Forecasting the risk of early death among patients suffering from lung cancer with brain metastasis after radiotherapy using interpretable machine learning: a study based on the SEER database

Abstract

Background

Lung cancer with brain metastasis (LCBM) significantly shortens patient survival. Accurately predicting individual prognosis remains challenging. This study aimed to identify key prognostic factors in LCBM patients after radiotherapy for the development of an interpretable machine learning (ML) model to support clinical decision-making and precision medicine.

Methods

Based on clinicopathological data from the U.S. Surveillance, Epidemiology, and End Results (SEER) database, patients were divided into training (70%) and validation (30%) cohorts. Thirteen variables associated with early death were screened by least absolute shrinkage and selection operator (LASSO) regression for model construction. Seven ML-based models were compared using area under the curve (AUC) values, calibration and decision curves, specificity, precision, and F1-score. SHapley Additive exPlanations (SHAP) analysis was applied to interpret the optimal model.

Results

The Light Gradient Boosting Machine (LightGBM) model achieved satisfactory performance in the validation set, with an AUC of 0.776, and showed good accuracy and clinical utility. SHAP analysis revealed that chemotherapy was associated with a lower risk of early death, while younger age and lower T stage were also associated with better outcomes. Conversely, bone, liver, and lung metastases were associated with a higher risk of early death.

Conclusions

This ML-based prediction model may help quantify the risk of early death in LCBM patients after radiotherapy, providing references for clinicians to improve prognostic evaluation and optimize treatment strategies.

Highlight box.

Key findings

• The Light Gradient Boosting Machine algorithm demonstrated satisfactory predictive performance, achieving an area under the curve of 0.776 in the validation cohort. Administration of chemotherapy emerged as the predominant protective factor associated with reduced early mortality. Conversely, the presence of extrathoracic metastases—specifically to bone, liver, or lung—constituted the principal risk factors for increased mortality.

What is known and what is new?

• Lung cancer with brain metastasis (LCBM) is associated with an unfavorable prognosis, with survival outcomes influenced by conventional clinicopathological parameters, including age and tumor stage.

• This investigation presents the inaugural comparative evaluation of seven machine learning algorithms specifically among patients with LCBM after radiotherapy. The application of SHapley Additive exPlanations interpretability analysis quantifies variable-specific contributions, identifying chemotherapy as the predominant modifiable prognostic factor.

What is the implication, and what should change now?

• The principal implication is enhanced prognostic precision for individual patients. Clinically, these findings support more systematic integration of chemotherapy with radiotherapy among eligible patients. Therapeutic strategies should be individualized according to patient-specific risk profiles generated by the model.

• The developed model facilitates personalized risk stratification, enabling clinicians to identify high-risk patients warranting intensified surveillance. Essential subsequent actions include prospective validation of these findings and potential incorporation into clinical guidelines to promote combined modality treatment for improved survival outcomes.

Introduction

Lung cancer (LC) serves as the chief cause of deaths related to cancers around the globe, and lung cancer with brain metastasis (LCBM) accounts for a relatively high proportion of those deaths, approximately 50% of all cancers with brain metastases (1-3). With the continuous innovation in treatment techniques for primary tumors and the continuous advancement in imaging technologies, the survival duration of patients has been considerably extended, triggering an increase year with year in diagnosed cases of LCBM. Data indicate that roughly 30% of patients suffering from non-small cell lung cancer (NSCLC) have brain metastasis at first diagnosis. As the NSCLC progresses, that proportion further increases, and ultimately, about 60% of patients develop NSCLC with brain metastasis. The prognoses of patients with LCBM are extremely poor, with a median overall survival (OS) limited to 4–9 months for untreated NSCLC patients with brain metastasis (4). On the basis of the Graded Prognostic Assessment index, the median OS for NSCLC patients with brain metastasis is 7 months. The median OS by whole-brain radiation therapy (WBRT) is 1–6 months. Patients undergoing WBRT have a 1-year survival rate of roughly 10–20%.

