orcid":"","institution":"NHO Kinki Chuo Chest Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Hyungeun","middleName":"","lastName":"Yoon","suffix":""}],"badges":[],"createdAt":"2024-04-27 15:24:22","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4334704/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4334704/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":56121973,"identity":"e56e8bd6-9675-4c02-a6e8-ab3492121cd0","added_by":"auto","created_at":"2024-05-08 20:30:23","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":221124,"visible":true,"origin":"","legend":"\u003cp\u003eKaplan-Meier curve showing the probability of recurrence-free survival among patients after lung cancer resection according to the PD-L1 expression levels.\u003c/p\u003e","description":"","filename":"FIGURE1.png","url":"https://assets-eu.researchsquare.com/files/rs-4334704/v1/cee3f5c7ab97a15d7eb6aa85.png"},{"id":56121972,"identity":"c0468f15-bf25-4d7d-95e7-2c99978d917c","added_by":"auto","created_at":"2024-05-08 20:30:23","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1826932,"visible":true,"origin":"","legend":"\u003cp\u003eHeatmap showing the correlation coefficients between different features in a dataset. Colors range from positive correlation (red) to negative correlation (blue).\u003c/p\u003e","description":"","filename":"FIGURE2.png","url":"https://assets-eu.researchsquare.com/files/rs-4334704/v1/c670eb635ff719851d875245.png"},{"id":56121974,"identity":"5f740257-5e10-4d9c-9681-39ea3ce7c47f","added_by":"auto","created_at":"2024-05-08 20:30:23","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":228865,"visible":true,"origin":"","legend":"\u003cp\u003e(A) Variables with the most significant impact on the prediction of postoperative recurrence ranked in order of importance.\u003c/p\u003e\n\u003cp\u003e(B) The distribution of the influence of each variable on the prediction of postoperative recurrence is presented. The numerical characteristics of the variables are visually represented by colors, with larger values shown in red and smaller values shown in blue. Negative SHAP values (spread to the left) suggest a decrease in the probability of postoperative recurrence, while positive values (spread to the right) suggest an increase in the probability.\u003c/p\u003e","description":"","filename":"FIGURE3.png","url":"https://assets-eu.researchsquare.com/files/rs-4334704/v1/786fccc8d3905f5e1a2aad40.png"},{"id":56121975,"identity":"d7f057cc-bcb4-4c43-a206-8ec424a86a3e","added_by":"auto","created_at":"2024-05-08 20:30:23","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":159724,"visible":true,"origin":"","legend":"\u003cp\u003eA partial dependence plot depicting the association between the PD-L1 expression levels and the prediction of postoperative recurrence. The x-axis represents the PD-L1 expression levels, while the y-axis represents the contribution to the prediction of postoperative recurrence (partial dependency).\u003c/p\u003e","description":"","filename":"FIGURE4.png","url":"https://assets-eu.researchsquare.com/files/rs-4334704/v1/2fc9fcc96fdd30751b6ebbf3.png"},{"id":56121971,"identity":"722fc657-4b1b-47f2-ba62-24ecac1b43c5","added_by":"auto","created_at":"2024-05-08 20:30:22","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":678967,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4334704/v1/da84b6f0-5076-4353-af04-b37077ece997.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Dynamic changes in postoperative risk of recurrence of non-small cell lung cancer according to variations in PD-L1 expression levels","fulltext":[{"header":"Introduction","content":"\u003cp\u003eNon-small cell lung cancer (NSCLC) remains a highly lethal malignancy worldwide\u003csup\u003e1\u003c/sup\u003e. While surgical resection is generally an effective therapeutic approach for NSCLC. Approximately 15\u0026ndash;38% of patients who undergo NSCLC resection experience recurrence\u003csup\u003e2\u003c/sup\u003e. Predicting and preventing postoperative recurrence remains a major clinical challenge. Therefore, there is an urgent need to clinically assess the risk of postoperative recurrence of NSCLC and identify novel predictive factors.\u003c/p\u003e \u003cp\u003ePD-L1 (programmed death ligand 1) and PD-1 (programmed death 1) are factors associated with the immune evasion mechanism of tumors\u003csup\u003e3\u003c/sup\u003e. PD-L1 expressed on tumor cells interacts with PD-1 expressed on immune cells, suppressing the activity of immune cells and promoting immune evasion by tumor cells\u003csup\u003e4\u003c/sup\u003e. Immune checkpoint inhibitors (ICIs) disrupt this immune evasion mechanism and exert immunological anti-tumor effects\u003csup\u003e5,6\u003c/sup\u003e. Clinical studies reported that in NSCLC, higher PD-L1 expression levels are correlated with increased efficacy of immune checkpoint inhibitors against tumors\u003csup\u003e7,8\u003c/sup\u003e. Paradoxically, this suggests that the increased expression of PD-L1 may lead to greater immune suppression, potentially increasing the risk of postoperative recurrence in NSCLC. Although this hypothesis is supported by several reports, including our previous studies\u003csup\u003e9,10\u003c/sup\u003e. conflicting reports suggest that the expression of PD-L1 is a favorable prognostic factor in NSCLC\u003csup\u003e11,12\u003c/sup\u003e. Thus, a consensus on the role of PD-L1 in the risk of postoperative recurrence in NSCLC remains elusive and further discussion is warranted.\u003c/p\u003e \u003cp\u003eTo address this gap, we conducted a study to elucidate the association between the risk of postoperative recurrence and the PD-L1 expression in NSCLC. Using machine learning, we constructed a postoperative recurrence prediction model based on the clinical and pathological features of patients who underwent NSCLC resection. By analyzing this prediction model, we evaluated the contribution of differences in the expression of PD-L1 to postoperative recurrence and explored the relationship between the increased PD-L1 expression and increased recurrence risk.\u003c/p\u003e \u003cp\u003eThe significance of our study is that it provides novel insights into the prediction of postoperative recurrence based on the expression of PD-L1 in NSCLC patients. Unexplored insights from a machine learning approach help improve the accuracy of the prediction of postoperative recurrence and may be useful for developing lung cancer treatment strategies tailored to PD-L1 expression levels. The results of our study could potentially provide new guidelines for recurrence prevention strategies in patients undergoing lung cancer resection.\u003c/p\u003e"},{"header":"Materials and methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003ePatients\u003c/h2\u003e \u003cp\u003eThe study population consisted of 647 patients who underwent lung cancer resection at the NHO Kinki Chuo Chest Medical Center (KCMC) between April 2017 and June 2022. Only patients with histologically confirmed pathologically complete resection (R0) were included. Cases with incomplete tumor removal (R1) were excluded. Patient background information was retrospectively collected from the electronic medical records. Patients whose electronic medical records were unavailable were also excluded. Histopathological diagnoses were performed by our institutional pathologists in accordance with the 2015 World Health Organization classification\u003csup\u003e13\u003c/sup\u003e. Eligible patients who provided their informed consent were treated with platinum-based adjuvant chemotherapy based on the guidelines of the Japanese Lung Cancer Association. Clinicopathological features, including age, sex, neutrophil-to-lymphocyte ratio (NLR) before surgery, pathological stage (American Joint Committee eighth edition), histological type, vascular invasion (v), lymphatic invasion (Ly), adjuvant chemotherapy, PD-L1 expression, and postoperative recurrence-free survival (RFS), were collected from the medical records as key features for the machine learning model and statistical analyses. Pathological information (i.e., pathological stage, v, Ly, and histological type) was collected based on pathology reports created by experienced pathologists.