Machine learning-based reconstruction of prognostic staging for gastric cancer patients with different differentiation grades: A multicenter retrospective study

Abstract

BACKGROUND

The prognosis of gastric cancer (GC) patients is poor, and an accurate prognostic staging system would help assess patients' prognostic status before treatment and determine appropriate treatment strategies.

AIM

To develop positive lymph node ratio (LNR) and machine learning (ML)-based staging systems for GC patients with varying differentiation.

METHODS

This multicenter retrospective cohort study included 11772 GC patients, with 5612 in the training set (Harbin Medical University Cancer Hospital) and 6160 in the validation set (Surveillance, Epidemiology, and End Results Program database). X-tile software identified optimal cutoff values for the positive LNR, and five ML models were developed using pT and LNR staging. Risk scores were divided into seven stages, constructing new staging systems tailored to different tumor differentiation levels.

RESULTS

In both the training and validation sets, regardless of the tumor differentiation level, LNR staging demonstrated superior prognostic stratification compared to pN. Extreme Gradient Boosting exhibited better predictive performance than the other four models. Compared to tumor node metastasis staging, the new staging systems, developed for patients with different degrees of differentiation, showed significantly better predictive performance.

CONCLUSION

The new positive lymph nodes ratio staging and integrated staging systems constructed for GC patients with different differentiation grades exhibited better prognostic stratification capabilities.

Key Words: Gastric cancer; Machine learning; Positive lymph nodes ratio; Prognostic staging system; Tumor differentiation

Core Tip: This study introduces new machine learning-based gastric cancer (GC) staging systems, which incorporate the positive lymph node ratio and pT stages, tailored for well/moderately differentiated GC and poorly differentiated GC patients, respectively. These novel systems demonstrate superior prognostic accuracy compared to the traditional American Joint Committee on Cancer tumor node metastasis staging system, providing a more precise tool for predicting overall survival in resectable GC patients and guiding personalized treatment strategies.

INTRODUCTION

Gastric cancer (GC) is one of the most prevalent and lethal types of cancer worldwide, with over 1 million new cases and approximately 769000 deaths attributed to it in 2020[1]. In recent years, advanced treatments, such as minimally invasive surgery, targeted therapy, and immunotherapy, have been widely adopted in clinical practice, extending the survival time of GC patients to some extent and improving their quality of life[2-4]. Concurrently, accurate staging plays a crucial role in devising individualized treatment plans, optimizing treatment outcomes, and assessing patient prognosis[5].

The eighth edition of the tumor node metastasis (TNM) staging system developed by the American Joint Committee on Cancer (AJCC) provides a standardized staging framework for GC. This system comprehensively evaluates the depth of tumor invasion, the number of regional lymph node metastases (LNM), and distant metastases, serving as the primary reference for prognostic evaluation and treatment guidance in patients with GC[6]. However, several studies have highlighted limitations in this staging system. First, tumor differentiation, which significantly influences the prognosis of GC patients, is inadequately addressed in the current TNM staging system[7,8]. Thus, neglecting tumor differentiation may lead to incomplete and inaccurate prognosis assessment. Second, the impact of the number of lymph nodes examined on N staging is a critical issue. Variations in patient conditions, extent of lymph node dissection, and surgical technique may affect the number of nodes examined, potentially leading to stage migration, also known as the Will Rogers phenomenon[9,10]. Therefore, studies have explored the value of the positive lymph node ratio (LNR), which analyzes the ratio of metastatic to total examined lymph nodes, in the prognostic assessment of patients with GC. The results indicate that the LNR serves as an independent prognostic factor, improving the stratification of patient prognosis compared with the traditional N staging system[11-13]. Although previous studies have replaced the N staging system with the LNR staging system and developed new staging systems based on the eighth edition of the AJCC TNM staging system[14], the application of machine learning (ML) algorithms to further refine GC staging systems remains limited.

This study aimed to identify the optimal cutoff points for the LNR in patients with GC at different differentiation levels. These cutoff points were used to categorize the LNR into five stages. ML prediction models for overall survival (OS) were subsequently constructed for GC patients. Furthermore, on the risk scores generated by the best models, we developed new GC staging systems for well/moderately differentiated GC (WDGC) and poorly differentiated GC (PDGC) patients and compared their predictive performance with that of the eighth edition of the AJCC TNM staging system.

