Development and validation of a machine learning model for predicting 3-year overall survival in metastatic nasopharyngeal carcinoma: a SEER database and web visualization study

Introduction

Nasopharyngeal carcinoma (NPC) is a malignant tumor that originates from the epithelium of the nasopharynx. It is particularly prevalent in East Asia and Southeast Asia. Due to its hidden primary site and vague clinical manifestations, approximately 5–10% of NPC patients are diagnosed with distant metastasis at their initial presentation (1-3). Although the long-term survival rate for patients with stage metastatic NPC (M1-NPC) is extremely low, with an overall survival (OS) of only 20–30% at 3 years, there still exists significant heterogeneity in survival outcomes among these patients (4). Previous studies have demonstrated that some M1-NPC patients can achieve long-term disease-free survival following aggressive multimodal therapy (5-8).

The 8^th edition of the American Joint Committee on Cancer (AJCC) tumor-node-metastasis (TNM) staging system is currently the most comprehensive anatomical staging system for cancer. In this system, “M1” refers to patients with primary metastatic NPC who already have distant metastases at the time of diagnosis. However, the M1 category is typically classified using a “catch-all” approach and is generally grouped together in the same category. Unfortunately, TNM staging based solely on the “M1” classification cannot accurately predict the survival outcome of patients. There have been numerous studies in the past aimed at further optimizing the staging of M1-NPC (6,9-11). One study employed recursive partitioning analysis (RPA) to divide patients with metachronous and synchronous liver and bone metastases into two subgroups, M1a and M1b, and found a statistically significant difference in median OS (mOS) between the two subgroups (9). Another study focused on the number and organ of metastases, ultimately subdividing M1 into M1a (oligometastasis without liver metastases) and M1b (liver metastases or multiple metastases) based on survival outcomes (12). What’s more, Zeng et al. developed a prognostic nomogram model for OS (1- and 3-year OS) in M1-NPC, with multivariate analysis revealing that M-stage subclassification (M1a defined as ≤5 metastatic lesions, M1b as >5 lesions without liver metastasis, and M1c as >5 lesions with liver metastasis), primary tumor radiotherapy, and immunotherapy were significantly correlated with OS. The area under the receiver operating characteristic (ROC) curves (AUCs) for 1- and 3-year OS curves of the nomogram model were 0.799 [95% confidence interval (CI): 0.717–0.882] and 0.766 (95% CI: 0.667–0.866), respectively, while the concordance index (C-index) reached 0.737 (95% CI: 0.692–0.782) (13). Wen et al. established a RPA-based stratification model incorporating metastatic lesion number, liver metastasis status, and post-treatment Epstein-Barr virus (EBV)-DNA levels to predict progression-free survival (PFS) in de novo M1-NPC patients receiving chemoimmunotherapy and local regional radiotherapy, demonstrating AUCs of 0.77, 0.722, and 0.829 for 1-, 2-, and 3-year PFS in the validation cohort, respectively (14). However, these studies present certain limitations, including the complexity of prediction systems that hinder clinical implementation and incomplete incorporation of clinicopathological factors, which may restrict the widespread adoption and application of these refined staging approaches. Consequently, there remains an unmet need for a more comprehensive, clinically practical, and visually intuitive stratification methodology to optimize patient management.

In recent years, machine learning (ML) algorithms have been widely used in various tumor prediction models, leading to significant improvements in the efficiency of medical practice (15). Support vector machine (SVM), random forest (RF), extreme gradient boosting (XGBoost), gradient boosting decision tree (GBDT), and k-nearest neighbor (KNN) are five commonly employed ML algorithms. RF enhances model generalization through ensemble learning of multiple decision trees, demonstrating notable advantages in handling high-dimensional data, resisting noise interference, and providing built-in feature importance evaluation. SVM excels in high-dimensional feature space but suffers from high computational complexity and poor interpretability. XGBoost and GBDT achieve exceptional prediction accuracy through gradient boosting mechanisms yet require meticulous parameter tuning to prevent overfitting. KNN offers intuitive simplicity but experiences significant computational efficiency degradation with increasing data volume. Each algorithm possesses its own strengths and limitations, and the selection and fine-tuning of these algorithms should be tailored to different tumor types and available data in specific applications. The optimal choice of model depends on the specific research objectives and available resources. It is crucial to continue conducting research and validation to ensure the accuracy and reliability of the model.

The Surveillance, Epidemiology, and End Results (SEER) database in the United States is a large registry of cancer patients, containing a large number of samples and complete follow-up information. The database provides rich data on diagnosis, treatment, and survival follow-up of cancer cases. The aim of this study is to collect clinicopathological data and follow-up information of M1-NPC patients from the SEER database. Based on a comparison of different ML methods, we aim to establish a prognostic prediction model for M1-NPC patients. Additionally, we intend to develop a user-friendly online calculator that visually presents individualized survival predictions, enabling clinicians to efficiently assess patient prognosis and guide personalized treatment strategies.

By integrating advanced ML techniques with practical clinical tools, this study seeks to bridge the gap between prognostic research and real-world practice, ultimately improving therapeutic decision-making for M1-NPC patients. We present this article in accordance with the TRIPOD reporting checklist (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-977/rc).