\u003c/p\u003e \u003cp\u003eThis study was approved by the KCMC Institutional Ethics Committee (Approval No. 2023-31), which granted retrospective exemption from obtaining informed consent from all study participants, considering the anonymous nature of the study. The KCMC website offers information on opting out of participation in the study. All research methods adhered to the applicable guidelines and regulations.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003eThe RFS\u003c/h2\u003e \u003cp\u003eThe primary endpoint of this study was the assessment of postoperative recurrence through a clinical diagnosis. Recurrence-free survival (RFS) was defined as the period between lung cancer resection and clinical confirmation of recurrence. We considered the patients to be recurrence-free until it was definitively confirmed after surgery. Patients who received surgery underwent regular blood sampling and X-ray examinations every three\u0026ndash;six months. In cases where any abnormal findings suggestive of disease recurrence were observed, additional diagnostic tests such as magnetic resonance imaging (MRI) of the head, contrast-enhanced computed tomography (CT), positron emission tomography (PET), and pathological examination of tissue biopsy samples were conducted. Recurrence was diagnosed through a comprehensive evaluation of these test results during joint conferences involving general thoracic surgeons, oncologists, pathologists, and radiologists.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003ePD-L1 immunohistochemistry\u003c/h2\u003e \u003cp\u003eA pathologist examined all cancer cells detectable in the tissue samples extracted from the resected lung cancer specimens. The PD-L1 clone 22C3 pharmDx kit was used in conjunction with the Dako Automated Link 48 platform (Agilent Technologies, Santa, Clara, CA, USA) for immunohistochemical analysis to assess the expression of programmed death-ligand 1 (PD-L1). The tumor proportion score (TPS) for PD-L1 was calculated as the percentage of membranous staining, either complete or partial, observed in the tissue samples. The score ranges from 0\u0026ndash;100%, and was calculated following the standard 22C3 assay protocol. The tumor region was visually segmented into four areas, and the proportion of PD-L1 positive cells in each area was quantified, resulting in an average value for the clinical TPS.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003eMachine learning\u003c/h2\u003e \u003cp\u003eWe used a machine learning approach to develop the postoperative recurrence prediction model for lung cancer. We chose machine learning algorithms with high predictive accuracy, including bagging and boosting techniques from the ensemble learning family, as well as random forest, gradient boosting, light gradient boosting, and Ada boosting\u003csup\u003e14,15\u003c/sup\u003e. We selected variables for the machine learning algorithm based on factors previously reported to correlate with postoperative recurrence, in addition to the expression of PD-L1. The following variables were selected: the expression of PD-L1 (TPS), pathological stage (stage I\u0026ndash;III), invasion of blood vessels (v0\u0026ndash;v1)\u003csup\u003e16\u003c/sup\u003e, lymphatic vessel invasion (Ly0\u0026ndash;Ly1)\u003csup\u003e17\u003c/sup\u003e, histological classification of cancer (adenocarcinoma, squamous cell carcinoma, or others), neutrophil-to-lymphocyte ratio (NLR)\u003csup\u003e18\u003c/sup\u003e, adjuvant chemotherapy\u003csup\u003e19\u003c/sup\u003e, age, and sex. We assessed the correlations of individual variables within the machine learning model through a heat map analysis. To prevent overfitting, the entire dataset was randomly divided into three parts: a training set, validation set, and test set. The training set comprised 80% of the dataset, while the validation and test sets comprised 10% each.\u003c/p\u003e \u003cp\u003ePython packages (RandomForestClassifier, lgb.LGBMClassifier, GradientBoostingClassifier, and AdaBoostClassifier) were used to construct the machine learning models. Training data were employed to train the model and explore hyperparameters. By employing optimal hyperparameters, the performance of the model was assessed using validation data. Finally, a conclusive evaluation of the model was conducted using the test data. We employed Bayesian hyperparameter optimization using the Python library Optuna, to select the hyperparameters for each machine learning model designed to predict postoperative recurrence. For each model, we conducted a series of 100 trials to determine the optimal hyperparameter configuration. Subsequently, we selected the hyperparameters linked to the most effective configuration as the optimal settings for each of the investigated models (Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e). The predictive performance of each model was evaluated quantitatively using metrics such as accuracy, F1 score, Brier score, receiver operating characteristic area under the curve (ROC AUC), and precision-recall AUC (PR AUC). The model with the best predictive performance was used as the final model. The predictions were interpreted through the application of Shapley Additive Explanations (SHAP) values\u003csup\u003e20\u003c/sup\u003e. SHAP values are derived from Shapley values in coalition game theory, offering a robust and precise approach to quantifying the impact of individual variables on the predictions of a machine learning model. SHAP values were computed using the shap v0.28.5 Python library and visualizations were generated using Matplotlib v3.0.311. To investigate the influence of the PD-L1 expression on the model's predictions, we utilized the plot_partial_dependence function from the scikit-learn Python package and visualized the influence using partial dependence plots.