Due to the blood-brain barrier, conventional systemic treatment has difficulty in effectively reaching the areas of brain lesions to produce its efficacy (5,6). Local therapies targeting the brain, such as radiation therapy or neurosurgical resection, are usually adopted in clinical practices (7,8). As a key local therapy for LCBM, radiation therapy mainly includes stereotactic radiosurgery (SRS) and WBRT. In recent years, radiotherapy + immunotherapy and targeted therapy, to some extent, have improved the prognoses of patients with LCBM (9-11). Nevertheless, their risk of early mortality remains high, and intracranial hypertension, brain herniation, and damage to important brain functional areas are common causes (12,13). Currently, there is a lack of research on the risk factors for early death after radiotherapy. Therefore, accurately identifying high-risk factors and predicting survival risks are crucial for the optimization of clinical decisions.

In the field of developing prediction models, the nomogram has been the most extensively employed tool in clinical practice. Machine learning (ML), in recent years, has gained increasing attention from an increasing number of medical professionals due to its unique practicality, innovation, and excellent prediction accuracy. Previous studies have confirmed the potential of ML in medical prediction. However, the clinical application of ML is limited by insufficient interpretability, and its model decision process lacks intuitive interpretation, leading to a major obstacle for its promotion. This study innovatively combined an advanced ML algorithm with the SHapley Additive exPlanations (SHAP) framework. While retaining the predictive advantages of the algorithm, it also revealed key influencing factors through visual interpretation, solving the ‘black box’ problem of traditional models. Hence, it might provide an understandable basis for clinical decision-making and promote medical ML to become more interpretable and personalized.

This study selected patients with LCBM who received radiotherapy from the Surveillance, Epidemiology, and End Results (SEER) database as its study subjects and constructed an interpretable ML model for forecasting the risk of early mortality among patients suffering from LCBM after radiotherapy. This study is the first to employ the SHAP framework for the prediction of early mortality among patients suffering from LCBM, providing methodological guidance for similar studies and facilitating the development of prognostic models with greater clinical utility. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-0966/rc).

Methods

Collection of data

The data in this study were derived from the SEER database (http://seer.cancer.gov/), one of the authoritative large-scale databases of cancer registration around the U.S. The SEER database, supported by the National Cancer Institute, collects data from nearly 30 large-scale cancer registries across the country, covering approximately 18% of the national population. No approval was obtained for this study from an ethics committee or consent from patients, as the SEER data is publicly accessible and does not contain identifiable personal information of patients. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

SEER*Stat (version 8.4.3) was adopted to retrieve clinical data of patients who suffered from LCBM and had received radiotherapy, from the SEER database between 2010 and 2017. The data screening procedure is shown in Figure S1. Exclusion criteria were implemented to ensure data quality, including (I) unknown survival time; (II) unknown race and marital status; (III) unclear tumor, node, metastasis (TNM) stage; (IV) unclear surgery and lymph node biopsy positivity; and (V) unclear information on liver metastasis, bone metastasis, and lung metastasis.

Clinical and demographic data of patients suffering from LCBM undergoing radiotherapy were derived from the SEER database. Specifically, those data comprised age at diagnosis, sex, race, marital status, primary site, pathological type, grade, laterality, surgery, chemotherapy, T (tumor) stage, N (node) stage, bone metastasis, lung metastasis, liver metastasis, and lymph node positivity (lymph node biopsy positivity and lymph node detection positivity). Based on age, patients were divided into three groups (≤60, 61–70, >70 years). On the basis of pathological type, patients were roughly classified into three groups: adenocarcinoma, squamous cell carcinoma, and other types of cancers, referring to the International Classification of Diseases for Oncology Third Edition (ICD-O-3). Based on the degree of differentiation, patients were divided into three groups (grade I–II, grade III–IV, unknown). The tumor sequence represents the order of malignant tumors that can be reported in a lifetime and is classified as primary tumors and other tumors. Early death was the endpoint of this study, which meant the death of patients with LCBM within 3 months after radiotherapy (14-21).