Methods

Data sources and research subjects

A retrospective analysis was conducted using the SEER database (SEER*Stat 8.4.3; https://seer.cancer.gov) for this study. The study included data from the SEER Research Data, 17 Registries, Nov 2023 Sub [2000–2021]. The SEER program routinely collects follow-up information through active linkage to vital status records and other data sources. For each patient in the SEER database, the follow-up time is calculated from the date of initial diagnosis to the date of the last known contact (for OS) or the date of death. The survival status (alive or deceased) and the cause of death (cancer-specific or other causes) are also provided. The population of interest for this study was patients diagnosed with M1-NPC, and their clinical pathology and follow-up data were collected and analyzed. Since the SEER database is a public database and does not contain identifiable patient information, ethical approval from an ethics committee was not required.

Inclusion criteria:

• The primary tumor was located in the nasopharynx [according to International Classification of Diseases for Oncology, 3^rd edition (ICD-O-3) International Classification of Tumors C11/World Health Organization (WHO) 2008];
• TNM stage was M1;
• The diagnosis was made between 2010 and 2021;
• Complete follow-up information.

Exclusion criteria:

• Lack of information about whether the primary tumor was surgically resected and whether regional lymph nodes were resected;
• No information available regarding the presence of metastases in the brain, liver, bone, and lungs.

The inclusion and exclusion criteria and flow chart were shown in Figure 1 for details. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Figure 1 Inclusion and exclusion criteria and flow chart. AUC, area under the ROC curve; GBDT, gradient boosting decision tree; ICD-O-3, International Classification of Diseases for Oncology; KNN, k-nearest neighbor; NPV, negative predictive value; PPV, positive predictive value; RF, random forest; ROC, receiver operating characteristic; SEER, Surveillance, Epidemiology, and End Results; SHAP, SHapley Additive exPlanations; SVM, support vector machine; TNM, tumor-node-metastasis; WHO, World Health Organization; XGBoost, extreme gradient boosting.

Data collection and preprocessing

In this study, 19 variables were selected as input variables, while the presence or absence of death and survival time were used as output variables. The input variables included: age, sex, race, marital status, income, time from diagnosis to treatment, surgery (primary site), surgery (regional lymph nodes), radiotherapy, chemotherapy, lymph node size, tumor size, bone metastasis, brain metastasis, liver metastasis, lung metastasis, number of metastatic organs (including bone, brain, liver, and lung), tumor stage (T stage), and node stage (N stage).

The statistical analysis in this study was performed using Python version 3.12 (https://www.python.org/downloads/). Due to the presence of missing variables in the SEER database, including race (0.7%), marital status (5.3%), interval from diagnosis to treatment (16.1%), lymph node size (35.4%), and tumor size (44.4%), linear regression was applied to impute the missing values in the dataset. Linear regression imputation involves constructing a predictive model using a complete-case dataset or ML-based regression algorithms, whereby available observed attributes from subjects with missing values are substituted into the model to estimate and impute absent values. This approach effectively handles datasets with a substantial proportion of missing data while preserving key statistical properties—such as means, variances, and correlation structures—more accurately than simple imputation methods like mean or median substitution. By maintaining the covariance structure between variables, it reduces bias in parameter estimates and enhances the robustness and power of downstream analyses, including the training and performance of ML models. This method was chosen to ensure the robustness of the study (16,17).

To streamline the data as much as possible, we have referred to previous studies and transformed the continuous variables (age, income, interval from diagnosis to treatment, lymph node size, tumor size) into categorical variables (18). The specific variable assignment is detailed in Table S1.

Statistical analysis

The specific flow chart of this study is shown in Figure 1. First, univariate Cox regression analysis was performed on the included population to screen out the independent risk factors related to the 3-year OS of M1-NPC patients, and these factors were included in the follow-up study. The significance level was set at P<0.05. To ensure the robustness and generalizability of our model, we randomly partitioned the entire dataset into a training set and an internal testing set using a simple random sampling method with a fixed ratio (typically 70% for training and 30% for testing). A fixed random seed was set at 42 prior to splitting to guarantee the reproducibility of our results.

After verifying that there were no significant differences in baseline data, five ML models were used to establish and evaluate the training set. The parameters of each model were tuned by grid search and five-fold cross-validation to select the model with the best performance (19,20).

To establish clinical relevance, a prognostic model must demonstrate robust performance across key dimensions, including discrimination, calibration, clinical utility, and interpretability, transcending mere statistical significance. Firstly, the ROC curve is an intuitive way to assess sensitivity and specificity (21). The performance of the test depends on the AUC value, where a higher AUC value indicates better performance of the ML model. AUC values of <0.6, 0.6 to <0.7, 0.7 to <0.8, 0.8 to <0.9, and ≥0.9 were defined as failed, poor, fair, good, and excellent for discrimination, respectively (22-24). Secondly, calibration was examined visually and statistically using calibration curves, while clinical utility was quantified via decision curve analysis (DCA) to estimate net benefit across clinically relevant risk thresholds. Furthermore, model interpretability was enhanced through SHapley Additive exPlanations (SHAP) value analysis, enabling transparent visualization of feature contributions and facilitating clinical adoption (25).