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003eStatistical Analysis\u003c/h2\u003e \u003cp\u003eTo evaluate the contribution of the PD-L1 expression to postoperative recurrence, we used a multivariate Cox proportional hazards model. In this analysis, the outcome variable was postoperative recurrence, and the explanatory variable of interest was the PD-L1 expression in the resected lung cancer tissues. The PD-L1 expression was treated as a continuous variable that was independent of the threshold. Features, other than the PD-L1 expression used in the machine learning model were incorporated as confounders. It is important to note that in a multivariate Cox proportional hazards model, the number of variables that can be included is limited by the number of outcome events. In our study, the number of recurrence cases (151), allowed for up to 15 variables to be included in the model\u003csup\u003e21\u003c/sup\u003e. The primary variable in our study was the expression of PD-L1. We selected ten other confounding variables in our model. Specifically, we considered age (categorized as \u0026ge;\u0026thinsp;65 and \u0026lt;\u0026thinsp;65 years, with \u0026lt;\u0026thinsp;65 years as the reference group), sex (male and female, with male as the reference), pathological stage (categorized as stage I\u0026ndash;III, with stage I as the reference), histological type (categorized as adenocarcinoma, squamous cell carcinoma and others, with adenocarcinoma as the reference), invasion of lymphatic vessels (categorized as Ly0\u0026ndash;Ly1, with Ly0 as the reference), invasion of blood vessels (categorized as v0\u0026ndash;v1, with v0 as the reference), NLR (continuous variable), and adjuvant chemotherapy. We assessed the validity of the proportional hazard assumption in the Cox models by examining martingale residual plots. Multicollinearity among the variables in the multivariable Cox proportional hazards model was evaluated using a variance inflation factor (VIF) with a threshold of \u0026lt;\u0026thinsp;2 to determine its presence or absence.\u003c/p\u003e \u003cp\u003eStatistical analyses were performed using Easy R (EZR) (Saitama Medical Center, Saitama, Japan), which is a graphical user interface of R (The R Foundation for Statistical Computing, Vienna, Austria)\u003csup\u003e22\u003c/sup\u003e. EZR is an improved version of the R commander with additional biostatistical functions. P values of \u0026lt;\u0026thinsp;0.05 were considered to indicate statistical significance.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003ePatient characteristics\u003c/h2\u003e \u003cp\u003eThe cohort consisted of 647 patients who were randomly stratified into training (517 patients, 80%), validation (65 patients, 10%) and test groups (65 patients, 10%). The incidence rate of postoperative recurrence was 23% and 151 patients were observed. The median PD-L1 expression in resected lung cancer tissue for the entire cohort was 5% (range, 0\u0026ndash;39%). In the cohort, 73% were \u0026ge;\u0026thinsp;65 years of age, and 59% were male. The pathological subtypes of lung cancer identified in our study included adenocarcinoma (74%, 478 cases), squamous cell carcinoma (18%, 116 cases), and other subtypes (8%, 53 cases). The pathological stages were as follows: stage I, 71%; stage II, 17%; and stage III, 12%. The median RFS was 729 days (range: 360\u0026ndash;1157 days). No statistically significant differences were observed among the three subgroups for any of the variables, indicating a homogeneous distribution (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). In the total cohort, patients were classified into three groups based on their PD-L1 expression levels: no expression (\u0026lt;\u0026thinsp;1%), low expression (1\u0026ndash;49%), and high expression (50\u0026ndash;100%). The probability of postoperative RFS in these three groups was compared using Kaplan-Meier curves. The analysis revealed that RFS was significantly shorter in the high-, low, and no-expression groups (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eClinical characteristics and outcomes of 647 patients undergoing lung cancer resection\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCharacteristic\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTotal cohort\u003c/p\u003e \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;647)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTraining cohort\u003c/p\u003e \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;517)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eValidation cohort\u003c/p\u003e \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;65)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTest cohort\u003c/p\u003e \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;65)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eP value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"6\" nameend=\"c6\" namest=\"c1\"\u003e \u003cp\u003eContinuous variables, median (Q1, Q3)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePD-L1 expression (TPS [%])\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5 (0\u0026ndash;39)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e5 (0\u0026ndash;40)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5 (0\u0026ndash;40)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e3 (0\u0026ndash;20)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.83\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNLR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.8 (1.4\u0026ndash;2.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.8 (1.4\u0026ndash;2.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.8 (1.4\u0026ndash;2.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.9 (1.4\u0026ndash;2.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.79\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNumber of recurrent cases\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e151 (\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e129 (\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e10 (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e12 (\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.14\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRFS (day)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e729 (360\u0026ndash;1157)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e730 (363\u0026ndash;1172)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e736 (336\u0026ndash;1292)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e647 (368\u0026ndash;1105)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"6\" nameend=\"c6\" namest=\"c1\"\u003e \u003cp\u003eCategorical variables, n (%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge (\u0026ge;\u0026thinsp;65)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e472 (73)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e380 (74)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e50 (77)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e42 (65)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.24\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMale sex\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e381 (59)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e301 (58)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e39 (60)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e41 (63)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.74\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHistological types\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.35\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAdenocarcinoma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e478 (74)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e383 (74)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e43 (66)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e52 (80)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSquamous cell carcinoma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e116 (\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e90 (\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e17 (\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e9 (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOthers\u003csup\u003e\u0026dagger;\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e53 (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e44 (\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5 (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e4 (\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePathological