Machine learning models based on XGBoost algorithm to predict prognosis of lung cancer brain metastases

Introduction

According to the Global Cancer Statistics 2020, lung cancer caused over 1.7 million cancer deaths worldwide in 2020, ranking first among all cancers (1). As one of the common reasons for treatment failure, about 10–25% of these patients already have lung cancer brain metastasis (LCBM) at the time of diagnosis while approximately 40% of lung cancer patients will develop LCBM during the course of disease (2). Brain metastases are associated with a significant decrease in patient quality of life and survival.

The current treatment options for LCBM, including surgery, radiotherapy, chemotherapy, and systemic therapies, often provide only palliative care (3-5). Despite advances in treatment, the prognosis for LCBM patients remains poor. The natural overall survival (OS) period of LCBM is 3 to 5 months, and the survival period can be improved by radiotherapy and chemotherapy for 1 to 3 months (6-9).

The treatment modalities utilized for LCBM exert a profound influence on the patient’s quality of life and survival. However, it is evident that LCBM patients exhibit significant heterogeneity, and the prognosis should be individually assessed (10,11). To individualize treatment precisely, it is imperative to identify distinct subgroups with varying prognoses. Therefore, there is an urgent need to develop a predictive model that can forecast the survival of patients with LCBM and assess the impact of different treatment modalities on survival, ultimately guiding the selection of clinically optimal treatment strategies.

Prior early research has yielded various predictive models for LCBM patients’ prognoses (12-15). However, the accuracy of these models leaves room for improvement, with the area under the curve (AUC) or concordance index (c-index) less than 0.75 (16). The advancement of LCBM patient survival prediction models has involved incorporating lung cancer driver gene mutation status, tyrosine kinase inhibitor (TKI) application, and select immunological features. These additions have improved model performance to some extent (10,11,14), but the models are limited by the potential need for complex gene testing and biomarker detection.

Recent advances in machine learning offer new opportunities for improving the prognosis and management of LCBM. Chen et al. successfully predicted the survival of LCBM patients by employing magnetic resonance imaging (MRI) radiomics and deep learning algorithms, with the model exhibiting commendable performance (17). However, that research was restricted in its capacity to solely differentiate whether a patient’s survival duration surpassed the median survival time.

Extreme gradient boosting (XGBoost) is a leading machine learning based boosting algorithm renowned for its remarkable efficiency, adaptability, and portability. Its ability to mitigate overfitting effectively boosts model performance during validation, guaranteeing superior outcomes (18,19). As a result, XGBoost has achieved widespread success across diverse studies and domains (20).

This research aims to build a machine learning model based on the XGBoost algorithm to predict the prognosis of synchronous LCBM patients using the Surveillance, Epidemiology, and End Results (SEER) database. Such a model, if validated in independent cohorts with complete molecular and treatment data, could potentially inform treatment decisions and contribute to improved patient outcomes. We present this article in accordance with the STARD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-0997/rc).

Methods

Data source

Data for this study came from SEER 17 registries (SEER November 2023 Submission dataset on SEER database). A total of 15,216 LCBM patients diagnosed between 2010 and 2014 were identified via SEER*Stat for model construction. A further 3,181 LCBM patients diagnosed in 2015 from SEER were used for temporal validation. Data on lung and bronchus cancer patients were selected. Inclusion criteria: (I) lung and bronchus cancer was the only cancer; (II) brain metastasis was present at diagnosis (i.e., synchronous brain metastases only); (III) diagnosis was per International Classification of Diseases for Oncology, 3rd Edition (ICD-O-3). Of note, the SEER database does not contain data on molecular markers (e.g., EGFR, ALK, PD-L1 status), targeted therapy, immunotherapy, performance status, or number of brain metastases, which limits the ability to fully account for contemporary treatment paradigms. The study used SEER public data for retrospective analysis. With personal information removed, it met ethical exemption, needed no institutional review board (IRB) approval or informed consent, and complied with SEER regulations. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Data process and algorithm

To address the remaining missing values and ensure data completeness, Multiple Imputation by Chained Equations (MICE) was applied, as it effectively preserves the multivariate relationships among features.

XGBoost is an efficient and scalable gradient boosting algorithm, boasting core advantages of high efficiency and anti-overfitting. It excels in machine learning tasks like classification, regression, and ranking, with significant application value in related fields. In addition to the XGBoost algorithm, we also used Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), and K-Nearest Neighbor (KNN) algorithms to construct comparative models, and ultimately selected the optimal one.

Feature selection and model validation

Univariate and multivariate Cox regression analyses were performed on clinical and pathological characteristics obtained from the SEER database. Characteristics that were statistically significant in the multivariate Cox analysis—including age at diagnosis, sex, race, primary site of tumor, histological type, income, surgical therapy for the primary site, radiotherapy, chemotherapy, liver metastases, bone metastases, tumor (T) stage and lymph node (N) stage—were incorporated into the machine learning model to predict 6-month, 1-, 2- and 3-year LCBM patients’ OS. The proportional hazards assumption was assessed using Schoenfeld residuals and was not violated for the variables included in the model. In this model, survival status was used as the response variable, where 1 represented survival and 0 represented death. Patients were randomized into training and test datasets in an 8:2 ratio.