Next, the two best-performing base models were combined into one Ensemble Model using stacking. Finally, the performance of each model on the test set was evaluated, including Accuracy, Sensitivity, Specificity, positive predictive value (PPV), negative predictive value (NPV), and AUC. Additionally, a ROC curve was plotted.

Furthermore, we constructed a prognostic model for M1-NPC patients based on the best model. This model divided patients into high-risk and low-risk death groups, with a 3-year time OS frame. Survival curves were generated for these two groups.

For the convenience of clinicians, we have developed an online calculator based on the prediction model. This calculator provides quick prediction results based on inputted patient characteristics.

Results

Demographic characteristics and parameter screening

According to the inclusion criteria, a total of 689 M1-NPC patients were included in this study. Among all patients, the proportions of demographic characteristics were as follows: age: ≥60 (45.72%) or <60 years old (54.28%); sex: female (22.79%) or male (77.21%); race: Asian or Pacific Islander (41.96%), White (42.84%), American Indian/Alaska Native (2.49%) or Black (12.72%); marital status: single/unmarried (28.68%), married (55.06%), divorced/separated (9.82%), or widowed (6.44%); income: <40,000 (1.02%), 40,000–79,999 (47.90%), 80,000–119,999 (42.53%) or ≥120,000 US dollar (USD) (8.56%); time from diagnosis to treatment: <30 (56.06%), 30–59 (28.89%) or ≥60 days (15.05%); surgery of primary site (4.50%); surgery of regional lymph node (4.93%); lymph node size: 0 (18.88%), <1 (75.96%), or ≥1 cm (5.17%); tumor size: <30 (22.45%), 30–59 (57.18%), or ≥60 mm (20.37%); radiotherapy (56.46%); chemotherapy (75.04%); bone metastasis (51.81%); brain metastasis (7.84%); liver metastasis (28.74%); lung metastasis (33.38%); number of metastases (bone, lung, brain, and liver): 0 (15.09%), 1 (55.44%), 2 (22.93%), 3 (5.66%), and 4 (0.87%); T0 (1.31%), T1 (22.50%), T2 (12.19%), T3 (21.48%), T4 (28.88%), or Tx (13.64%); N0 (14.95%), N1 (28.59%), N2 (31.79%), N3 (20.61%), or Nx (4.06%).

First, a univariate Cox regression analysis was performed for all enrolled populations (Table 1). Using 3-year OS as the outcome indicator, the results showed that age, race, time from diagnosis to treatment, radiotherapy, chemotherapy, lymph node size, liver metastasis, lung metastasis, number of metastatic organs (bone, liver, brain, lung), T stage, and N stage demonstrated significant statistical differences in univariate survival analysis (P<0.05). Therefore, we considered these factors as independent risk factors for 3-year OS in M1-NPC patients.

In the subsequent process of model construction, we included the aforementioned factors for consideration. The 689 patients were randomly divided into a training set and a testing set in a ratio of 7:3. The training set consisted of 482 cases, while the testing set had 207 cases. The median survival time for both groups was 17 months. By comparing the P values of age, race, time from diagnosis to treatment, radiotherapy, chemotherapy, lymph node size, liver metastasis, lung metastasis, number of metastases (bone, liver, brain, and lung), T stage, and N stage between the training set and testing set, the P values were found to be 0.219, 0.306, 0.906, 0.517, 0.480, 0.978, 0.785, 0.368, 0.413, 0.25, and 0.613, respectively. All P values were greater than 0.05, indicating that there were no baseline differences between the training and testing set (Table S2).

Training model prediction

In the training set, five ML algorithms (SVM, RF, XGBoost, GBDT, KNN) were applied in this study to analyze the data, and the best predictive model was selected. Grid search and 5-fold cross-validation were used to optimize the parameters of each model, and the best model was chosen based on the AUC score (Table 2). By comparing the AUC values, all the predictive models showed good performance (AUC >0.9). In particular, RF and KNN achieved an AUC of 0.9998, which was the best among all the ML models.

Subsequently, we created an ensemble model called soft voting using RF and KNN models and added it to the model dictionary. The results showed that in the training set, the soft voting ensemble model exhibited comparable performance to RF and KNN (AUC =0.998).

Testing model prediction

In the testing set, we evaluated the six ML models above, and they are ranked based on their AUC values as follows: RF (AUC =0.72), GBDT & KNN (AUC =0.67), ensemble (AUC =0.64), XGboost (AUC =0.63), and SVM (AUC =0.56) (Table 2). According to the AUC ranking, RF demonstrated the best predictive ability. In this study, the RF model achieved an AUC value of 0.72, which is generally classified as “fair” in performance. With an AUC value greater than 0.7, RF is generally considered to have strong enough discrimination, indicating its high accuracy in classifying different categories (26). Additionally, we plotted the ROC curve to assess the performance of the ML models (Figure 2).

Figure 2 ROC comparison of SVM, RF, XGBoost, GBDT, KNN, and ensemble model in testing test. AUC, area under the ROC curve; GBDT, gradient boosting decision tree; KNN, k-nearest neighbor; RF, random forest; ROC, receiver operating characteristic; SVM, support vector machine; XGBoost, extreme gradient boosting.

Calibration and clinical use

When comparing calibration curves across multiple ML models, it is generally necessary to evaluate which curve aligns most closely with the ideal calibration line. A higher degree of alignment indicates better accuracy in the model’s predictive performance. Additionally, it is essential to integrate these findings with a comparative analysis of ROC curves across the models in order to comprehensively identify the optimal model (Figure S1).