stage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStage I\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e456 (71)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e364 (70)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e45 (69)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e47 (72)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStage II\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e112 (\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e90 (\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e11 (\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e11 (\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStage III\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e79 (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e63 (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e9 (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e7 (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVascular invasion\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.89\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ev0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e200 (\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e162 (\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e19 (\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e19 (\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ev1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e447 (69)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e355 (69)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e46 (71)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e46 (71)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLympho-vessel invasion\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.35\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLy0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e246 (38)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e193 (37)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e23 (35)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e30 (46)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLy1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e401 (62)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e324 (63)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e42 (65)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e35 (54)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAdjuvant chemotherapy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e66 (\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e48 (\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e11 (\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e7 (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.16\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"6\"\u003e\u003csup\u003e\u0026dagger;\u003c/sup\u003eHistological types excluding adenocarcinoma and squamous cell carcinoma in non-small-cell lung cancer were as follows: among 53 cases, 17 had pleomorphic carcinoma, 16 had large-cell neuroendocrine carcinoma, 12 had adenosquamous carcinoma, 5 had large-cell carcinoma, and 3 had carcinoid tumors.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003eEvaluation of machine learning models\u003c/h2\u003e \u003cp\u003eThe correlation between the features to be input into the machine learning model was examined using a heatmap (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). It was observed that the correlation coefficients between features from different categories were all \u0026lt;\u0026thinsp;0.7, which suggests that there was no strong correlation. Machine learning models, including random forest, gradient boosting, light gradient boosting and Ada boosting algorithms, were trained using the training dataset for postoperative recurrence prediction. The performance of each model was then assessed using a validation dataset. Model performance was assessed by measuring the ROC AUC, PR AUC, accuracy, F1 score and Brier score. Among all prediction models, the random forest model showed the highest performance with an ROC AUC of 0.97 (95% CI 0.90\u0026ndash;1.00), PR AUC of 0.86 (95% CI 0.72\u0026ndash;1.00), Accuracy of 0.92 (95% CI 0.81\u0026ndash;1.00), F1 score of 0.76 (95% CI 0.58\u0026ndash;0.94) and Brier score of 0.08 (95% CI 0.01\u0026ndash;0.16) (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). Based on the overall performance, the random forest model was selected as the superior model. Finally, the performance of the random forest model was assessed on the test data and showed the following predictive accuracy: ROC AUC of 0.90 (95% CI 0.78\u0026ndash;1.00), PR AUC of 0.78 (95% CI 0.61\u0026ndash;0.94), accuracy of 0.92 (95% CI 0.81\u0026ndash;1.00), F1 score of 0.74 (95% CI 0.55\u0026ndash;0.92) and Brier score of 0.10 (95% CI 0.02\u0026ndash;0.18) (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePredictive performance of various machine learning models in identifying postoperative lung cancer recurrence in the validation cohort\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eMachine learning model\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"5\" nameend=\"c6\" namest=\"c2\"\u003e \u003cp\u003eAssessment metrics (95% CI)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eROC AUC\u003c/b\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003ePR AUC\u003c/b\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003eAccuracy\u003c/b\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003eF1 score\u003c/b\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003eBrier score\u003c/b\u003e\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.97 (0.90\u0026ndash;1.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.86 (0.72\u0026ndash;1.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.92 (0.81\u0026ndash;1.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.76 (0.58\u0026ndash;0.94)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.08 (0.01\u0026ndash;0.16)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGradient boosting\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.92 (0.81\u0026ndash;1.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.73 (0.54\u0026ndash;0.92)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.89 (0.76\u0026ndash;1.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.67 (0.47\u0026ndash;0.86)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.08 (0.01\u0026ndash;0.14)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLight gradient boosting\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.93 (0.82\u0026ndash;1.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.62 (0.42\u0026ndash;0.82)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.89 (0.76\u0026ndash;1.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.67 (0.47\u0026ndash;0.86)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.07 (0.01\u0026ndash;0.15)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAda boosting\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.93 (0.82\u0026ndash;1.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.61 (0.41\u0026ndash;0.81)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.89 (0.76\u0026ndash;1.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.64 (0.44\u0026ndash;0.83)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.21 (0.08\u0026ndash;0.34)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"6\"\u003eAbbreviations: CI, confidence interval; PR AUC, area under the precision-recall curve; ROC AUC, area under the receiver operating characteristic curve.