MATERIALS AND METHODS

Patients

This multicenter retrospective study included a total of 11772 patients diagnosed with GC confirmed by gastroscopy and pathological examination. Among them, 5612 GC patients who underwent radical gastrectomy with standard D2/D2+ lymph node dissection at Harbin Medical University Cancer Hospital and diagnosed between January 2011 and December 2018 were assigned to the training set. A further 6160 patients from the Surveillance, Epidemiology, and End Results Program database (http://seer.cancer.gov/seerstat/) and diagnosed with GC between January 2008 and December 2015 were allocated to the validation set.

The inclusion criteria were as follows: (1) Diagnosed with GC by pathology; (2) Owing to ethical considerations and the rarity of GC in the pediatric population, aged 18 years and older; (3) Complete survival information was available; and (4) Underwent radical GC surgery. The exclusion criteria were as follows: (1) Distant metastasis of GC; (2) Incomplete clinical or pathological data; (3) Patients who underwent surgery with fewer than 16 Lymph nodes removed; and (4) A survival time of less than 1 month, these patients were excluded because their short survival times may not accurately reflect long-term treatment effects. Furthermore, this exclusion criterion helps reduce data anomalies and outliers, improving the stability and representativeness of the research results. The patient enrollment flowchart is shown in Figure 1. The study design and implementation strictly adhered to the Declaration of Helsinki and relevant ethical guidelines.

Figure 1  Flow chart of the study.

Data extraction

Clinical and pathological characteristics, including sex, age, LNM status, LNR, pT stage, pN stage, TNM stage, tumor location, and tumor differentiation, were collected from medical records. All patients were staged according to the AJCC TNM staging system (8^th edition).

The LNR stage and TNM staging system

Initially, according to the AJCC pN staging system (8^th edition), an LNR equal to 0 was defined as LNR0[11]. Using data from the training set, we subsequently calculated two optimal cutoff values for WDGC and PDGC patients using X-tile software, dividing patients into LNR1, LNR2, and LNR3. Finally, we calculated one optimal cutoff value for LNR3 GC patients with different differentiation levels using X-tile software and further divided the patients into LNR3a and LNR3b groups.

Construction of ML models and new TNM staging system

In this study, we developed eXtreme gradient boosting (XGBoost), logistic regression (LR), support vector machine (SVM), random forest (RF), decision tree (DT) models based on the pT staging system and the new LNR staging system to predict the OS of WDGC and PDGC patients. After model establishment, we identified the best model by comprehensively evaluating different predictive models. Using the quantile method, we subsequently categorized the risk scores generated by the best model into 7 stages corresponding to stages I A to IIIC of the eighth edition AJCC TNM staging system. To gain a deeper understanding of the relationship between different risk stages and clinical variables, we conducted statistical analysis using contingency tables to examine the pT and LNR stages associated with each risk stage, with the goal of constructing new GC staging systems. We compared their predictive performances with that of the traditional TNM staging system.

Study outcomes and follow-up

The primary endpoint of this study was the OS of patients. OS was defined as the time from diagnosis to death from any cause. We set the follow-up endpoint at 5 years and recorded the occurrence of outcome events if patients died from any cause within 5 years. For the training set data, patients were followed up every 3-6 months via telephone. The last follow-up date was January 1, 2024.

Statistical analysis

Categorical variables are described as frequencies and percentages, whereas continuous variables are described as medians and quartiles[15]. The receiver operating characteristic (ROC) curve and its area under the curve (AUC) were used to compare the predictive performance of different features or predictive models. In addition, to more accurately evaluate the predictive ability of the models, we also use sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV) to conduct an in-depth assessment of the models' performance, thereby selecting the best model. Univariate and multivariate Cox proportional hazards regression analyses were used to identify features associated with OS in GC patients and compare the hazard ratios (HR) of different features. Additionally, the likelihood ratio χ² test in the Cox proportional hazards regression model was used to compare the homogeneity of the TNM staging system and new TNM staging systems on the basis of the predictive models; the Akaike information criterion (AIC) value, concordance index (C-index) associated with the Cox regression model, and decision curve analysis (DCA) were used to compare and evaluate the predictive performance of the TNM system and new GC staging systems. A lower AIC value, higher C-index, and larger area under the decision curve (AUDC) indicated the models had more accurate predictions[16]. Python version 3.9 (https://docs.python.org/3.9/) was used for the web-based calculator, and all other statistical analyses were conducted using R software version 4.2.3 (R Core Team (2023)). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria (https://www.R-project.org/). In this study, P < 0.05 was considered statistically significant.