Figure S2 presents the DCA curves for both the training and testing datasets. When comparing DCA curves across models, a curve positioned higher within the same threshold range indicates greater net benefit. While DCA primarily evaluates the clinical utility of multiple models by quantifying their net benefit in support of decision-making, it remains essential to also consider the model’s discriminative ability and accuracy in the context of clinical application.

Model interpretation and feature importance

To enhance our understanding of the contribution of features in the RF prediction model to the prediction results, we utilized SHAP to illustrate the ranking of feature importance. In Figure 3A, the risk factors were evaluated based on the average absolute SHAP value. Figure 3B presents all the features included in our model. The feature ranking is indicated on the Y-axis, showcasing the significance of the predictive model. On the X-axis, the SHAP value serves as a standardized index reflecting the influence of a specific feature within the model. In each row of the feature importance, the attributions of all patients to the outcome were represented by dots of various colors. The red dots indicate a high-risk value, while the blue dots represent a low-risk value.

Figure 3 The model’s interpretation. (A) The importance ranking of the 11 variables according to the mean (|SHAP value|); (B) the importance ranking of the 11 risk factors with stability and interpretation using the optimal model. The higher SHAP value of a feature is given, the higher risk of death the patient would have. The red part in feature value represents higher value. N, node; SHAP, SHapley Additive exPlanations; T, tumor.

The features are ranked in descending order of importance as follows: time from diagnosis to treatment, T stage, N stage, race, number of metastases, lung metastasis, chemotherapy, age, lymph node size, radiotherapy, and liver metastasis. This ranking indicates that in this prediction model, time from diagnosis to treatment, T stage, and N stage are considered the most important features, contributing the most to the prediction results. Conversely, features such as lymph node size, radiotherapy, and liver metastasis have relatively lower importance in the prediction. In terms of treatment, chemotherapy is more important than radiotherapy for patients with M1-NPC.

Survival curves plotting

In this study, we constructed a 3-year OS prediction model for M1-NPC patients based on RF. According to the 3-year OS, patients were classified into high- or low-risk groups based on a threshold of 50% mortality probability. The Kaplan-Meier method was used to plot the survival curves to visualize the survival probabilities between different risk groups (Figure 4). The results showed that RF could effectively differentiate between high- and low-risk groups: compared to the high-risk group, the low-risk group had significantly longer survival time with a significant statistical difference (P=3.003e−10).

Figure 4 Survival curves for high- and low-risk groups. Group 0: high-risk group; group 1: low-risk group.

Network calculator building

In this study, an online calculator was developed based on the results of the RF model. The interface is shown in Figure 5. As shown in the screenshot, the 3-year OS probability for this sample was 10.00%. Clinicians can input the 11 predictive factor indices of M1-NPC patients and generate the 3-year survival probability with just one click.

Figure 5 Screenshot of the online prediction calculator. OS, overall survival.

Discussion

In this study, our objective was to develop an ML model based on RF and SHAP to stratify the 3-year OS risk of M1-NPC patients with catch-all classification. Subsequently, an online calculator was constructed for convenient visualization purposes.

Distant metastasis is the most important cause of death in NPC patients, and patients diagnosed with M1-NPC have a poorer prognosis compared to those who develop distant metastasis after comprehensive treatment. In the population with initial distant metastasis, the median survival time is approximately 13–34 months (27,28). However, there is significant heterogeneity within this population, and the current AJCC TNM 8^th edition categorizes all M1-NPC patients into a single group, lacking the ability to accurately differentiate the prognoses of M1-NPC patients. Therefore, analysis of M1-NPC based on prognostic factors is of great significance, aiming to predict the survival outcomes of M1-NPC patients early and then formulate individualized plans for modifying treatment methods and adjusting therapy intensity according to patient needs.

In our study, we employed multiple ML algorithms to construct models, and the results showed that RF performed the best in terms of predictive performance. RF is a powerful ML algorithm that has demonstrated excellent performance in many tumor prediction models (29-32). It was originally proposed by Leo Breiman and Adele Cutler in 2001. Cutler concluded that RF has several advantages compared to other ML algorithms: (I) high classification accuracy; (II) assessment of feature importance; (III) capability to analyze various types of data; and (IV) modeling complex interactions among explanatory variables (32).

Although the RF model achieved an AUC of 0.72—indicating only fair discriminatory accuracy—it still demonstrates significant clinical utility in this study. Specifically, the model effectively stratifies patients into distinct risk groups with statistically significant differences (P<0.001), thereby offering valuable support for clinical decision-making. Based on further analysis, we suggest that the relatively small sample size in the testing set may be the main factor limiting the model’s stability, as performance in such contexts often improves with larger testing cohorts. In addition, we attempted to construct an ensemble model to improve the prediction performance. Unfortunately, the ensemble model did not improve the predictive performance of RF and KNN. Possible reasons for this are: (I) the performance improvement of an ensemble model often relies on the diversity of the base models. Different types of models have complementary strengths in capturing different patterns and noise in the data. However, in the current implementation of the ensemble model, both RF and KNN models share similar characteristics, resulting in insufficient complementarity between the models. (II) Ensemble models may be more sensitive to noise in the data, especially when there is no significant difference in performance among the base models. The ensemble process attempts to integrate the results of all base models, but if the prediction errors of the base models are similar, the ensemble process may exacerbate the impact of these errors, leading to a decrease in overall performance (33).