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePredictive performance of a random forest model in identifying postoperative lung cancer recurrence in the test cohort\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eMachine learning model\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"5\" nameend=\"c6\" namest=\"c2\"\u003e \u003cp\u003eAssessment metrics (95% CI)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eROC AUC\u003c/b\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003ePR AUC\u003c/b\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003eAccuracy\u003c/b\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003eF1 score\u003c/b\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003eBrier score\u003c/b\u003e\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.90 (0.78\u0026ndash;1.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.78 (0.61\u0026ndash;0.94)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.92 (0.81\u0026ndash;1.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.74 (0.55\u0026ndash;0.92)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.10 (0.02\u0026ndash;0.18)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"6\"\u003eAbbreviations: CI, confidence interval; PR AUC, area under the precision-recall curve; ROC AUC, area under the receiver operating characteristic curve.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eImportance and dependence of features in machine learning predictions\u003c/h2\u003e \u003cp\u003eTo assess the importance of each feature used in the prediction model, a SHAP analysis was performed using a random forest model. Among the features analyzed, the average SHAP values were highest for RFS, pathological stage I, pathological stage III, and vascular invasion, followed by PD-L1 expression (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA). The distribution of the SHAP scores was also analyzed for each feature (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eB). With respect to the PD-L1 expression, the red dots were predominantly located in the positive SHAP value region, while the blue dots were located in the negative SHAP value region. Thus, an increase in the expression of PD-L1 was associated with an increase in its contribution to model predictions. To further explore the impact of the PD-L1 expression on the model predictions, partial dependence plots were generated (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). The dependence on predictions greatly increased when PD-L1 was expressed in comparison to when it was not, and a subsequent linear increase in dependency was observed with increasing PD-L1 expression levels.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eMultivariate Cox proportional hazards analysis\u003c/h2\u003e \u003cp\u003eWe then examined the association between PD-L1 expression and postoperative recurrence using a statistical approach, specifically the multivariate Cox proportional hazards model. The results of the multivariate Cox proportional hazards analysis are presented in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e. The multicollinearity of each explanatory variable was assessed using the variance inflation factor (VIF). The VIF was less than 2 for all variables, indicating the absence of multicollinearity (Table \u003cspan refid=\"MOESM2\" class=\"InternalRef\"\u003eS2\u003c/span\u003e). A proportional hazards analysis was performed using a martingale residual plot. The generally horizontal trend of this curve confirmed the fulfilment of the proportional hazards assumption (Figure \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e). An increase in the expression of PD-L1 was associated with a significantly increased risk of postoperative recurrence (adjusted hazard ratio, 1.006; 95% CI 1.000\u0026ndash;1.011).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eResults of a multivariate Cox proportional hazard analysis of RFS according to the expression of PD-L1.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eUnadjusted HR (95% CI), \u003cem\u003eP\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003csup\u003e\u0026dagger;\u003c/sup\u003eAdjusted HR (95% CI), \u003cem\u003eP\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePD-L1 expression (TPS [%])\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.013 (1.008\u0026ndash;1.018), \u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.006 (1.000\u0026ndash;1.011), 0.04\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"3\"\u003e\u003csup\u003e\u0026dagger;\u003c/sup\u003eIn the multivariate analysis, the hazard ratio was adjusted for age, sex, histological type, pathological stage, vascular invasion, lymphovessel invasion, and adjuvant chemotherapy.\u003c/td\u003e\u003c/tr\u003e \u003ctr\u003e\u003ctd colspan=\"3\"\u003eAbbreviations: CI, confidence interval; HR, hazard ratio; PD-L1, programmed death ligand 1; RFS, recurrence-free survival.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eWe conducted a single-center retrospective observational study of 647 postoperative NSCLC patients to investigate the association between the expression of PD-L1 and postoperative recurrence. In our cohort, the recurrence rate was 23.3%, consistent with the previously reported rates of 20\u0026ndash;26% in larger cohorts studying lung cancer\u003csup\u003e23,24\u003c/sup\u003e. Our study showed that groups with higher PD-L1 expression levels had shorter RFS in the conventional classification based on PD-L1 expression levels (no expression [\u0026lt;\u0026thinsp;1%], low expression [1\u0026ndash;49%], and high expression [50\u0026ndash;100%]). In addition, using a machine learning model with a random forest algorithm and a multivariate Cox proportional hazards model with a statistical analysis, we investigated the impact of the PD-L1 expression on the postoperative recurrence of NSCLC. Our results showed a nonlinear increase in the risk of postoperative recurrence based on the PD-L1 expression level.\u003c/p\u003e \u003cp\u003eIncluding our previous study and several conventional studies, there have been reports linking the PD-L1 expression in NSCLC to an increased risk of postoperative recurrence\u003csup\u003e10,25\u0026ndash;28\u003c/sup\u003e. However, the results of our study may help to further develop this association. In previous studies, a statistical approach demonstrated an association between the expression of PD-L1 in NSCLC and postoperative recurrence. This association was similarly observed in the multivariate Cox proportional hazards analysis in our study, which is consistent with previous research. In contrast to previous studies, our study statistically demonstrated an increase in postoperative recurrence risk corresponding to each 1% increase in the PD-L1 expression level by conducting statistical analysis treating the PD-L1 expression as a continuous variable rather than as a categorical variable. Furthermore, we used a machine learning approach to evaluate the detailed effect of the PD-L1 expression on postoperative recurrence. Specifically, we revealed in detail the dynamic changes in postoperative recurrence depending on the PD-L1 expression level using a SHAP analysis and partial dependence plots. Our machine-learning approach has discovered a new finding that has not been reported in previous statistical analyses. This finding highlights the potential for dynamic variations in postoperative recurrence of NSCLC associated with PD-L1 expression