RESULTS

Patient characteristics

A total of 5612 GC patients were included in the training set, with a five-year survival rate of 65.9%. In the validation set, 6160 GC patients were included, with a five-year survival rate of 49.7%. Additionally, in the training set, the median age of the patients was 59 years, with 2829 patients diagnosed with PDGC (50.4%). Compared with WDGC patients, those with PDGC were significantly more likely to be female, have a higher pT stage, have a higher pN stage, and have greater LNM (P < 0.05, Supplementary Table 1). In the validation set, the mean age of the patients was 67 years, with 4150 patients diagnosed with PDGC (67.3%). Similarly, patients with PDGC were more likely to have a higher pT stage, higher pN stage, and more LNM than patients with WDGC (Supplementary Table 2).

Comparison between the 8^th edition of the AJCC pN staging system and the LNR staging system

According to the analysis results from X-tile software, for WDGC patients, the three optimal cutoff points for LNR1–LNR3b were 0.108, 0.350, and 0.800 (Supplementary Figure 1). For PDGC patients, the three optimal cutoff points for LNR1-LNR3b were 0.174, 0.550, and 0.830 (Supplementary Figure 2). Based on these optimal cutoff points, we categorized patients into LNR0-LNR3b in the training and validation sets and compared their prognostic ability with that of the 8^th AJCC pN stage system via ROC analysis. For WDGC patients, in the training set, the AUC for the LNR stage and pN stage were 0.733 (95%CI: 0.712-0.753) and 0.710 (95%CI: 0.689-0.730), respectively; the accuracies were 0.750 and 0.714, respectively (Supplementary Figure 3A). In the validation set, the AUC for the LNR stage and pN stage were 0.669 (95%CI: 0.647-0.691) and 0.659 (95%CI: 0.637-0.681), respectively; the accuracies were 0.646 and 0.646, respectively (Supplementary Figure 3B). Similar findings were observed in PDGC patients, where in the training set, the AUC for the LNR stage and pN stage were 0.738 (95%CI: 0.719-0.757) and 0.719 (95%CI: 0.699-0.738), respectively; the accuracies were 0.727 and 0.676 (Supplementary Figure 3C), respectively. In the validation set, the AUC for the LNR stage and pN stage were 0.718 (95%CI: 0.704-0.733) and 0.705 (95%CI: 0.690-0.720), respectively; the accuracies were 0.660 and 0.649, respectively (Supplementary Figure 3D).

To further compare the stratification ability of the LNR and pN staging systems for the prognosis of GC patients, we performed univariate Cox proportional hazards regression analysis for OS based on data from the training set, separately for WDGC and PDGC patients. The results indicated that sex, age, tumor location, pT stage, pN stage, and LNR stage were significantly associated with OS in both groups (Supplementary Table 3). Considering the potential collinearity between the LNR stage and pN stage, we used these two staging systems separately to adjust for the other four features and performed multivariate Cox regression analysis. This process allowed the independent prognostic value of the LNR and pN stage to be evaluated while controlling for other variables. Compared with the LNR0 stage or pN0 stage, the LNR1 stage had a smaller HR than the pN1 stage in both the WDGC and PDGC patients in the training set. However, the HR for LNR2 stage to LNR3b stage disease were consistently greater than those for pN2 stage to pN3b stage disease (Figure 2). Similar observations were made in the validation set (Figure 3).

Figure 2 Multivariable cox proportional hazards regression analysis of lymph node ratio staging system and pN staging system for overall survival in gastric cancer patients with different degrees of differentiation in the training set. A: Lymph node ratio staging system; B: pN staging system. WD: Well-differentiated; MD: Moderately differentiated; PD: Poorly differentiated; GC: Gastric cancer.