Based on the SEER database, we conducted univariate Cox regression analysis and identified 11 variables for inclusion in our model, including T stage, N stage, race, lymph node size, number of metastases, time from diagnosis to treatment, age, radiotherapy, liver metastasis, lung metastasis, and chemotherapy. Some of these results have been reported in previous studies (27,34-36).

When all variables are ranked according to their relative importance, the top 4 variables are time from diagnosis to treatment, T stage, N stage, and race. The present study appears to have yielded a puzzling result regarding the effect of time from diagnosis to treatment initiation on survival outcome. The findings showed that delayed interval time improved survival rather than worse survival, which is contradictory to our consensus. Ramos et al. focused on the association of delay with survival in two systematic reviews published, and in their analysis, more studies reported improved survival with longer delay intervals (four out of 26 studies) rather than worse survival (two out of 26 studies) (37,38). A study published by Castelo et al. showed that younger patients with shorter total intervals (<108 days) had worse OS and cancer-specific survival (CSS) (39). Tørring et al. proposed the concept of a U-shaped association in which mortality was highest in patients with the shortest time from symptom onset to diagnosis and in those with an interval greater than the 70^th percentile (40). These results provide support for the conclusions of this study. Reasons for this observation may be attributed to underlying disease biology. Specifically, patients with longer treatment intervals may have more indolent disease progression, whereas those with aggressive or highly symptomatic disease are likely to seek medical attention earlier due to distressing symptoms, thus resulting in shorter delay intervals. This pattern raises the possibility that the observed intervals may be influenced by differences in clinical urgency and disease aggressiveness (41-44). Some previous studies have reached a consistent conclusion that T stage or N stage may be an independent prognostic factor related to M1-NPC patients’ survival (10,45-47). A study has shown that the incidence of metastatic lymph nodes in the distance of the M1-NPC is about 24% (47). The impact of lymph node metastasis on survival prognosis in NPC may be attributed to factors such as increased lymph node stiffness, modulation of the tumor immune microenvironment, and the potential synergy with radiotherapy and immunotherapy (48,49). Further research is needed to fully understand the mechanisms involved. Due to the regional nature of NPC, few previous studies have focused on exploring the impact of ethnicity on prognosis. In our study, among the 689 included patients, 41.96% were Asian or Pacific Islander, 2.49% were American Indian/Alaska Native, 12.72% were Black, and 42.84% were White. Compared to Asian or Pacific Islander, the hazard ratio for 3-year OS was 1.43 (95% CI: 1.17–1.75) for Whites, indicating that being White is a risk factor for 3-year OS compared to Asians or Asian or Pacific Islander.

Among them, the number of organs involved and the site of metastases have been hot topics in prognostic research for M1-NPC (12,27,35). The results of our study indicate that liver and lung metastasis, as well as the number of metastatic organs, are significant risk factors for the survival prognosis of M1-NPC patients. According to the importance evaluation, the relative rankings of these three factors are number of metastatic organs, presence of lung metastasis, and presence of liver metastasis. In our study, we analyzed the number of metastases (bone, liver, brain, lung) and the site of metastases separately. The proportions of patients with bone, liver, brain, and lung metastasis were 51.81%, 28.74%, 7.84%, and 33.38% respectively, and the proportions of patients with 0–4 metastatic lesions in these four organs were 15.09%, 55.44%, 22.93%, and 5.66%, and 0.87% respectively. This study provides evidence that the risk of mortality rises with an increasing number of metastatic organs. In accordance with certain previous studies, the presence or absence of liver metastasis has been considered as one of the criteria for further subdivision of M1-stage NPC into M1a and M1b sub-stages (12,50). In relation to lung metastasis, Pan et al. identified that lung metastasis was an independent prognostic factor for M1-NPC (27). The results of these similar previous studies provide further evidence for the present study.

The median survival time for all patients in the study was 17 months, which demonstrated that even in patients with distant metastasis at initial diagnosis, timely systemic treatment offers a chance of survival beyond 1 year. Regarding treatment modalities, this study identified both radiotherapy and chemotherapy as favorable prognostic factors. These findings are consistent with and extend previous research in this field (51). However, in this study, the role of radiotherapy in improving long-term survival rates appeared to be relatively limited compared to the systemic effects of chemotherapy. We speculate that this may be due to the fact that local radiotherapy does not target all metastatic lesions. Currently, two ongoing phase III trials (NCT05128201 and NCT04421469) are investigating the consolidative role of radiotherapy targeting all metastatic lesions in M1-NPC, and we eagerly await the results of these trials.