levels. In conventional studies, the relationship between the expression of PD-L1 and the postoperative prognosis has been examined using categorical classifications such as no expression (\u0026lt;\u0026thinsp;1%), low expression (1\u0026ndash;49%), and high expression (\u0026ge;\u0026thinsp;50%). However, our study, leveraging machine learning techniques, revealed a more nuanced, continuous association between PD-L1 expression levels and recurrence risk that could not be captured by these traditional categories. The nonlinear increase in the risk of recurrence with even minimal PD-L1 levels (as low as 1%) and the linear escalation of risk when the expression increased beyond 1% suggest the importance of considering the PD-L1 expression as a continuous variable rather than discrete category. This finding underscores the potential for dynamic variations in the postoperative recurrence risk across the spectrum of PD-L1 expression, providing a more granular understanding of the relationship between the expression of PD-L1 and the recurrence of NSCLC. The results of this study may contribute to its clinical application. It has been reported that adjuvant chemotherapy with immune checkpoint inhibitors during the perioperative period of lung cancer improves a patient\u0026rsquo;s prognosis, even with PD-L1 expression levels as low as 1%, and this effect becomes more pronounced when the expression levels exceeds 50%\u003csup\u003e29\u003c/sup\u003e. The nonlinear and dynamic changes in the risk of recurrence based on PD-L1 expression levels may encourage the active introduction of perioperative immune checkpoint inhibitors for PD-L1-positive lung cancer patients as a clinical decision-making strategy.\u003c/p\u003e \u003cp\u003eThe results of this study can be explained from an immunological perspective. We showed that the contribution to recurrence increased non-linearly and sharply when PD-L1 was expressed, even at 1%, in comparison to PD-L1-negative cases. This trend suggests that, even the minimal expression of PD-L1 may have a significant impact on postoperative recurrence in NSCLC. Even when cancer cells express PD-L1 at 1%, subtle interactions between PD-L1-expressing cancer cells and the surrounding immune cells may lead to local immune suppression. This local suppression of immune cell activity may increase the immune resistance of cancer cells, activate local immune escape mechanisms and potentially increase the risk of postoperative recurrence. This possibility is supported by clinical studies showing the higher efficacy of immune checkpoint inhibitors in cases where PD-L1 is expressed, even at 1%, in comparison to cases with no expression\u003csup\u003e30,31\u003c/sup\u003e. In contrast, in cases with PD-L1 expression levels of \u0026ge;\u0026thinsp;1%, a linear increase in the contribution to postoperative recurrence was observed with increasing expression levels. This mechanism suggests that, as the number of cancer cells expressing PD-L1 increases, the suppression of immune cell activity increases, resulting in more immune-resistant cancer cells and further activation of immune escape mechanisms. This possibility is supported by clinical trials that demonstrate the higher efficacy of immune checkpoint inhibitors in patients with high PD-L1 expression levels (\u0026ge;\u0026thinsp;50%) in comparison to those with low expression levels (1\u0026ndash;49%) and PD-L1-negative cases (\u0026lt;\u0026thinsp;1%)\u003csup\u003e31\u0026ndash;33\u003c/sup\u003e. In addition to direct interactions with immune cells, the involvement of PD-L1 in the formation of the tumor microenvironment could be a potential mechanism by which different PD-L1 expression levels exert a nonlinear influence on the risk of recurrence. PD-L1 is capable of nuclear translocation and has been reported to directly bind to DNA and regulate the transcriptional induction of genes involved in the tumor microenvironment, such as immune responses and inflammation. In other words, when PD-L1 is expressed, the higher its expression level, the more likely it is that signaling pathways associated with tumor immune evasion are activated, potentially contributing to the establishment of an immunosuppressive tumor microenvironment. The formation of the tumor microenvironment may be closely related to the survival and proliferation of residual cancer cells, possibly increasing the risk of postoperative recurrence\u003csup\u003e34\u003c/sup\u003e. The dynamic changes in the increased risk of recurrence as a function of PD-L1 expression levels revealed in our study suggest significant variations in the risk of recurrence, even within the PD-L1 expression categories of low (1\u0026ndash;49%) and high (50\u0026ndash;100%) that are commonly used in clinical practice. Our study is the first to reveal the machine-learning-based dynamic variation of postoperative recurrence risk dependent on the continuous range of PD-L1 expression levels from low to high expression.\u003c/p\u003e \u003cp\u003eThis study was associated with several limitations. First, as this was a retrospective observational study conducted at a single institution, caution should be exercised in generalizing the results. Given its retrospective nature, it is a possible that clinical factors other than the expression of PD-L1 were not considered. Future validation through additional studies, such as prospective investigations or multicenter collaborations, is essential to elucidate the impact of these factors. Second, the performance of machine learning models is strictly limited by predictive accuracy and requires careful consideration in clinical applications. The results of this study, derived from a limited dataset, require further validation in clinical research using other cohorts or larger datasets to determine whether similar trends can be observed in recurrence prediction models. Overall, these considerations highlight the need for cautious interpretation and for future research to improve the robustness and applicability of our findings.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn conclusion, our study using machine learning and statistical analysis revealed a significant, nonlinear association between the expression of PD-L1 and the risk of postoperative recurrence in NSCLC. We demonstrated that even minimal PD-L1 expression levels (as low as 1%) were associated with an increased risk of recurrence, suggesting the potential impact of subtle immune interactions. Furthermore, a continuous increase in the expression of PD-L1 beyond 1% corresponded to a linear increase in the risk of recurrence. These novel findings suggest a dynamic relationship between the expression of PD-L1 and postoperative recurrence, and provide valuable insights into personalized therapeutic strategies for NSCLC. Our findings highlight the importance of considering the continuum of PD-L1 expression levels in real-world clinical practice, as opposed to the traditional classification of PD-L1 expression levels into low versus high. This study contributes to a better understanding of the immunological factors influencing the recurrence of NSCLC, paving the way for tailored treatment interventions based on the expression of PD-L1.