To ensure the stability of the model, we use ten-fold cross-validation in the training set to find the optimal hyper-parameters. To assess the predictive validity and accuracy, we formed predicted receiver operating characteristic (ROC) curves and computed the corresponding AUC as well as confusion matrix. Pairwise comparisons of test-set AUC values between models were performed using the DeLong test. The Holm step-down procedure was applied to adjust for multiple comparisons and control the family-wise error rate at α=0.05, with 4 independent comparisons conducted at each follow-up time point.

Calibration curves were plotted to evaluate the calibration performance of all five models in the internal validation set, by comparing predicted survival probabilities against actual observed survival probabilities. Decision curve analysis was further performed to assess the clinical net benefit of each model across a range of clinically relevant high-risk thresholds.

Statistical analysis

Univariate Cox proportional hazards regression was used to identify variables associated with overall survival, and variables reaching statistical significance in univariate analysis were entered into a multivariate Cox proportional hazards regression model to identify independent prognostic factors. Baseline characteristics between surgical and non-surgical patients were compared using the Chi-squared test.

Propensity score matching (PSM) was employed to balance covariates between treated and control groups in this observational study. Importantly, the variables included in the propensity score model were selected based on their statistical significance in the multivariate Cox regression (i.e., a predictive modeling framework). While this approach reduces measured confounding, residual confounding from unmeasured variables—including performance status, precise disease burden, and molecular profiles—remains a limitation.

To examine the association between surgical intervention and LCBM patient prognosis, 1:1 nearest-neighbor PSM was performed with a caliper of 0.2 standard deviations of the logit of the propensity score, matching patients who received surgical therapy for the primary site to those who did not. Kaplan-Meier (K-M) survival analysis was then applied to compare OS between matched groups. Additionally, patients were stratified by sex, race, histologic type, and N-stage; 1:1 PSM was re-applied to balance baseline characteristics within each stratum, followed by K-M survival analysis. Given the observational nature of the data, these analyses should be interpreted as exploratory and associative rather than causal.

All statistical computations were performed using R (version 4.3.2), with statistical significance set at two-tailed P<0.05.

Code availability

The custom code used in this study is available in GitHub at https://github.com/Linastro/Machine-learning-models-based-on-XGBoost/tree/V1.0.1. The version tagged in the repository corresponds to DOI: https://zenodo.org/records/15542887. For inquiries regarding the code, please contact the corresponding author.

Results

Clinical characteristics of LCBM patients

The machine learning model was developed using a dataset of 15,216 LCBM patients. As summarized in Table 1, the cohort characteristics revealed: 18.4% (n=2,804) of patients were aged <54 years while 17.1% (n=2,601) were >75 years. The population was nearly equally distributed by sex (47.5% female, n=7,225) and predominantly White (77.2%). The upper lung lobe was the most common primary site (51.8%), and adenocarcinoma represented the majority of histologic types (68.0%). Approximately 32.0% of patients reported an annual household income exceeding US$80,000. Tumor staging distribution showed T1–T4 proportions of 10.8%, 29.3%, 4.6%, and 42.7% respectively, with 12.7% classified as T0/TX. As for N-stage, N0–N3 stages accounted for 21.2%, 8.3%, 45.2%, and 19.4%, while 6.0% were NX. Treatment modalities included surgery (3.2%), radiotherapy (75.9%), and chemotherapy (53.6%). Metastatic involvement was observed in 34.7% (bone) and 17.0% (liver) of cases.

Univariate and multivariate Cox regression analysis

Univariate Cox regression analysis identified all characteristics included in the study as statistically significant factors associated with LCBM patients’ OS (Table 2).

Multivariate Cox regression analysis was performed to adjust for measured confounders and identify variables independently associated with LCBM patients’ survival (Table 2). Increasing age was associated with progressively shorter survival [hazard ratio (HR) 1.181–1.573, all P<0.001]. Male sex was associated with worse survival compared to female [HR 1.162, 95% confidence interval (CI): 1.128–1.197, P<0.001]. Chemotherapy (HR 0.354, 95% CI: 0.342–0.366), radiotherapy (HR 0.798, 95% CI: 0.770–0.826), and surgery (HR 0.568, 95% CI: 0.519–0.622) were all associated with improved survival (all P<0.001); however, these associations should be interpreted with caution, as they likely reflect, at least in part, confounding by indication. Bone metastases (HR 1.242, 95% CI: 1.202–1.284) and liver metastases (HR 1.390, 95% CI: 1.334–1.448) were associated with worse survival. Race (Black vs. White: HR 0.967, 95% CI: 0.923–1.012, P=0.14) and certain income strata showed non-significant or modest effects.

Establishing and evaluating predictive models

Based on the above results of multivariate cox analysis, we developed an XGBoost-based predictive model to predict 6-month, 1-year, 2-year, and 3-year OS in LCBM patients. The cohort was grouped into training and testing sets at an 8:2 ratio. To enhance model robustness, ten-fold cross-validation was implemented on the training set for iterative hyperparameter optimization. ROC curve analysis with corresponding AUC calculations was subsequently performed for both training and test cohorts.