The 9^th edition of the AJCC TNM staging for NPC in 2025 classifies M1 stage patients into two subcategories: M1a (≤3 metastatic lesions) and M1b (>3 metastatic lesions), based on 5-year OS (52). Our study differs in several aspects: (I) different research perspectives: Our study focuses on the number of metastatic organs in the bone, liver, brain, and lung, while the AJCC 9^th edition focuses on the number of metastatic lesions. (II) Consideration of liver involvement: similar to the results of previous studies, liver involvement was an independent survival factor for patients with M1-NPC. However, due to the low proportion of liver-involved cases (16%), the AJCC 9^th edition did not incorporate liver involvement into the staging changes, which is a limitation. (III) Statistical methods: Our study employs the RF algorithm instead of traditional Cox regression analysis to better explain the importance of each characteristic in predicting the outcome. While the 9^th edition provides important and normative guidance for reclassifying the M1 stage, it is undeniable that our study offers a different and visual perspective on the subdivision of M1-NPC.

Above all, in this study, we established an online calculator to predict the 3-year probability of survival based on 11 variables: T stage, N stage, race, lymph node size, number of metastases, time from diagnosis to treatment, age, radiotherapy, liver metastasis, lung metastasis, and chemotherapy. This calculator can potentially aid in clinical decision making by facilitating timely adjustment of treatment plans and therapeutic intensity, thereby assisting in guiding patient management.

There are limitations to our study. First, the SEER database is limited to specific regions within the United States. As a result, it may not fully represent the global or non-United States cancer patient population, introducing potential geographical bias into our findings. Therefore, the results should be interpreted with caution when generalizing to populations outside the United States. Second, while our model benefits from the comprehensive clinical and pathological variables available in the SEER database, we acknowledge the absence of several established prognostic factors that could further refine predictive accuracy. These include promising molecular biomarkers such as microRNA panels (53), well-documented EBV-DNA load parameters, detailed chemotherapy metrics (including cycle number and regimen intensity), comprehensive metastatic lesion characterization, and advanced quantitative imaging features [computed tomography (CT)/magnetic resonance imaging (MRI) radiomics] (54). Finally, it should be noted that the learning ability of the model is limited due to the restricted amount of data and the absence of external validation. Future studies should incorporate multi-center data, include previously missing variables, and perform external validation to enhance the model’s reliability and generalizability.

Conclusions

In summary, we utilized RF to construct a predictive model for the 3-year OS outcome of M1-NPC patients and developed an accompanying online calculator to assist clinicians in assessing patient’s survival risk. By evaluating individual risk, clinicians can take appropriate intervention measures in advance to prolong patient survival.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-977/rc