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eK.K. conceived and designed the study. K.K. and H.S. collected the patient data. K.K. performed the data analysis. K.K., H.S., T.I., T.T., K.O. and H.Y. interpreted the analyzed data. K.K. wrote the manuscript. All authors reviewed and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe database used in this study is not available to the public. Participants in our study did not agree for their data to be publicly shared. The Python code used in this study is available on request from the corresponding author, [email protected], upon reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eKensuke Kojima received a specific grant from Grant-in-Aid for Clinical Research from the Osaka Foundation for the Prevention of Cancer and Cardiovascular Disease, a public funding agency.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConflict of interest\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAuthors have no conflict of interest to declare.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthical approval statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was approved by the Institutional Review Board (IRB) of the National Hospital Organization Kinki Chuo Chest Medical Center (KCMC) (approval number: 2023-31) 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.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePatient consent statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDue to the retrospective nature of this study, informed consent from the patients was not required.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eSung, H. \u003cem\u003eet al\u003c/em\u003e. Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. \u003cem\u003eCA Cancer J Clin\u003c/em\u003e. \u003cstrong\u003e71\u003c/strong\u003e, 209\u0026ndash;249 (2021).\u003c/li\u003e\n\u003cli\u003eSubotic, D., Van, Schil, P. \u0026amp; Grigoriu, B. Optimising treatment for post-operative lung cancer recurrence. \u003cem\u003eEur Respir J\u003c/em\u003e. \u003cstrong\u003e47\u003c/strong\u003e, 374\u0026ndash;378 (2016).\u003c/li\u003e\n\u003cli\u003eReck, M. \u003cem\u003eet al\u003c/em\u003e. Pembrolizumab versus chemotherapy for PD-L1-positive non-small-cell lung cancer. \u003cem\u003eN Engl J Med\u003c/em\u003e. \u003cstrong\u003e375\u003c/strong\u003e, 1823\u0026ndash;1833 (2016).\u003c/li\u003e\n\u003cli\u003eHan, Y., Liu, D. \u0026amp; Li, L. PD-1/PD-L1 pathway: current researches in cancer. \u003cem\u003eAm J Cancer Res\u003c/em\u003e. \u003cstrong\u003e10\u003c/strong\u003e, 727\u0026ndash;742 (2020).\u003c/li\u003e\n\u003cli\u003eDaud, AI. \u003cem\u003eet al\u003c/em\u003e. Programmed death-ligand 1 expression and response to the anti-programmed death 1 antibody pembrolizumab in melanoma. \u003cem\u003eJ Clin Oncol\u003c/em\u003e. \u003cstrong\u003e3\u003c/strong\u003e, 4102\u0026ndash;4109 (2016).\u003c/li\u003e\n\u003cli\u003eGaron, EB. \u003cem\u003eet al\u003c/em\u003e. Pembrolizumab for the treatment of non-small-cell lung cancer. \u003cem\u003eN Engl J Med\u003c/em\u003e. \u003cstrong\u003e372\u003c/strong\u003e, 2018\u0026ndash;2028 (2015).\u003c/li\u003e\n\u003cli\u003eRosenberg, JE. \u003cem\u003eet al\u003c/em\u003e. Atezolizumab in patients with locally advanced and metastatic urothelial carcinoma who progressed following treatment with platinum-based chemotherapy: a single arm, phase 2 trial. \u003cem\u003eLancet\u003c/em\u003e. \u003cstrong\u003e387\u003c/strong\u003e, 1909\u0026ndash;1920 (2016).\u003c/li\u003e\n\u003cli\u003eBrahmer, J. \u003cem\u003eet al\u003c/em\u003e. Nivolumab versus docetaxel in advanced squamous-cell non-small-cell lung cancer. \u003cem\u003eN Engl J Med\u003c/em\u003e. \u003cstrong\u003e373\u003c/strong\u003e, 123\u0026ndash;135 (2015).\u003c/li\u003e\n\u003cli\u003eSun, JM. \u003cem\u003eet al\u003c/em\u003e. Prognostic Significance of PD-L1 in Patients with Non-Small-Cell Lung Cancer: A Large Cohort Study of Surgically Resected Cases. \u003cem\u003eJ Thorac Oncol\u003c/em\u003e. \u003cstrong\u003e11\u003c/strong\u003e, 1003\u0026ndash;1011 (2016).\u003c/li\u003e\n\u003cli\u003eKojima, K., Sakamoto, T., Kasai, T., Kagawa, T., Yoon, H. \u0026amp; Atagi, S. PD-L1 expression as a predictor of postoperative recurrence and the association between the PD-L1 expression and EGFR mutations in NSCLC. \u003cem\u003eSci Rep\u003c/em\u003e. \u003cstrong\u003e11\u003c/strong\u003e, 17522 (2021).\u003c/li\u003e\n\u003cli\u003eCooper, WA. \u003cem\u003eet al\u003c/em\u003e. PD-L1 expression is a favorable prognostic factor in early stage non-small cell carcinoma. \u003cem\u003eLung Cancer\u003c/em\u003e. \u003cstrong\u003e89\u003c/strong\u003e, 181\u0026ndash;188 (2015).\u003c/li\u003e\n\u003cli\u003eMotono, N., Mizoguchi, T., Ishikawa, M., Iwai, S., Iijima, Y. \u0026amp; Uramoto, H. PD-L1 expression in not a predictive factor for recurrence in resected non-small cell lung cancer. \u003cem\u003eLung\u003c/em\u003e. \u003cstrong\u003e201\u003c/strong\u003e, 95\u0026ndash;101 (2023).\u003c/li\u003e\n\u003cli\u003eTravis, WD. \u003cem\u003eet al\u003c/em\u003e. The 2015 world health organization classification of lung tumors: Impact of genetic, clinical and radiologic advances since the 2004 classification. \u003cem\u003eJ Thorac Oncol\u003c/em\u003e. \u003cstrong\u003e10\u003c/strong\u003e, 1243\u0026ndash;1260 (2015).\u003c/li\u003e\n\u003cli\u003eRahman, S., Irfan, M., Raza, M., Ghori, K. M., Yaqoob, S. \u0026amp; Awais, M. Performance analysis of boosting classifiers in recognizing activities of daily living. \u003cem\u003eInt J Environ Res Public Health\u003c/em\u003e. \u003cstrong\u003e17\u003c/strong\u003e, 1082 (2020).\u003c/li\u003e\n\u003cli\u003ePatel, H., Vock, D. M., Marai, G. E., Fuller, C. D., Mohamed, A. S. R. \u0026amp; Canahuate, G. Oropharyngeal cancer patient stratification using random forest based-learning over high-dimensional radiomic features. \u003cem\u003eSci Rep\u003c/em\u003e. \u003cstrong\u003e11\u003c/strong\u003e, 14057 (2021).\u003c/li\u003e\n\u003cli\u003eVieira, T. \u003cem\u003eet al\u003c/em\u003e. Blood vessel invasion is a major feature and a factor of poor prognosis in sarcomatoid carcinoma of the lung. \u003cem\u003eLung Cancer\u003c/em\u003e. \u003cstrong\u003e85\u003c/strong\u003e, 276\u0026ndash;281 (2014).\u003c/li\u003e\n\u003cli\u003eAl-Alao, B. S., Gately, K., Nicholson, S., McGovern, E., Young, V. K. \u0026amp; O\u0026rsquo;Byrne, K. J. Prognostic impact of vascular and lymphovascular invasion in early lung cancer. \u003cem\u003eAsian Cardiovasc Thorac Ann\u003c/em\u003e. \u003cstrong\u003e22\u003c/strong\u003e, 55\u0026ndash;64 (2014).\u003c/li\u003e\n\u003cli\u003eMizuguchi, S., Izumi, N., Tsukioka, T., Komatsu, H. \u0026amp; Nishiyama, N. Neutrophil-lymphocyte ratio predicts recurrence in patients with resected stage 1 non-small cell lung cancer. \u003cem\u003eJ Cardiothorac Surg\u003c/em\u003e. \u003cstrong\u003e13\u003c/strong\u003e, 78 (2018).\u003c/li\u003e\n\u003cli\u003ePirker, R. Adjuvant chemotherapy in patients with completely resected non-small cell lung cancer. \u003cem\u003eTrans Lung Cancer Res\u003c/em\u003e. \u003cstrong\u003e3\u003c/strong\u003e, 305\u0026ndash;310 (2014).\u003c/li\u003e\n\u003cli\u003eLundberg, S. M. \u0026amp; Lee, S-I. A unified approach to interpreting model predictions. Proceedings of the 31st international conference on neural information processing systems; 2017.\u003c/li\u003e\n\u003cli\u003ePeduzzi, P., Concato, J., Feinstein, A. R. \u0026amp; Holford, T. R. Importance of events per independent variable in proportional hazards regression analysis. Ⅱ. Accuracy and precision of regression estimates. \u003cem\u003eJ Clin Epidemiol\u003c/em\u003e. \u003cstrong\u003e48\u003c/strong\u003e, 1503\u0026shy;\u0026ndash;1510 (1995).