XGBoost model showed good performance in predicting LCBM patient survival, with AUC values for 6-month (test set: AUC =0.816), 2-year (test set: 0.796; train set: 0.805) and 3-year (test set: 0.810; train set: 0.817) survival prediction (Figure 1). Compared to traditional machine learning algorithms (LR, KNN, SVM, RF), the XGBoost model performed best, with AUC values at all time points higher than the other algorithms (ranging from 0.796 to 0.816) (Table 3).

Figure 1 Validation of XGBoost models from test and train data via ROC curve and AUC at 6 months, 1 year, 2 years, and 3 years. AUC, area under the curve; ROC, receiver operating characteristic; XGBoost, extreme gradient boosting.

DeLong test pairwise comparisons between XGBoost and the four competing models showed statistically significant differences across all four follow-up endpoints (all Holm-adjusted P<0.05, Table 4). The performance gaps were largest between XGBoost and SVM/KNN, intermediate with RF, and smallest with LR. The statistical significance of these differences increased with longer follow-up duration.

We further formed confusion matrix (Figure 2) to assess the predictive validity and accuracy of the XGBoost model. The 6-month survival prediction model was calculated to have an accuracy of 0.745. The 1-year survival prediction model was calculated to have an accuracy of 0.768. The 2-year survival prediction model was calculated to have an accuracy of 0.890. The 3-year survival prediction model was calculated to have an accuracy of 0.940.

Figure 2 Confusion matrix of the XGBoost model-predicted results in the test data. Cell colors represent the number of patients in each classification category, with darker shades indicating higher counts. XGBoost, extreme gradient boosting.

We generated calibration curves to evaluate the calibration performance of all five machine learning models in the internal validation set (Figure 3). The XGBoost model exhibited the best calibration across all four follow-up time points, with predicted probabilities most closely matching observed survival probabilities and its curve lying closest to the ideal diagonal reference line.

Figure 3 Calibration curves of five machine learning prognostic models for 6-month to 3-year survival prediction. KNN, K-Nearest Neighbor; LR, Logistic Regression; RF, Random Forest; SVM, Support Vector Machine; XGBoost, extreme gradient boosting.

Notably, the calibration performance of all models generally improved with longer follow-up duration. The XGBoost model maintained consistent and excellent calibration characteristics throughout the 6-month to 3-year follow-up period. These results, together with its superior discriminatory ability, confirm that the XGBoost model outperforms traditional machine learning algorithms in both discrimination and calibration for survival prediction in LCBM patients.

We performed decision curve analysis to evaluate the clinical net benefit of all five machine learning models (Figure 4). The XGBoost model achieved the highest net benefit across all four follow-up time points within the clinically relevant high-risk threshold range. The SVM model showed the lowest clinical net benefit, with a rapid decline in net benefit at lower thresholds. The LR, RF, and KNN models provided intermediate net benefit, all consistently inferior to the XGBoost model. The clinical advantage of the XGBoost model became more evident with longer follow-up duration. These findings, together with its superior discrimination and calibration, confirm that the XGBoost model is the optimal prognostic tool for predicting survival in LCBM patients.

Figure 4 Decision curve analysis of five machine learning prognostic models for 6-month to 3-year survival prediction. KNN, K-Nearest Neighbor; LR, Logistic Regression; RF, Random Forest; SVM, Support Vector Machine; XGBoost, extreme gradient boosting.

In order to further validate our XGBoost models, we obtained 3181 LCBM patients diagnosed in 2015 as the temporal validation dataset from the SEER database. Then, we used our XGBoost models to predict 6-month, 1-, 2- or 3-year OS of this external cohort. It showed that our predictive models kept performing well on the temporal validation dataset, demonstrating high robustness (6-month: AUC =0.822; 1-year: AUC =0.790; 2-year: AUC =0.786; 3-year: AUC =0.780) (Figure 5).

Figure 5 External validation of XGBoost models via ROC curve and AUC at 6 months, 1 year, 2 years, and 3 years. AUC, area under the curve; ROC, receiver operating characteristic; XGBoost, extreme gradient boosting.

In addition, we evaluated the importance of clinical features in the model. It is important to note that feature importance in the XGBoost model reflects predictive contribution within this dataset and should not be interpreted as reflecting causal or time-dependent biological effects. The results (Figure 6) showed that chemotherapy was the variable with the highest predictive importance across all time horizons, which likely reflects treatment selection patterns and the strong association between receiving chemotherapy and overall prognosis. In the short- and medium-term models (6-month, 1-year and 2-year), age at diagnosis was the second most important variable, whereas race was the second most important in the 3-year model. Interestingly, radiotherapy contributed substantially only in the 6-month model, while its contribution was negligible in the medium- and long-term models. Conversely, surgery, race, and histologic type showed negligible contribution in the 6-month model but greater importance in the 1-, 2-, and 3-year models. These temporal patterns in feature importance should be considered hypothesis-generating given the observational nature of the data and the potential influence of treatment selection patterns on variable rankings.