Statistical analysis

Statistical analysis was performed using R (version 4.5.1) with various packages, comprising Survival, cmprsk, mstate, riskRegression, pec, dcurves, readxl, dplyr, Glmnet, compareGroups, rms, pROC, caret, rmda, e1071, rpart, neuralnet, randomForest, xgboost, Matrix, lightgbm, shapviz, and dplyr. Categorical variables in this study were described leveraging the number (N) and percentage (%). The Chi-squared test or Fisher’s exact test was employed for comparisons among groups. Continuous variables that were normally distributed were displayed as mean ± standard deviation, with comparisons among groups conducted through the independent sample t-test. Continuous variables without normal distribution were described as median (quartile), with comparisons among groups performed via the Mann-Whitney U test.

All patients were randomly classified into a training set (70%) and a validation set (30%) with the random seed set to 123. The least absolute shrinkage and selection operator (LASSO) was leveraged to initially identify relevant predictive features and determine the prediction variables significantly associated with early death. Subsequently, the variables screened out by LASSO were incorporated into multiple ML algorithms, including logistic regression (LR), naive Bayes (NB), support vector machine (SVM), decision tree (DT), random forest (RF), eXtreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM), to construct prediction models. The predictive performance of each model was measured by the area under the curve (AUC) values. Detailed parameter settings and optimization strategies for the seven ML-based models are presented in Table S1. The models were compared and assessed through various indicators, comprising precision, sensitivity, specificity, negative predictive value (NPV), positive predictive value (PPV), F1 score, AUC, and Youden index. The utility of the decision models was evaluated by quantifying the net benefit of different threshold probabilities to perform the decision curve analysis (DCA). Calibration curves were plotted to assess the predictive accuracy of the models, and calibration slope/intercept or the Brier score were used for assessment. A calibration slope close to the ideal value of 1, or a Brier score below 0.25, indicated satisfactory calibration performance. After the best model was determined, the analyses of the SHAP summary plot and the SHAP force plot were employed to comprehensively examine the output results of the models, both globally and locally. The significance level was set at P<0.05.

Results

Clinically and demographically pathological characteristics

This study included 19,826 patients in total, and they were classified into a training set (N=13,878) and an internal validation set (N=5,948) at a 7:3 ratio. To ensure the rationality of data distribution, their baseline characteristics were compared. The results indicated that no significant difference was noted in key features between the training set and the validation set (Table 1), which laid a reliable data foundation for subsequent training and validation of the models.

Table 1. Baseline characteristics of participants.