Figure 3 Multivariable cox proportional hazards regression analysis of lymph node ratio staging system and pN staging system for overall survival in gastric cancer patients with different degrees of differentiation in the validation set. A: Lymph node ratio staging system; B: pN staging system. WD: Well-differentiated; MD: Moderately differentiated; PD: Poorly differentiated; GC: Gastric cancer.

Prediction performance of the ML models

Using data from the training set, we constructed five ML predictive models based on the pT stage and LNR stage for WDGC and PDGC patients and validated their predictive ability in the validation set. We performed ROC analysis on the risk scores generated by the five models for each patient. The results revealed that for WDGC patients, in the training set, the AUC for XGBoost, SVM, RF, LR, and DT were 0.773 (95%CI: 0.754-0.793), 0.703 (95%CI: 0.679-0.726), 0.725 (95%CI: 0.705-0.744), 0.758 (95%CI: 0.737-0.778), and 0.717 (95%CI: 0.698-0.736), respectively (Supplementary Figure 4A). In the validation set, the AUC for XGBoost, SVM, RF, LR, and DT were 0.709 (95%CI: 0.687-0.732), 0.623 (95%CI: 0.598-0.648), 0.646 (95%CI: 0.626-0.665), 0.705 (95%CI: 0.682-0.728), and 0.643 (95%CI: 0.624-0.662), respectively (Supplementary Figure 4B). We then compared the sensitivity, specificity, accuracy, PPV, and NPV of the five models. In the training set, XGBoost had the highest specificity, accuracy, and PPV, with values of 0.677, 0.765, and 0.743, respectively. In the validation set, XGBoost had the highest sensitivity, accuracy, and NPV, with values of 0.611, 0.670, and 0.674, respectively (Supplementary Table 4). On the basis of the overall performance of each model in both the training and validation sets, XGBoost performed the best and was selected as the optimal model. For PDGC patients, in the training set, the AUC for XGBoost was 0.769 (95%CI: 0.751-0.787), which was greater than those of the other models (Supplementary Figure 4C). In the validation set, the AUC for XGBoost was 0.738 (95%CI: 0.723-0.753), which was also greater than those of the other models (Supplementary Figure 4D). In addition, we found that in the training set, XGBoost had the highest accuracy, and XGBoost, SVM, and RF had the highest sensitivity, specificity, and PPV, with values of 0.577, 0.836, and 0.684, respectively. In the validation set, XGBoost had the highest specificity and accuracy, with values of 0.845 and 0.660, respectively. LR had the highest sensitivity and NPV, with values of 0.576 and 0.606, respectively (Supplementary Table 5). After a comprehensive evaluation of all the models' predictive performance, XGBoost was chosen as the best model for predicting OS in PDGC patients. Finally, we compared the AUC differences between XGBoost and the other four models using DeLong's test. The results revealed that, for both WDGC and PDGC patients, the AUC of XGBoost and LR were similar in the validation set. In all other cases, the AUC of XGBoost was significantly greater than those of the other models (Supplementary Tables 6 and 7).

Comparison between the 8^th edition of the AJCC TNM staging system and the new staging systems

First, for WDGC patients, we calculated percentiles from the risk scores derived from XGBoost in the training set, dividing the data into 0%, 14.3%, 28.6%, 42.9%, 57.1%, 71.4%, 85.7%, and 100% percentiles. We subsequently identified the pT stage and LNR stage corresponding to each risk stage and developed the xTRM-W staging system exclusive to WDGC patients, where x represents XGBoost, T and M have the same meanings as in the 8^th edition of the AJCC TNM staging system, R represents the LNR stage, and W represents well/moderately differentiated stages. Using the same method, we also developed a new staging system for PDGC patients, named the xTRM-P staging system, where P represents poor differentiation. Additionally, these two new staging systems were applied to their respective validation sets to confirm their generalizability and practicality. Figure 4 shows the detailed distributions of the 8^th edition AJCC TNM staging system, the xTRM-W staging system, and the xTRM-P staging system. Two Sankey diagrams depict the migration patterns for the two new staging systems compared with those of the 8^th edition AJCC TNM staging system (Figure 5).

Figure 4 Specific staging content of the 8^th edition American Joint Committee on Cancer tumor node metastasis staging system, xTRM-W staging system, and xTRM-P staging system. A: American Joint Committee on Cancer tumor node metastasis staging system; B: xTRM-W staging system; C: xTRM-P staging system. LNR: Lymph node ratio.