Peer Review File: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-977/prf

Funding: This work was supported by the Jiangsu Province Entrepreneurship and Innovation Doctoral Talent Program (No. 2019303073386ER19 to S.Z.) and the Jiangsu Province People’s Hospital Clinical Capability Enhancement Project (No. JSPH-MC-2021-17 to S.Z.).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-977/coif). S.Z. reports that this work was supported by the Jiangsu Province Entrepreneurship and Innovation Doctoral Talent Program (No. 2019303073386ER19) and the Jiangsu Province People’s Hospital Clinical Capability Enhancement Project (No. JSPH-MC-2021-17). The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Lin M, Zhang XL, You R, et al. Evolutionary route of nasopharyngeal carcinoma metastasis and its clinical significance. Nat Commun 2023;14:610. [Crossref] [PubMed]
2. You R, Liu YP, Huang PY, et al. Efficacy and Safety of Locoregional Radiotherapy With Chemotherapy vs Chemotherapy Alone in De Novo Metastatic Nasopharyngeal Carcinoma: A Multicenter Phase 3 Randomized Clinical Trial. JAMA Oncol 2020;6:1345-52. [Crossref] [PubMed]
3. Chen YP, Chan ATC, Le QT, et al. Nasopharyngeal carcinoma. Lancet 2019;394:64-80. [Crossref] [PubMed]
4. Hong S, Zhang Y, Yu G, et al. Gemcitabine Plus Cisplatin Versus Fluorouracil Plus Cisplatin as First-Line Therapy for Recurrent or Metastatic Nasopharyngeal Carcinoma: Final Overall Survival Analysis of GEM20110714 Phase III Study. J Clin Oncol 2021;39:3273-82. [Crossref] [PubMed]
5. Hui EP, Leung SF, Au JS, et al. Lung metastasis alone in nasopharyngeal carcinoma: a relatively favorable prognostic group. A study by the Hong Kong Nasopharyngeal Carcinoma Study Group. Cancer 2004;101:300-6. [Crossref] [PubMed]
6. Teo PM, Kwan WH, Lee WY, et al. Prognosticators determining survival subsequent to distant metastasis from nasopharyngeal carcinoma. Cancer 1996;77:2423-31. [Crossref] [PubMed]
7. Fandi A, Bachouchi M, Azli N, et al. Long-term disease-free survivors in metastatic undifferentiated carcinoma of nasopharyngeal type. J Clin Oncol 2000;18:1324-30. [Crossref] [PubMed]
8. Choo R, Tannock I. Chemotherapy for recurrent or metastatic carcinoma of the nasopharynx. A review of the Princess Margaret Hospital experience. Cancer 1991;68:2120-4. [Crossref] [PubMed]
9. Chan SK, O’Sullivan B, Huang SH, et al. An Exploratory Study of Refining TNM-8 M1 Categories and Prognostic Subgroups Using Plasma EBV DNA for Previously Untreated De Novo Metastatic Nasopharyngeal Carcinoma. Cancers (Basel) 2022;14:1923. [Crossref] [PubMed]
10. Shen L, Dong J, Li S, et al. M1 stage subdivision and treatment outcome of patients with bone-only metastasis of nasopharyngeal carcinoma. Oncologist 2015;20:291-8. [Crossref] [PubMed]
11. Zeng L, Tian YM, Huang Y, et al. Retrospective analysis of 234 nasopharyngeal carcinoma patients with distant metastasis at initial diagnosis: therapeutic approaches and prognostic factors. PLoS One 2014;9:e108070. [Crossref] [PubMed]
12. Lu TZ, Zeng FJ, Hu YJ, et al. Anatomic prognostic factors and their potential roles in refining M1 classification for de novo metastatic nasopharyngeal carcinoma. Cancer Med 2023;12:22091-102. [Crossref] [PubMed]
13. Zeng XY, Li Y, Ma J, et al. Prognostic nomogram for synchronous metastatic nasopharyngeal carcinoma: a retrospective multicentre study. Radiat Oncol 2025;20:42. [Crossref] [PubMed]
14. Wen D, Gu L, Long H, et al. Recursive partitioning analysis model for de novo metastatic nasopharyngeal carcinoma treated with locoregional radiotherapy following chemoimmunotherapy. ESMO Open 2024;9:103960. [Crossref] [PubMed]
15. Swanson K, Wu E, Zhang A, et al. From patterns to patients: Advances in clinical machine learning for cancer diagnosis, prognosis, and treatment. Cell 2023;186:1772-91. [Crossref] [PubMed]
16. Liu L, Zhang R, Shi Y, et al. Automated machine learning for predicting liver metastasis in patients with gastrointestinal stromal tumor: a SEER-based analysis. Sci Rep 2024;14:12415. [Crossref] [PubMed]
17. Beyad Y, Maeder M. Multivariate linear regression with missing values. Anal Chim Acta 2013;796:38-41. [Crossref] [PubMed]
18. Yu C, Zhang Y. Establishment of prognostic nomogram for elderly colorectal cancer patients: a SEER database analysis. BMC Gastroenterol 2020;20:347. [Crossref] [PubMed]
19. Wu P, Zhang C, Tang X, et al. Pan-cancer characterization of cell-free immune-related miRNA identified as a robust biomarker for cancer diagnosis. Mol Cancer 2024;23:31. [Crossref] [PubMed]
20. Adnan M, Alarood AAS, Uddin MI, et al. Utilizing grid search cross-validation with adaptive boosting for augmenting performance of machine learning models. PeerJ Comput Sci 2022;8:e803. [Crossref] [PubMed]
21. Obuchowski NA, Bullen JA. Receiver operating characteristic (ROC) curves: review of methods with applications in diagnostic medicine. Phys Med Biol 2018;63:07TR01. [Crossref] [PubMed]
22. El Khouli RH, Macura KJ, Barker PB, et al. Relationship of temporal resolution to diagnostic performance for dynamic contrast enhanced MRI of the breast. J Magn Reson Imaging 2009;30:999-1004. [Crossref] [PubMed]
23. Pitamberwale A, Mahmood T, Ansari AK, et al. Biochemical Parameters as Prognostic Markers in Severely Ill COVID-19 Patients. Cureus 2022;14:e28594. [Crossref] [PubMed]
24. Mandrekar JN. Receiver operating characteristic curve in diagnostic test assessment. J Thorac Oncol 2010;5:1315-6. [Crossref] [PubMed]
25. Ponce-Bobadilla AV, Schmitt V, Maier CS, et al. Practical guide to SHAP analysis: Explaining supervised machine learning model predictions in drug development. Clin Transl Sci 2024;17:e70056. [Crossref] [PubMed]
26. White N, Parsons R, Collins G, et al. Evidence of questionable research practices in clinical prediction models. BMC Med 2023;21:339. [Crossref] [PubMed]