Figure 6 Ranking the importance of clinical features in XGBoost prognostic modeling. N, lymph node; T, tumor; XGBoost, extreme gradient boosting.

Benefits of surgical treatment in LCBM patients

The therapeutic role of surgical intervention for the primary site in LCBM patients remains controversial. To explore this, we first compared baseline characteristics between surgical and non-surgical patients: Table 5 shows significant baseline differences before PSM, which were largely resolved after PSM adjustment. K-M survival analysis in the PSM-adjusted cohort showed that surgery was associated with improved OS, with a 33.7% reduction in mortality risk (HR 0.663, 95% CI: 0.587–0.750, Figure S1). However, potential immortal time bias should be noted, as patients who underwent surgery necessarily survived long enough to receive the intervention.

We further conducted exploratory subgroup analyses to examine whether the observed association between surgery and survival varied across clinical subgroups. By sex, both females and males showed an association with improved survival, with a larger magnitude observed in females (46.9% vs. 22.2% mortality risk reduction; Figure S2). By race, Black patients showed no significant association between surgery and survival, while White patients and those of other races showed notable survival differences (36.4% and 43.1% mortality risk reduction, respectively; Figure S3). By histologic type, the association was observed in patients with adenocarcinoma (38.1% mortality risk reduction), whereas patients with squamous cell carcinoma and other histologic types did not show a significant association (Figure S4). By N-stage, significant associations were observed in N0 and N2-stage patients (41.9% and 31.6% mortality risk reduction, respectively), but not in N1 or N3-stage patients (Figure S5). For concomitant metastases, surgery was not significantly associated with improved survival in patients with bone or liver metastases, but was associated with improved survival in those without—showing 36.8% (no bone metastases) and 38.5% (no liver metastases) mortality risk reduction (Figure S6). These subgroup analyses should be interpreted as exploratory given the risks of multiple comparisons and reduced statistical power in stratified analyses.

Discussion

The brain is the most lethal metastatic site for lung cancer, yet reliable prognostic tools for LCBM remain insufficient in clinical practice. Traditional models [e.g., recursive partitioning analysis (RPA), graded prognostic assessment (GPA), basic score for brain metastases (BS-BM)] and previous SEER-based LCBM models have exhibited suboptimal performance (AUC/c-index <0.75) (12-15,21-24). This study developed and compared five machine learning models for LCBM survival prediction. The XGBoost model achieved the highest AUC values, ranging from 0.796 to 0.816 in the test set, which compares favorably with previously reported models. DeLong test with Holm correction confirmed statistically significant differences in AUC between XGBoost and all four competing algorithms across all follow-up time points (all adjusted P<0.05). Furthermore, the XGBoost model also demonstrated superior calibration and higher clinical net benefit compared to traditional machine learning algorithms. These comprehensive performance evaluations provide robust evidence supporting the superiority of the XGBoost model as a promising prognostic tool for LCBM patients.

Previous notable models predict LCBM patient survival based on factors such as age, Karnofsky Performance Status (KPS), primary tumor control status, presence of extracranial metastases, the number of brain metastatic lesions, lung cancer driver gene mutation status, TKI drug application, and select immunological features (12-15,21). In this study, younger age, female sex, non-White/non-Black race, upper lobe primary site, adenocarcinoma histology, higher household income, lower T-stage and N-stage, and absence of liver and bone metastases were identified as variables associated with more favorable prognosis. While chemotherapy, radiotherapy, and surgery also showed associations with improved survival, these treatment-related variables are subject to confounding by indication and should not be interpreted as independent causal factors.

Ever since the RPA study in 1997, multiple previous studies have unanimously and conclusively confirmed that age is associated with the prognosis of LCBM patients (12,25). Moreover, the risk of developing LCBM decreases with age (26). In our research, we ascertained that the older the age at diagnosis, the poorer the prognosis for LCBM patients, which is consistent with the findings of prior studies.

The prognostic role of sex in LCBM remains controversial: some studies report better survival in females (22-24,27,28), while a meta-analysis suggested sex is not a valid prognostic marker (29). In our study, female LCBM patients showed a greater magnitude of association between surgery and survival compared to males, consistent with previous surgical cohort findings (30,31). It has been hypothesized that sex-based differences may relate to differences in driver mutation prevalence—such as higher EGFR mutation rates in female adenocarcinoma patients—and potential hormonal influences on tumor biology (32,33); however, these mechanistic hypotheses cannot be directly assessed in this study given the absence of molecular and endocrine data in the SEER database. Further research incorporating molecular profiling is needed to clarify the role of sex in LCBM prognosis.

In the U.S., better healthcare access among high-income and White populations may contribute to differences in LCBM prognosis (34). Prior studies have identified African American race as associated with higher mortality in brain metastases (HR =1.31, 95% CI: 1.12–1.53), though this disparity diminishes after adjusting for treatment access (24). In our multivariate analysis, the association between Black race and prognosis was not statistically significant (HR 0.967, 95% CI: 0.923–1.012, P=0.14), and the discussion of racial differences should be tempered accordingly. Notably, only White and non-African American patients showed an association between surgery and improved survival in subgroup analyses, though the sample size for Black patients was limited. Previous findings on racial differences in surgical outcomes are inconsistent: a 224-patient analysis found no significant survival difference between African Americans and Whites after stage matching (11.2 vs. 12.4 months, P=0.21) (35). The observed differences, if real, may relate to complex interactions between socioeconomic factors, healthcare access, and potential biological differences, requiring further investigation in studies with adequate representation and comprehensive covariate adjustment.