Figure 5 The distribution differences between the xTRM-W staging system, xTRM-P staging system, and the 8^th edition American Joint Committee on Cancer tumor node metastasis staging system. A: The distribution differences between the xTRM-W staging system and the 8^th edition American Joint Committee on Cancer tumor node metastasis (AJCC TNM) staging system; B: The distribution differences between the xTRM-P staging system and the 8^th edition AJCC TNM staging system. AJCC: American Joint Committee on Cancer.

We subsequently used multiple evaluation metrics to compare the predictive performance of the 8^th edition AJCC TNM staging system and the xTRM-W staging system in WDGC patients in the training and validation sets, including time-dependent ROC, DCA, likelihood ratio χ², AIC, and the C-index. The results indicated that in both the training and validation sets, the xTRM-W staging system had a greater AUC for predicting 1-, 3-, and 5-year mortality rates than the AJCC TNM staging system (Figure 6A-D); the xTRM-W staging system showed greater net benefit across most threshold ranges than the pTNM staging system (Figure 6E and F). Additionally, in both the training and validation sets, the xTRM-W staging system exhibited better homogeneity (higher likelihood ratio χ²), a larger C-index, and a smaller AIC value than the pTNM staging system (Table 1). Similar trends were observed in PDGC patients, where in both the training and validation sets, the xTRM-P staging system generally had a larger AUC and AUDC than the pTNM staging system (Figure 7); the xTRM-P staging system had a greater likelihood ratio χ² and C-index and a smaller AIC value (Table 2). These results indicate that for GC patients with different differentiation levels, the xTRM-W and xTRM-P staging systems have better prognostic ability than the traditional pTNM staging system.

Figure 6 Time-dependent receiver operating characteristic curve analysis and decision curve analysis of the xTRM-W staging system and the 8^th edition American Joint Committee on Cancer tumor node metastasis staging system. A: Time-dependent receiver operating characteristic curve (ROC) analysis of the xTRM-W staging system in the training set; B: Time-dependent ROC analysis of the 8^th edition American Joint Committee on Cancer tumor node metastasis (AJCC TNM) staging system in the training set; C: Time-dependent ROC analysis of the xTRM-W staging system in the validation set; D: Time-dependent ROC analysis of the 8^th edition AJCC TNM staging system in the validation set; E: Decision curve analysis (DCA) of the xTRM-W staging system and the 8^th edition AJCC TNM staging system in the training set; F: DCA of the xTRM-W staging system and the 8^th edition AJCC TNM staging system in the validation set. ROC: Receiver operating characteristic curve; DCA: Decision curve analysis.

Figure 7 Time-dependent receiver operating characteristic curve analysis and decision curve analysis of the xTRM-P staging system and the 8^th edition American Joint Committee on Cancer tumor node metastasis staging system. A: Time-dependent receiver operating characteristic curve (ROC) analysis of the xTRM-P staging system in the training set; B: Time-dependent ROC analysis of the 8^th edition American Joint Committee on Cancer tumor node metastasis (AJCC TNM) staging system in the training set; C: Time-dependent ROC analysis of the xTRM-P staging system in the validation set; D: Time-dependent ROC analysis of the 8^th edition AJCC TNM staging system in the validation set; E: Decision curve analysis (DCA) of the xTRM-P staging system and the 8^th edition AJCC TNM staging system in the training set; F: DCA of the xTRM-P staging system and the 8^th edition AJCC TNM staging system in the validation set. ROC: Receiver operating characteristic curve; DCA: Decision curve analysis.

Table 1 Comparison of the predictive performance of the xTRM-W staging system and the 8^th edition tumor node metastasis staging system.

Group	Classification	Likelihood ratio χ2	C-index	AIC value
Training set	xTRM-W	668.580	0.742	12323.976
	pTNM	479.560	0.714	12512.994
Validation set	xTRM-W	324.290	0.666	12489.603
	pTNM	275.700	0.655	12538.197

TNM: Tumor node metastasis.

Table 2 Comparison of the predictive performance of the xTRM-P staging system and the 8^th edition tumor node metastasis staging system.