Characteristic	All (N=19,826)	Test (N=5,948)	Train (N=13,878)	P (overall)
Age (years)				0.35
≤60	6,928 (34.9)	2,078 (34.9)	4,850 (34.9)
>70	5,821 (29.4)	1,784 (30.0)	4,037 (29.1)
61–70	7,077 (35.7)	2,086 (35.1)	4,991 (36.0)
Sex				0.42
Female	9,652 (48.7)	2,922 (49.1)	6,730 (48.5)
Male	10,174 (51.3)	3,026 (50.9)	7,148 (51.5)
Race				0.80
Other	4,154 (21.0)	1,239 (20.8)	2,915 (21.0)
White	15,672 (79.0)	4,709 (79.2)	10,963 (79.0)
Marital status				0.6
Married	11,024 (55.6)	3,290 (55.3)	7,734 (55.7)
Other	8,802 (44.4)	2,658 (44.7)	6,144 (44.3)
Site				0.25
Lower	5,076 (25.6)	1,569 (26.4)	3,507 (25.3)
Other	3,587 (18.1)	1,059 (17.8)	2,528 (18.2)
Upper	11,163 (56.3)	3,320 (55.8)	7,843 (56.5)
Histology				0.22
Adenocarcinoma	10,758 (54.3)	3,281 (55.2)	7,477 (53.9)
Other	7,117 (35.9)	2,084 (35.0)	5,033 (36.3)
Squamous cell carcinoma	1,951 (9.84)	583 (9.80)	1,368 (9.86)
Grade				0.27
I–II	2,133 (10.8)	652 (11.0)	1,481 (10.7)
III–IV	5,740 (29.0)	1,675 (28.2)	4,065 (29.3)
Unknown	11,953 (60.3)	3,621 (60.9)	8,332 (60.0)
Laterality				0.30
Other	8,703 (43.9)	2,645 (44.5)	6,058 (43.7)
Right	11,123 (56.1)	3,303 (55.5)	7,820 (56.3)
Surgery				0.054
No	17,488 (88.2)	5,206 (87.5)	12,282 (88.5)
Yes	2,338 (11.8)	742 (12.5)	1,596 (11.5)
Chemotherapy				0.043
No	6,735 (34.0)	2,083 (35.0)	4,652 (33.5)
Yes	13,091 (66.0)	3,865 (65.0)	9,226 (66.5)
Examined				0.27
No	17,579 (88.7)	5,251 (88.3)	12,328 (88.8)
Yes	2,247 (11.3)	697 (11.7)	1,550 (11.2)
Bone				0.14
No	13,169 (66.4)	3,905 (65.7)	9,264 (66.8)
Yes	6,657 (33.6)	2,043 (34.3)	4,614 (33.2)
Liver				0.43
No	16,267 (82.0)	4,860 (81.7)	11,407 (82.2)
Yes	3,559 (18.0)	1,088 (18.3)	2,471 (17.8)
Lung				0.92
No	15,001 (75.7)	4,497 (75.6)	10,504 (75.7)
Yes	4,825 (24.3)	1,451 (24.4)	3,374 (24.3)
Sequence number				0.90
One primary only	16,261 (82.0)	4,882 (82.1)	11,379 (82.0)
Other	3,565 (18.0)	1,066 (17.9)	2,499 (18.0)
First malignant				0.69
No	2,901 (14.6)	880 (14.8)	2,021 (14.6)
Yes	16,925 (85.4)	5,068 (85.2)	11,857 (85.4)
AJCC T				0.96
T1	2,722 (13.7)	816 (13.7)	1,906 (13.7)
T2	5,861 (29.6)	1,749 (29.4)	4,112 (29.6)
T3	4,892 (24.7)	1,462 (24.6)	3,430 (24.7)
T4	6,351 (32.0)	1,921 (32.3)	4,430 (31.9)
AJCC N				0.77
N0	4,642 (23.4)	1,405 (23.6)	3,237 (23.3)
N1	1,856 (9.36)	567 (9.53)	1,289 (9.29)
N2	9,348 (47.2)	2,771 (46.6)	6,577 (47.4)
N3	3,980 (20.1)	1,205 (20.3)	2,775 (20.0)
OS				0.39
Alive	13,116 (66.2)	3,908 (65.7)	9,208 (66.3)
Death	6,710 (33.8)	2,040 (34.3)	4,670 (33.7)
Time (months)	13.0±18.5	12.9±18.4	13.0±18.5	0.83

Data are presented as n (%) or mean ± standard deviation. AJCC, American Joint Committee on Cancer; N, node; OS, overall survival; T, tumor.

Further results of the analyses of the baseline characteristics of patients from the SEER database were as follows: the male proportion was 51.3%, and the age range was mainly between 61 and 70 years old (35.7%). The majority of the population was White (79.0%), and 55.6% of participants were married. Regarding tumor location, the lesions mostly occurred in the right lungs (56.1%) and were more frequently observed in the upper lobe (56.3%). In terms of stage, stage T4 (32.0%) and stage N2 (47.2%) were relatively common. Additionally, 82.0% of the patients have no liver metastasis, 75.7% have no lung metastasis, and 66.4% have no bone metastasis. Regarding pathological type, patients with adenocarcinoma were the most common (54.3%), followed by patients with squamous cell carcinoma (9.84%), and the remaining patients were with other pathological types (35.9%). In terms of degree of differentiation, 10.8% of the patients were at grade I–II, 29.0% were at grade III–IV, and 60.3% were unknown. Regarding therapy, the majority of the patients did not receive surgical treatment (88.2%), and approximately 66.0% of the patients received chemotherapy.