Clearly, the prognosis of patients varies according to different histological types (36). In this investigation, we discovered that patients with adenocarcinoma had a better prognosis than patients with other histological types, and histology ranked relatively high in the model’s sensitivity analysis. In the subgroup analysis of surgical OS benefit by lung cancer histological type, this study found that only LCBM patients with adenocarcinoma histology could benefit from surgery. Another previous study based on the SEER database (n=95,191) further confirmed that patients with adenocarcinoma brain metastasis who underwent surgery had a 5-year OS rate of 18.7%, which was significantly higher than that of patients with squamous cell carcinoma (9.3%), a finding consistent with the conclusions of this study (24). The differences may be attributed to several factors. On one hand, the biological characteristics of the metastatic lesions play a role. Brain metastases from lung adenocarcinoma often present as solitary or oligometastatic lesions, predominantly located in the superficial regions of the brain parenchyma, resulting in a higher rate of complete surgical resection (exceeding 80%). In contrast, brain metastases from squamous cell carcinoma are frequently associated with extensive vascular invasion or diffuse lesions, making surgical resection more challenging (37). Additionally, squamous cell carcinoma, lacking driver gene mutations, cannot benefit from targeted therapies as adenocarcinoma patients can, thereby increasing the risk of postoperative recurrence.

Surgical intervention for brain metastasis may provide a potential avenue for extending survival in LCBM patients (37). Thoracic management has been identified as an important predictor of survival. However, whether surgical therapy for the primary site is associated with prolonged survival in LCBM patients has long been debated. Through PSM, this study found that surgery was associated with a significant reduction in mortality risk in the overall cohort (HR =0.663, 95% CI: 0.587–0.750). It is important to emphasize that this association does not establish causality, and the decision to perform surgery is influenced by patient selection factors that cannot be fully accounted for in this observational analysis.

Previous studies have indicated that pulmonary resection and neurosurgical intervention have provided substantial survival benefits for LCBM patients (38). Nevertheless, the benefits of surgical treatment for LCBM patients may vary heterogeneously based on different subgroup patients (39). The relationship between surgical OS benefit in LCBM patients and variables such as gender, ethnicity, and histological type has already been discussed earlier.

Previous studies have found that for patients with N0–N1 disease, early surgical intervention of the primary tumor may reduce the risk of secondary metastasis (40). In contrast, patients with N2 disease have significantly worse postoperative survival compared with those with N0–N1 disease. A study involving 204 patients with III-N2 disease showed that the median OS in the surgical group was 28.3 months, which was better than that in the radiotherapy group (21.6 months), but the 5-year survival rate was only 34.1% (41). Notably, for patients with multi-station N2 metastasis, the median OS after surgery dropped sharply to 18.9 months (42). In the present exploratory subgroup analysis, the association between surgery and survival appeared to vary by N-stage. However, no significant association was observed in patients with N1 disease, which may relate to the relatively small proportion of patients in this subgroup. These N-stage-stratified results should be interpreted as exploratory and require validation in larger cohorts.

Extracranial metastases are another important aspect to consider. Previous studies have identified liver metastasis as the worst prognostic metastatic site for patients with non-small cell lung cancer (NSCLC), with a HR of 1.68 (95% CI: 1.58–1.79), which significantly diminishes the survival benefit of surgery (43). Another study demonstrated that brain metastasis patients with concomitant bone metastasis had a significantly higher risk of postoperative recurrence (HR =2.1, P=0.008), highlighting the need for a comprehensive assessment of the overall disease burden (37). In the present study, subgroup analysis after PSM revealed that LCBM patients with liver or bone metastasis did not achieve a survival benefit in terms of OS from surgery, consistent with previous findings.

Apart from surgery, radiotherapy is also an effective strategy for local control of brain metastases. Radiotherapy modalities include whole brain radiotherapy (WBRT), γ-knife, stereotactic radiotherapy (SRS), among others (6-8,44). A study from the Lancet (45) demonstrated that for LCBM patients who are not eligible for surgery or SRS and who have no extracranial metastases, WBRT may provide potential survival advantages. A predictive model investigation from 2022 also endorsed the viewpoint (8). The survival gains of SRS for LCBM patients also rely on factors such as the existence of extracranial metastases (6), and its local control of brain metastases is comparable to that of brain metastasis tumor resection surgery (38). γ-knife surgery represents another safe and effective therapeutic option that can ameliorate the quality of life for LCBM patients (44). In this investigation, LCBM patients who were treated with radiotherapy demonstrated a 20.2% decrease in the risk of mortality, validating that LCBM can derive advantages from radiotherapy.