Group	Classification	Likelihood ratio χ2	C-index	AIC value
Training set	xTRM-P	732.740	0.723	15938.971
	pTNM	559.260	0.704	16112.451
Validation set	xTRM-P	989.680	0.680	34678.546
	pTNM	893.260	0.680	34774.967

TNM: Tumor node metastasis.

Web-based calculator

In addition to developing the xTRM-W staging system and xTRM-P staging system for GC patients at different differentiation levels, we also developed web-based calculators on the XGBoost algorithm. The web-based calculations for WDGC and PDGC patients are available at https://xgbhigh.streamlit.app/ and https://xgblow.streamlit.app/, respectively. These web-based calculators include the pT stage (assigned values of 1-6 corresponding to T1a-T4b) and the LNR stage (assigned values of 0-4 corresponding to LNR0-LNR3b).

DISCUSSION

This study classified all patients into WDGC and PDGC groups based on their differentiation status. We subsequently introduced LNR staging systems into each group to replace the traditional pN staging system and compared their predictive performance and HRs. Compared with the AJCC pN staging system, the LNR staging system demonstrated better stratification ability for patient prognosis. Next, we developed three ML predictive models based on pT staging and LNR staging for GC patients with different differentiation statuses. Through comparison, XGBoost was selected as the optimal model. Finally, on XGBoost-derived risk scores, we developed new xTRM-W and xTRM-P staging systems for WDGC and PDGC patients, respectively. A comprehensive evaluation revealed that both new staging systems outperformed the AJCC TNM staging system in predicting patient outcomes in both the training and validation sets.

The 8^th edition of the AJCC TNM staging system, introduced in 2017, refines pN3 into pN3a and pN3b, and studies have shown that it is superior to the 7^th edition in predicting the prognosis of GC patients[17]. However, with the increasing use of postoperative adjuvant chemotherapy, especially in stage III GC patients, survival times have increased, narrowing the prognostic differences between stages and limiting the predictive power of the TNM system[18,19]. Additionally, the 8^th edition does not consider tumor differentiation, although it is an established independent prognostic factor for GC[20-22]. Compared with WDGC patients, PDGC patients are at a greater risk of recurrence and metastasis compared to WDGC patients, suggesting the need for adjuvant radiotherapy[23]. This study also revealed that PDGC patients had more lymph node metastasis than WDGC patients did. Therefore, the GC staging system should be further refined by including tumor differentiation to better assess patient prognosis and support personalized treatment.

The LNR was first proposed in 2002 and has since been used to evaluate the prognosis of GC patients[24]. Subsequent research has consistently shown the LNR to be an independent prognostic factor for GC patients, with LNR staging, stratified by LNM and the number of lymph nodes examined, better predicting prognosis than traditional pN staging[25,26], which is consistent with the results of this study. The LNR consists of LNM and the number of lymph nodes examined. As the sole component of the pN staging system developed by the AJCC, LNM has always been one of the most important factors affecting prognosis and is a crucial reference for selecting treatment options for GC patients[27,28]. Additionally, the number of lymph nodes examined is significantly correlated with survival in GC patients. Studies have indicated that increasing the number of lymph nodes examined significantly prolongs survival in GC patients; the optimal survival rate is observed when the number of negative lymph nodes exceeds 20[29]. Moreover, GC patients who undergo examination of more than 40 Lymph nodes have a significantly greater 5-year survival rate than other patients[30]. Therefore, these reports suggest obtaining more lymph nodes for examination during surgery. These results further support the LNR as a more precise tool for assessing prognosis in GC patients because it reflects both the extent of LNM and the comprehensiveness of lymph node examination.

Some studies have compared the LNR with pN staging in GC patients[12]. The optimal cutoff value for the LNR remains controversial. In addition to differences in study cohorts, these studies used different staging methods. For example, some studies have divided GC patients with LNR greater than 0.37 into high-risk groups on the basis of the median LNR[31]. Other researchers used X-tile software to divide the LNR into three stages, finding optimal cutoff values of 0.05 and 0.43[32]. Compared with previous studies, this study provided a more detailed division of the LNR. First, we identified the optimal cutoff values for the LNR separately in GC patients with different degrees of differentiation. Second, using data from the training set and referring to the 8^th edition of the AJCC N staging system, patients with an LNR of 0 were categorized into the LNR0 stage. Subsequently, lymph node-positive GC patients were further classified into LNR1-LNR3 stages using X-tile software, with patients in the LNR3 stage further divided into LNR3a and LNR3b stages. These calculated cutoff values were then applied to the validation set data. The results showed that LNR staging had better predictive performance than pN staging in both the training and validation sets.