Variable selection for patients with LCBM after radiotherapy

A model for risk forecast on the foundation of LASSO regression was established, with 18 feature parameters included. In LASSO regression, incorporating an L1 regularization term (absolute value penalty term) into the ordinary least squares regression could shrink some coefficients to zero, and the most important features or variables were selected. This helped prevent overfitting, improve the generalization ability of the prediction model, and perform feature selection. Especially in the processing of multiple correlated features, that approach could help enhance the interpretability and performance of the prediction model. Coefficient shrinkage in LASSO regression was achieved through the minimization of the loss function and the L1 regularization term. This encouraged some coefficients to be shrunk to zero, thereby productively eliminating the corresponding features (Figure S2). To establish a prediction model based on ML algorithms for patients with LCBM after radiotherapy, the 13 determined features, comprising age, sex, race, pathological type, surgery, chemotherapy, lymph node biopsy, T stage, N stage, bone metastasis, liver metastasis, lung metastasis, and tumor sequence, were used as the independent variables.

Model comparison

After the selection of the feature variables, seven ML algorithms, comprising LR, NB, SVM, DT, RF, XGBoost, and LightGBM, were used for model construction. Figure 1 shows the receiver operating characteristic (ROC) curves of the seven models in the training set and validation set, respectively. The results demonstrated that the XGBoost model in the training set yielded an AUC of 0.795, which outperformed the other 6 models and became a satisfactory prediction model. The LightGBM model in the validation set yielded an AUC of 0.776, outperforming the other 6 models. The performance rankings of the other 6 models were as follows: the XGBoost model (AUC =0.775), the LR model (AUC =0.774), the NB model (AUC =0.772), the SVM model (AUC =0.765), the RF model (AUC =0.765), and the DT model (AUC =0.732). The detailed performance indicators for the 7 models are shown in Table S2. In the validation set, the LightGBM model exhibited satisfactory predictive performance among the similar models, with a precision rate of 75.15%, a sensitivity rate of 64.9%, a specificity rate of 80.5%, a PPV of 63.47%, an NPV of 81.46%, an F1 score of 64.18%, and a Youden index of 0.454. These results indicated that the LightGBM model was a satisfactory prediction model.

Figure 1.

ROC curve comparison of the training set (A) and the validation set (B) in multiple machine learning algorithms. AUC, area under the curve; KNB, K-nearest neighbor; LightGBM, Light Gradient Boosting Machine; ROC, receiver operating characteristic; SVM, support vector machine; XGBoost, eXtreme Gradient Boosting.

The calibration curve and DCA were used to comprehensively verify the predictive performance of the LightGBM model (Figure 2). Firstly, the calibration curve was used to visualize the calibration accuracy of the predictive performance of the model. The calibration curve of the test group nearly perfectly overlapped with the ideal diagonal line, indicating the highest degree of correspondence between the predictive probability of this model and the actual results. Additionally, the calibration slope was close to the ideal value of 1, and the Brier scores were 0.174 and 0.177 in the training and validation sets, respectively, indicating that the calibration performance of the model was satisfactory. Further, the DCA, a tool of clinical decision support, was employed to quantitatively assess the clinical utility and net benefit of the LightGBM model. Through the DCA, it was found that the LightGBM model achieved a clinical net benefit within the threshold probability range of 0.17 to 0.82.

Figure 2.

Calibration curves and DCA of the LightGBM model. (A) Calibration curve for the training set; (B) DCA for the training set; (C) calibration curve for the test set; (D) DCA for the test set. DCA, decision curve analysis; LightGBM, Light Gradient Boosting Machine.

SHAP-based interpretation of the LightGBM model

In the study of predicting early mortality risk in patients with LCBM after radiotherapy, the contribution of each feature to the prediction model was clarified as a key factor for identifying high-risk populations and formulating intervention strategies. Based on the LightGBM algorithm, SHAP values were calculated to systematically quantify feature importance, and the model was interpreted from both global (describing the overall function of the model) and local (explaining individual predictions) perspectives, in order to enhance clinical utility.

For global interpretation, a violin plot (Figure 3) was generated to illustrate the relationship between feature values and SHAP values. The distribution of feature values was reflected by the horizontal spread, while the correlation between SHAP values and feature values was indicated by vertical position and color (yellow for higher feature values and purple for lower values). Negative SHAP values were predominantly observed in patients who had undergone chemotherapy or surgery, who were younger, who had earlier T or N stages, or who were free of bone, liver, or lung metastases, suggesting that these factors were associated with decreased mortality risk and ameliorated prognosis.

Figure 3.