Systemic therapy, comprising chemotherapy, constitutes the foundation of treatment for patients with metastatic lung cancer (3). In this study, the receipt of chemotherapy emerged as the variable with the highest predictive importance for patient survival. However, this finding should be interpreted with caution: patients who did not receive chemotherapy may have had contraindications, poor performance status, or limited access to care—factors not fully captured in the SEER database—and this represents a form of confounding by indication. Considering the recent developments in targeted therapy and immunotherapy, the systemic treatment of LCBM has evolved substantially. Studies have shown that lung cancer patients with driver gene mutations exhibit a higher incidence of brain metastases, but their prognosis improves when TKI therapy is combined with treatment (3). Consequently, some models have integrated driver gene mutation status into their frameworks (11). However, for patients without actionable mutations, chemotherapy combined with immunotherapy continues to demonstrate disease control benefits. The absence of molecular and immunotherapy data in SEER represents an important limitation of the current model.

This study has several important limitations that warrant careful consideration. First, due to the constraints of the SEER database, the model did not incorporate driver gene mutation status (EGFR, ALK, ROS1, etc.), targeted therapy, or immunotherapy data, which are critical in the contemporary management of LCBM. The SEER database also lacks information on performance status, the number and size of brain metastases, and detailed treatment sequencing, all of which influence prognosis and treatment decisions. Second, the study population was restricted to patients with synchronous brain metastases at initial diagnosis and does not capture patients who develop brain metastases during the disease course, limiting generalizability. Third, the external validation cohort, although temporally distinct (diagnosed in 2015), was derived from the same SEER registries and should be more accurately described as temporal validation rather than true external validation; the model’s performance in geographically and demographically independent populations remains unknown. Fourth, the inclusion of treatment variables (chemotherapy, radiotherapy, surgery) in the predictive model limits its utility as a baseline prognostic tool, as these variables reflect post-diagnosis treatment decisions rather than pre-treatment patient characteristics. Fifth, despite the use of PSM, residual confounding from unmeasured variables (particularly performance status and disease burden) and potential immortal time bias cannot be excluded, and the observed associations with surgical treatment should be interpreted as associative rather than causal. Sixth, the subgroup analyses involve multiple comparisons and reduced sample sizes, and should be considered exploratory rather than definitive. Finally, the study population was primarily derived from North American registries, and the model’s applicability to populations in other regions with different healthcare systems and demographic profiles requires further validation.

Despite these limitations, this study demonstrates the feasibility of applying machine learning approaches to LCBM survival prediction using routinely collected clinical data, and identifies important directions for future research, including the incorporation of molecular and treatment data and validation in independent cohorts.

Conclusions

In conclusion, this study developed an XGBoost-based machine learning model for predicting short- and long-term survival outcomes in patients with synchronous LCBM using SEER data. The model achieved moderate-to-good discrimination (AUC 0.780–0.822 in temporal validation) and compared favorably with previously reported models, though differences from simpler algorithms were modest. Chemotherapy emerged as the variable with the highest predictive importance, though this finding should be interpreted in the context of confounding by indication. Exploratory subgroup analyses suggest that the association between primary site surgery and survival may vary across clinical subgroups defined by sex, race, histologic type, N-stage, and extracranial metastasis status. These findings should be considered hypothesis-generating. Future studies incorporating molecular markers, contemporary systemic therapies, and prospective external validation are needed to develop clinically actionable prognostic tools for LCBM. Treatment decisions for LCBM patients should continue to be individualized based on comprehensive clinical assessment rather than model predictions alone.

Acknowledgments

During the preparation of this work, the authors used Doubao AI Service for part of the language translation work. After using this service, the authors reviewed and edited the content as needed and takes full responsibility for the content of the published article.

Footnote

Reporting Checklist: The authors have completed the STARD reporting checklist. Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-0997/rc