\u003c/li\u003e\n\u003cli\u003eKanda, Y. Investigation of the freely available easy-to-use software \u0026lsquo;EZR\u0026rsquo; for medical statistics. \u003cem\u003eBone Marrow Transplant\u003c/em\u003e. \u003cstrong\u003e48\u003c/strong\u003e, 452\u0026ndash;458 (2013).\u003c/li\u003e\n\u003cli\u003eDemicheli, R. \u003cem\u003eet al\u003c/em\u003e. Recurrence dynamics for non-small-cell lung cancer: effect of surgery on the development of metastases. \u003cem\u003eJ Thorac Oncol\u003c/em\u003e. \u003cstrong\u003e7\u003c/strong\u003e, 723\u0026ndash;730 (2012).\u003c/li\u003e\n\u003cli\u003eLou, F., Huang, J., Sima, C. S., Dycoco, J., Rusch, V. \u0026amp; Bach, P. B. Patterns of recurrence and second primary lung cancer in early-stage lung cancer survivors followed with routine computed tomography surveillance. \u003cem\u003eJ Thorac Cardiovasc Surg\u003c/em\u003e. \u003cstrong\u003e145\u003c/strong\u003e, 75\u0026ndash;81 (2013).\u003c/li\u003e\n\u003cli\u003eSamejima, H., Kojima, K., Fujiwara, A., Tokunaga, T., Okishio, K. \u0026amp; Yoon, H. The combination of PD-L1 expression and the neutrophil-to-lymphocyte ratio as a prognostic factor of postoperative recurrence in non-small-cell lung cancer: a retrospective cohort study. \u003cem\u003eBMC Cancer\u003c/em\u003e. \u003cstrong\u003e23\u003c/strong\u003e, 1107 (2023).\u003c/li\u003e\n\u003cli\u003eMiyazawa, T. \u003cem\u003eet al\u003c/em\u003e. PD-L1 expression in non-small-cell lung cancer including various adenocarcinoma subtypes. \u003cem\u003eAnn Thorac Cardiovasc Surg\u003c/em\u003e. \u003cstrong\u003e25\u003c/strong\u003e, 1\u0026ndash;9 (2019).\u003c/li\u003e\n\u003cli\u003eCha, Y. J., Kim, H. R., Lee, C. Y., Cho, B. C. \u0026amp; Shim, H. S. Clinicopathological and prognostic significance of programmed cell death ligand-1 expression in lung adenocarcinoma and its relationship with p53 status. \u003cem\u003eLung Cancer\u003c/em\u003e. \u003cstrong\u003e97\u003c/strong\u003e, 73\u0026ndash;80 (2016).\u003c/li\u003e\n\u003cli\u003eHanda, Y. \u003cem\u003eet al\u003c/em\u003e. Prognostic impact of programmed death-ligand 1 and surrounding immune status on stage I lung cancer. \u003cem\u003eClin Lung Cancer\u003c/em\u003e. \u003cstrong\u003e21\u003c/strong\u003e, e302\u0026ndash;e314 (2020).\u003c/li\u003e\n\u003cli\u003eForde, P. M. \u003cem\u003eet al\u003c/em\u003e. Neoadjuvant nivolumab plus chemotherapy in resectable lung cancer. \u003cem\u003eN Engl J Med\u003c/em\u003e. \u003cstrong\u003e386\u003c/strong\u003e, 1973\u0026ndash;1985 (2022).\u003c/li\u003e\n\u003cli\u003eHerbst, R. S. \u003cem\u003eet al\u003c/em\u003e. Pembrolizumab versus docetaxel for previously treated, PD-L1-positive, advanced non-small-cell lung cancer (KEYNOTE-010): a randomized controlled trial. \u003cem\u003eLancet\u003c/em\u003e. \u003cstrong\u003e387\u003c/strong\u003e, 1540\u0026ndash;1550 (2016).\u003c/li\u003e\n\u003cli\u003eMok, T. S. K. \u003cem\u003eet al\u003c/em\u003e. Pembrolizumab versus chemotherapy for previously untreated, PD-L1-expressing, locally advanced or metastatic non-small-cell lung cancer (KEYNOTE-042): a randomized, open-label, controlled, Phase 3 trial. \u003cem\u003eLancet\u003c/em\u003e. \u003cstrong\u003e393\u003c/strong\u003e, 1819\u0026ndash;1830 (2019).\u003c/li\u003e\n\u003cli\u003eJotte, R. \u003cem\u003eet al\u003c/em\u003e. Atezolizumab in combination with carboplatin and nab-paclitaxel in advanced squamous NSCLC (IMpower131): Results from a randomized phase III trial. \u003cem\u003eJ Thorac Oncol\u003c/em\u003e. \u003cstrong\u003e15\u003c/strong\u003e, 1351\u0026ndash;1360 (2020).\u003c/li\u003e\n\u003cli\u003eHerbst, R. S.\u003cem\u003e et al\u003c/em\u003e. Atezolizumab for first-line treatment of PD-L1-selected patients with NSCLC. \u003cem\u003eN Engl J Med\u003c/em\u003e. \u003cstrong\u003e383\u003c/strong\u003e, 1328\u0026ndash;1339 (2020).\u003c/li\u003e\n\u003cli\u003eGao, Y. \u003cem\u003eet al\u003c/em\u003e. Acetylation-dependent regulation of PD-L1 nuclear translocation dictates the efficacy of anti-PD-1 immunotherapy. \u003cem\u003eNat Cell Biol\u003c/em\u003e. \u003cstrong\u003e22\u003c/strong\u003e, 1064\u0026ndash;1075 (2020).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Machine learning, Random forest, Multivariate Cox proportional hazard model, Non-small cell lung cancer, Recurrence-free survival","lastPublishedDoi":"10.21203/rs.3.rs-4334704/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4334704/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe accurate prediction of postoperative recurrence is important for optimizing treatment strategies for non-small cell lung cancer (NSCLC). Previous studies have identified the PD-L1 expression in NSCLC as a risk factor for postoperative recurrence. This study aimed to examine the contribution of the PD-L1 expression in predicting postoperative recurrence using machine learning. The clinical data of 647 NSCLC patients who underwent surgical resection were collected and stratified into training (80%), validation (10%), and testing (10%) datasets. Machine learning models were trained on the training data using clinical parameters including the PD-L1 expression. The top-performing model was assessed on the test data using a SHAP analysis and partial dependence plots to quantify the contribution of the PD-L1 expression. A multivariate Cox proportional hazards model was used to validate the association between the PD-L1 expression and postoperative recurrence.\u003c/p\u003e \u003cp\u003eThe random forest model demonstrated the highest predictive performance with the SHAP analysis highlighting the PD-L1 expression as an important feature, and the multivariate Cox analysis indicating a significant increase in the risk of postoperative recurrence with each increment in the PD-L1 expression. These findings suggest that variations in the PD-L1 expression may provide valuable information for clinical decision-making in lung cancer treatment strategies.\u003c/p\u003e","manuscriptTitle":"Dynamic changes in postoperative risk of recurrence of non-small cell lung cancer according to variations in PD-L1 expression levels","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-05-08 20:30:18","doi":"10.21203/rs.3.rs-4334704/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2024-06-03T09:23:45+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-05-20T16:28:38+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"296783385496830619101559918175638621693","date":"2024-05-10T08:21:01+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"266664059895805906851600830517335077059","date":"2024-05-10T08:10:26+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"122057877291519471835565302468195190577","date":"2024-05-10T08:08:39+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2024-05-10T08:04:54+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2024-05-10T07:58:01+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2024-05-02T13:43:05+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2024-04-30T07:43:48+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2024-04-27T15:21:26+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"583bd22e-db3f-4282-b811-f649c50bcae4","owner":[],"postedDate":"May 8th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2024-07-01T14:52:31+00:00","versionOfRecord":[],"versionCreatedAt":"2024-05-08 20:30:18","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-4334704","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4334704","identity":"rs-4334704","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}