This study is the first to use pT staging and the newly introduced LNR staging system to construct ML models for optimizing the TNM staging system for GC. ML is an artificial intelligence algorithm that automatically constructs predictions on the basis of given data without the need for explicit programming, yielding precise results under uncertain conditions[33]. In recent years, this method has been widely used to predict the occurrence and progression of various diseases, achieving outstanding results[34,35]. Among these methods, XGBoost is a new type of ML algorithm with an optimized implementation of gradient boosting. Numerous studies have shown that the predictive performance of XGBoost is superior to those of other predictive models[36,37]. The results of this study are consistent with these findings, as XGBoost was selected as the optimal model, regardless of patient differentiation status, on the basis of comprehensive evaluation.

In recent years, improvements in living standards and the diversification of treatment modalities have significantly improved the survival rates of GC patients[38,39]. However, GC remains the fifth leading cause of cancer-related deaths worldwide[1]. To address this problem, an increasing number of studies are attempting to select the most relevant features related to the prognosis of GC patients using various data reduction methods and construct predictive models to help physicians and patients better understand the severity of their disease and the risk of death before treatment, thereby selecting the most appropriate treatment regimen for patients[40-42]. In contrast to previous studies, this study skipped the feature selection process. We chose to replace the pN staging system with the new LNR staging system based on the 8^th edition of the AJCC TNM staging system. Additionally, we did not consolidate the T1a and T1b stages into a single T1 stage, as in the 8^th edition pTNM staging system, but retained the T1a and T1b stages as two independent stages to provide a more accurate staging system. This approach was beneficial because GC patients with stage T1b disease were more likely to have shorter survival times than those with stage T1a disease were in our study (Supplementary Figure 5). Using complete pT staging and the new LNR staging, we subsequently constructed ML models for OS in GC patients with different differentiation statuses and selected XGBoost as the optimal model. On the basis of XGBoost-derived patient risk scores, we developed new xTRM-W and xTRM-P staging systems for WDGC and PDGC patients, respectively. These two new GC staging systems are similar in form to the traditional TNM staging system; on the basis of postoperative pathology information for resectable GC patients, physicians can easily determine the specific stage of patients. Nevertheless, both new staging systems demonstrated significantly better predictive performance than did the traditional TNM staging system. Additionally, we developed web-based calculators based on the XGBoost algorithm for GC patients with different differentiation statuses, which can assess mortality risk on the basis of patient staging. Therefore, the two new staging systems based on the XGBoost algorithm are convenient and effective tools that have the potential to help physicians assess the prognosis of GC patients more accurately.

This study also has limitations. First, this study was retrospective and analyzed patient data from a single center in the training set, which inevitably increases the risk of selection bias, necessitating further expansion of data sources and multicenter prospective studies to validate the reliability of our results. Second, the results of this study included only OS for GC patients. In future studies, we hope to construct new separate GC staging systems for progression-free survival and cancer-specific survival to assess the prognosis of patients with GC more effectively. Additionally, the new xTRM-W staging system and xTRM-P staging system developed in this study are applicable only to resectable GC patients and cannot be used to assess the prognosis of stage IV GC patients effectively. Therefore, in future research, we will focus on identifying prognostic markers applicable to stage IV GC patients and constructing new staging systems to assess the risk of death and disease progression in GC patients with distant metastases.

CONCLUSION

In both the training and validation sets, the prognostic stratification of the LNR staging system was superior to that of the pN staging system, regardless of the degree of tumor differentiation. Compared with the AJCC TNM staging system, the new xTRM-W and xTRM-P staging systems developed in this study demonstrated better predictive performance. They have the potential to help physicians identify GC patients at increased risk of death and guide the selection of treatment strategies.

ACKNOWLEDGEMENTS

We thank Professor Ying-Wei Xue and Professor Hai-Bin Song for their support and supervision of this study.