Peer Review File: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-0997/prf

Funding: This work was supported by the Key Clinical Projects of Peking University Third Hospital (grant Nos. BYSYRCYJ2023001 and BYSYZD2025049).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-0997/coif). The 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. Sung H, Ferlay J, Siegel RL, et al. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin 2021;71:209-49. [Crossref] [PubMed]
2. Andre F, Grunenwald D, Pujol JL, et al. Patterns of relapse of N2 nonsmall-cell lung carcinoma patients treated with preoperative chemotherapy: should prophylactic cranial irradiation be reconsidered? Cancer 2001;91:2394-400.
3. Pellerino A, Bruno F, Rudà R, et al. Systemic Therapy for Lung Cancer Brain Metastases. Curr Treat Options Oncol 2021;22:110. [Crossref] [PubMed]
4. Zhao J, Li X. Clinical Practice Guidelines for the Management of Brain Metastases from Non-small Cell Lung Cancer with Actionable Gene Alterations in China (2025 Edition). Chinese Journal of Lung Cancer 2025;28:1-21.
5. Khan A, Zhao HF, Meng H, et al. Stereotactic radiotherapy for brain metastases of non-small cell lung cancer: A comprehensive review. World J Radiol 2025;17:111076. [Crossref] [PubMed]
6. Zindler JD, Jochems A, Lagerwaard FJ, et al. Individualized early death and long-term survival prediction after stereotactic radiosurgery for brain metastases of non-small cell lung cancer: Two externally validated nomograms. Radiother Oncol 2017;123:189-94. [Crossref] [PubMed]
7. Zhang J, Jin J, Ai Y, et al. Computer Tomography Radiomics-Based Nomogram in the Survival Prediction for Brain Metastases From Non-Small Cell Lung Cancer Underwent Whole Brain Radiotherapy. Front Oncol 2020;10:610691. [Crossref] [PubMed]
8. Trikhirhisthit K, Setakornnukul J, Thephamongkhol K. Added survival benefit of whole brain radiotherapy in brain metastatic non-small cell lung cancer: Development and external validation of an individual prediction model. Front Oncol 2022;12:911835. [Crossref] [PubMed]
9. Tanzhu G, Peng H, Chen L, et al. Tumor treating fields combined with ionizing radiation inhibit the malignant phenotype of lung cancer brain metastasis cells by suppressing DNA damage repair pathways. Brain Stimul 2025;18:1426-40. [Crossref] [PubMed]
10. Li J, Jing W, Zhai X, et al. Estimating Survival in Patients with Non-Small-Cell Lung Cancer and Brain Metastases: A Verification of the Graded Prognostic Assessment for Lung Cancer Using Molecular Markers (Lung-molGPA). Onco Targets Ther 2021;14:1623-31. [Crossref] [PubMed]
11. Cacho-Díaz B, Cuapantécatl LD, Garcilazo-Reyes YJ, et al. Targeted Therapies and Utility of the Lung-molGPA in Non-Small-Cell Lung Cancer Patients with Brain Metastases. Oncology 2022;100:542-54. [Crossref] [PubMed]
12. Gaspar L, Scott C, Rotman M, et al. Recursive partitioning analysis (RPA) of prognostic factors in three Radiation Therapy Oncology Group (RTOG) brain metastases trials. Int J Radiat Oncol Biol Phys 1997;37:745-51. [Crossref] [PubMed]
13. Lorenzoni J, Devriendt D, Massager N, et al. Radiosurgery for treatment of brain metastases: estimation of patient eligibility using three stratification systems. Int J Radiat Oncol Biol Phys 2004;60:218-24. [Crossref] [PubMed]
14. Sperduto PW, Berkey B, Gaspar LE, et al. A new prognostic index and comparison to three other indices for patients with brain metastases: an analysis of 1,960 patients in the RTOG database. Int J Radiat Oncol Biol Phys 2008;70:510-4. [Crossref] [PubMed]
15. Sperduto PW, Kased N, Roberge D, et al. Summary report on the graded prognostic assessment: an accurate and facile diagnosis-specific tool to estimate survival for patients with brain metastases. J Clin Oncol 2012;30:419-25. [Crossref] [PubMed]
16. Zhu LH, Fan XW, Sun L, et al. New prognostic system specific for epidermal growth factor receptor-mutated lung cancer brain metastasis. Front Oncol 2023;13:1093084. [Crossref] [PubMed]
17. Chen BT, Jin T, Ye N, et al. Predicting Survival Duration With MRI Radiomics of Brain Metastases From Non-small Cell Lung Cancer. Front Oncol 2021;11:621088. [Crossref] [PubMed]
18. Sharma K, Tiwari PK, Sinha SK. Estimation of Hematocrit Volume Using Blood Glucose Concentration through Extreme Gradient Boosting Regressor Machine Learning Model. J Chem Inf Model 2025;65:1736-46. [Crossref] [PubMed]
19. Zhang Z, Wu J, Wang J, et al. Extreme gradient boosting-based explainable machine learning model for predicting significant fibrosis in autoimmune hepatitis. QJM 2026;119:34-41. [Crossref] [PubMed]
20. Huang Z, Hu C, Chi C, et al. An Artificial Intelligence Model for Predicting 1-Year Survival of Bone Metastases in Non-Small-Cell Lung Cancer Patients Based on XGBoost Algorithm. Biomed Res Int 2020;2020:3462363. [Crossref] [PubMed]
21. Chen K, Yu X, Zhang F, et al. Applicability of the lung-molGPA index in non-small cell lung cancer patients with different gene alterations and brain metastases. Lung Cancer 2018;125:8-13. [Crossref] [PubMed]