27. Pan CC, Lu J, Yu JR, et al. Challenges in the modification of the M1 stage of the TNM staging system for nasopharyngeal carcinoma: A study of 1027 cases and review of the literature. Exp Ther Med 2012;4:334-8. [Crossref] [PubMed]
28. Liu Y, Zuo ZC, Zeng XY, et al. Establishing subdivisions of M1 stage nasopharyngeal carcinoma based on decision tree classification: A multicenter retrospective study. Oral Oncol 2024;153:106834. [Crossref] [PubMed]
29. Jin Y, Lan A, Dai Y, et al. Development and testing of a random forest-based machine learning model for predicting events among breast cancer patients with a poor response to neoadjuvant chemotherapy. Eur J Med Res 2023;28:394. [Crossref] [PubMed]
30. Chong HH, Yang L, Sheng RF, et al. Multi-scale and multi-parametric radiomics of gadoxetate disodium-enhanced MRI predicts microvascular invasion and outcome in patients with solitary hepatocellular carcinoma ≤ 5 cm. Eur Radiol 2021;31:4824-38. [Crossref] [PubMed]
31. Zhang Y, Zhang Z, Wei L, et al. Construction and validation of nomograms combined with novel machine learning algorithms to predict early death of patients with metastatic colorectal cancer. Front Public Health 2022;10:1008137. [Crossref] [PubMed]
32. Cutler DR, Edwards TC Jr, Beard KH, et al. Random forests for classification in ecology. Ecology 2007;88:2783-92. [Crossref] [PubMed]
33. Brownlee J. Ensemble learning algorithms with Python: Make better predictions with bagging, boosting, and stacking. Machine Learning Mastery; 2021.
34. Shen LJ, Wang SY, Xie GF, et al. Subdivision of M category for nasopharyngeal carcinoma with synchronous metastasis: time to expand the M categorization system. Chin J Cancer 2015;34:450-8. [Crossref] [PubMed]
35. Shen L, Li W, Wang S, et al. Image-based Multilevel Subdivision of M1 Category in TNM Staging System for Metastatic Nasopharyngeal Carcinoma. Radiology 2016;280:805-14. [Crossref] [PubMed]
36. Raslan S, Rodriguez E, Zheng C, et al. Association Between Mortality Due to Nasopharyngeal Carcinoma and Race in the United States From 2007 to 2016. Asian Pac J Cancer Prev 2023;24:915-21. [Crossref] [PubMed]
37. Ramos M, Esteva M, Cabeza E, et al. Lack of association between diagnostic and therapeutic delay and stage of colorectal cancer. Eur J Cancer 2008;44:510-21. [Crossref] [PubMed]
38. Ramos M, Esteva M, Cabeza E, et al. Relationship of diagnostic and therapeutic delay with survival in colorectal cancer: a review. Eur J Cancer 2007;43:2467-78. [Crossref] [PubMed]
39. Castelo M, Paszat L, Hansen BE, et al. Analysis of Time to Treatment and Survival Among Adults Younger Than 50 Years of Age With Colorectal Cancer in Canada. JAMA Netw Open 2023;6:e2327109. [Crossref] [PubMed]
40. Tørring ML, Frydenberg M, Hamilton W, et al. Diagnostic interval and mortality in colorectal cancer: U-shaped association demonstrated for three different datasets. J Clin Epidemiol 2012;65:669-78. [Crossref] [PubMed]
41. Neal RD. Do diagnostic delays in cancer matter? Br J Cancer 2009;101:S9-S12. [Crossref] [PubMed]
42. Crawford SC, Davis JA, Siddiqui NA, et al. The waiting time paradox: population based retrospective study of treatment delay and survival of women with endometrial cancer in Scotland. BMJ 2002;325:196. [Crossref] [PubMed]
43. Rupassara KS, Ponnusamy S, Withanage N, et al. A paradox explained? Patients with delayed diagnosis of symptomatic colorectal cancer have good prognosis. Colorectal Dis 2006;8:423-9. [Crossref] [PubMed]
44. Stapley S, Peters TJ, Sharp D, et al. The mortality of colorectal cancer in relation to the initial symptom at presentation to primary care and to the duration of symptoms: a cohort study using medical records. Br J Cancer 2006;95:1321-5. [Crossref] [PubMed]
45. Xu Y, Chen X, Zhang M, et al. Prognostic effect of parotid area lymph node metastases after preliminary diagnosis of nasopharyngeal carcinoma: a propensity score matching study. Cancer Med 2017;6:2213-21. [Crossref] [PubMed]
46. Pan CC, Lu J, Chen P, et al. Evaluation of the prognostic significance of refinement and stratification of distant metastasis status in 1016 cases of nasopharyngeal carcinoma. Zhonghua Zhong Liu Za Zhi 2013;35:595-9.
47. Lin M, Yang Q, You R, et al. Metastatic characteristics associated with survival of synchronous metastatic nasopharyngeal carcinoma in non-epidemic areas. Oral Oncol 2021;115:105200. [Crossref] [PubMed]
48. Liu Z, Yu Z, Chen D, et al. Pivotal roles of tumor-draining lymph nodes in the abscopal effects from combined immunotherapy and radiotherapy. Cancer Commun (Lond) 2022;42:971-86. [Crossref] [PubMed]
49. Sun XS, Wang JW, Han F, et al. Prognostic value of metastatic cervical lymph node stiffness in nasopharyngeal carcinoma: A prospective cohort study. Radiother Oncol 2023;189:109939. [Crossref] [PubMed]
50. Du XJ, Wang GY, Zhu XD, et al. Refining the 8th edition TNM classification for EBV related nasopharyngeal carcinoma. Cancer Cell 2024;42:464-473.e3.
51. Chen L, Mo DC, He JN, et al. The role of radiotherapy in distant nasopharyngeal carcinoma: a SEER analysis. Transl Cancer Res 2025;14:3047-56. [Crossref] [PubMed]
52. Pan JJ, Mai HQ, Ng WT, et al. Ninth Version of the AJCC and UICC Nasopharyngeal Cancer TNM Staging Classification. JAMA Oncol 2024;10:1627-35. [Crossref] [PubMed]
53. Li X, Ge Z, Chen W, et al. Potential biomarker for screening nasopharyngeal carcinoma: three-microRNA panel in serum. Transl Cancer Res 2025;14:3554-64. [Crossref] [PubMed]
54. Lai P, Chen X, Pei W, et al. Development and validation of an MRI radiomics-based interpretable machine learning model for predicting the progression-free survival in locally advanced nasopharyngeal carcinoma. Quant Imaging Med Surg 2025;15:5347-61. [Crossref] [PubMed]