22. Peng S, Xiao Y, Li X, et al. A nomogram for predicting overall survival rate in patients with brain metastatic non-small cell lung cancer. Medicine (Baltimore) 2022;101:e30824. [Crossref] [PubMed]
23. Yang F, Gao L, Wang Q, et al. Development and validation of nomograms to predict early death in non-small cell lung cancer patients with brain metastasis: a retrospective study in the SEER database. Transl Cancer Res 2023;12:473-89. [Crossref] [PubMed]
24. Zhang GH, Liu YJ, De Ji M. Risk Factors, Prognosis, and a New Nomogram for Predicting Cancer-Specific Survival Among Lung Cancer Patients with Brain Metastasis: A Retrospective Study Based on SEER. Lung 2022;200:83-93. [Crossref] [PubMed]
25. Berger A, Mullen R, Bernstein K, et al. Extended Survival in Patients With Non-Small-Cell Lung Cancer-Associated Brain Metastases in the Modern Era. Neurosurgery 2023;93:50-9. [Crossref] [PubMed]
26. An N, Jing W, Wang H, et al. Risk factors for brain metastases in patients with non-small-cell lung cancer. Cancer Med 2018;7:6357-64. [Crossref] [PubMed]
27. Jacot W, Quantin X, Boher JM, et al. Brain metastases at the time of presentation of non-small cell lung cancer: a multi-centric AERIO analysis of prognostic factors. Br J Cancer 2001;84:903-9. [Crossref] [PubMed]
28. Gupta S, Singh S, Chophy A, et al. Analysis of prognostic factors in patients with brain metastases affecting survival. J Egypt Natl Canc Inst 2022;34:45. [Crossref] [PubMed]
29. Sun DS, Hu LK, Cai Y, et al. A systematic review of risk factors for brain metastases and value of prophylactic cranial irradiation in non-small cell lung cancer. Asian Pac J Cancer Prev 2014;15:1233-9. [Crossref] [PubMed]
30. Wroński M, Arbit E, Burt M, et al. Survival after surgical treatment of brain metastases from lung cancer: a follow-up study of 231 patients treated between 1976 and 1991. J Neurosurg 1995;83:605-16. [Crossref] [PubMed]
31. Baysungur V, Tezel Ç, Kıral H, et al. Results of surgical resection in lung cancer with synchronous brain metastasis. Turk Gogus Kalp Damar Cerrahisi Derg 2019;27:192-8. [Crossref] [PubMed]
32. Perng PS, Hsu HJ, Lee JS, et al. Outcomes of surgery and subsequent therapy for central nervous system oligoprogression in EGFR-mutated NSCLC patients. World J Surg Oncol 2023;21:368. [Crossref] [PubMed]
33. Dmukauskas M, Cioffi G, Waite KA, et al. Sex Difference in Disease-Related Adverse Events Post-Diagnosis of Lung Cancer Brain Metastases in Medicare Individuals ≥ 66 Years of Age. Cancers (Basel) 2024;16:2986. [Crossref] [PubMed]
34. Williams CD, Salama JK, Moghanaki D, et al. Impact of Race on Treatment and Survival among U.S. Veterans with Early-Stage Lung Cancer. J Thorac Oncol 2016;11:1672-81.
35. ByunJKimMLeeSPredictors of Prognosis of Surgically Resected Brain Metastasis from Lung Cancer with a Focus on Brain Recurrence.2021. doi: .
36. Waqar SN, Samson PP, Robinson CG, et al. Non-small-cell Lung Cancer With Brain Metastasis at Presentation. Clin Lung Cancer 2018;19:e373-9. [Crossref] [PubMed]
37. Byun J, Kim JH, Kim M, et al. Survival Outcomes and Predictors for Recurrence of Surgically Treated Brain Metastasis From Non-Small Cell Lung Cancer. Brain Tumor Res Treat 2022;10:172-82. [Crossref] [PubMed]
38. Bougie E, Masson-Côté L, Mathieu D. Comparison Between Surgical Resection and Stereotactic Radiosurgery in Patients with a Single Brain Metastasis from Non-Small Cell Lung Cancer. World Neurosurg 2015;83:900-6. [Crossref] [PubMed]
39. Wang ZN, Jiang XB, Lu J, et al. Survival Benefit from Surgical Resection in Lung Cancer Patients with Brain Metastases: a Single-Center, Propensity-Matched Analysis Cohort Study. Ann Surg Oncol 2022;29:3684-93. [Crossref] [PubMed]
40. Yang Z, Chen H, Jin T, et al. The Impact of Time Interval on Prognosis in Patients with Non-Small Cell Lung Cancer Brain Metastases After Metastases Surgery. World Neurosurg 2023;180:e171-82. [Crossref] [PubMed]
41. Yang J, Jiang C, Wang Y, et al. Surgery is a promising strategy for improving overall survival for stage III-N2 non-small cell lung cancer patients. J Thorac Dis 2023;15:1426-32. [Crossref] [PubMed]
42. Yang Z, Wei S, Xia L, et al. Prognostic Factors of N2 Stage Non-small Cell Lung Cancer after Surgery. Chinese Journal of Lung Cancer 2020;23:15-20.
43. Wei S, Wei W, Wu B, et al. The Incidence and Effect of Different Organ Metastasis on the Prognosis of NSCLC. Thorac Cardiovasc Surg 2024;72:217-26. [Crossref] [PubMed]
44. Bragstad S, Flatebø M, Natvig GK, et al. Predictors of quality of life and survival following Gamma Knife surgery for lung cancer brain metastases: a prospective study. J Neurosurg 2018;129:71-83. [Crossref] [PubMed]
45. Mulvenna P, Nankivell M, Barton R, et al. Dexamethasone and supportive care with or without whole brain radiotherapy in treating patients with non-small cell lung cancer with brain metastases unsuitable for resection or stereotactic radiotherapy (QUARTZ): results from a phase 3, non-inferiority, randomised trial. Lancet 2016;388:2004-14. [Crossref] [PubMed]