Summary plots of visualized SHAP values. A point’s placement along the x-axis denotes the actual SHAP value, representing the impact of a specific feature on the model’s output for that particular patient. The yellow features (on the left) indicate features that increase the mortality risk, and the purple features indicate features that decrease the mortality risk. Features are organized along the y-axis based on their importance, determined by the mean of their absolute Shapley values. The higher a feature is positioned in the plot, the more significant its impact on the model. AJCC, American Joint Committee on Cancer; Hist, histology; N, node; SHAP, SHapley Additive exPlanations; Surg, surgery; T, tumor.

In the variable importance plot (Figure 4), features were ranked in descending order of the mean absolute SHAP values. Chemotherapy was identified as the top contributor, indicating its pivotal role in reducing early mortality risk. Age and T stage were ranked second and third, respectively, suggesting that a patient’s general condition and the extent of primary tumor invasion exerted significant effects on prognosis. Bone metastasis and liver metastasis were ranked fourth and fifth, respectively, confirming that distant metastasis was a high-risk factor for an increase in early mortality risk, and the impact of bone metastasis was found to be slightly greater than that of liver metastasis.

Figure 4.

SHAP importance graph for LightGBM. AJCC, American Joint Committee on Cancer; Hist, histology; LightGBM, Light Gradient Boosting Machine; N, node; SHAP, SHapley Additive exPlanations; Surg, surgery; T, tumor.

Local interpretation was performed via SHAP force plots to analyze individual patient predictions. In the LightGBM model, E[f(x)] was defined as the baseline prediction value, and the strength of feature contributions was reflected by the color bars in the force plot, with red indicating negative impacts and yellow denoting positive impacts. Higher SHAP values were indicative of higher early mortality risk. In the example shown in Figure 5, the SHAP value was −0.476, which was higher than the baseline, suggesting that the early mortality risk was relatively high.

Figure 5.

SHAP force plot. The color bar in the force plots reflects the feature contribution intensity: red (left arrow) signifies a negative impact (reduced SHAP value), whereas yellow (right arrow) denotes a positive impact (increased SHAP value). AJCC, American Joint Committee on Cancer; N, node; SHAP, SHapley Additive exPlanations; Surg, surgery.

Discussion

In this study, an interpretable ML-based tool was designed for risk stratification based on the SEER database, in order to forecast the risk of early mortality among patients suffering from LCBM after radiotherapy. A total of 13 influencing factors were included, such as age and pathological type. Ultimately, the LightGBM model was determined to show satisfactory performance in predicting the risk of early mortality among patients suffering from LCBM after radiotherapy (AUC =0.77; F1 score =81%). Its predictive performance provided reliable data support and decision-making basis for the formulation of strategies for individualized diagnoses and treatments of patients with LCBM. In addition, the SHAP was leveraged, and an importance ranking was generated according to the SHAP values. The role of each feature in early death was visually clarified, among which chemotherapy was closely related to early death. The SHAP force plot further verified that chemotherapy played a vital role in alleviating early death in patients.

In this study, 13 independent influencing factors were determined through LASSO regression, including age, sex, race, pathological type, surgery, chemotherapy, lymph node biopsy, T stage, N stage, bone metastasis, liver metastasis, lung metastasis, and tumor sequence. The 5 most influential factors were chemotherapy, age, T stage, bone metastasis, and liver metastasis. Among them, chemotherapy was the most significant influencing factor, aligning with the findings of previous studies that the survival rate of patients with LCBM after radiotherapy was closely related to chemotherapy (22-25). In the C-Brain trial, a median OS of 20.9 months was achieved, and an intracranial disease control rate as high as 98.5% was observed in patients with NSCLC brain metastases who were treated with SRS or WBRT combined with camrelizumab (immunotherapy) and platinum-based doublet chemotherapy (26). Similarly, in patients with small cell lung cancer (SCLC) brain metastases, a significant extension in OS (10 vs. 3.5 months) was demonstrated by Li et al. with WBRT combined with chemotherapy, whereas no survival benefit was observed with chemotherapy alone or WBRT alone (27). In a meta-analysis systematically reviewing and comparing the efficacy of immune checkpoint inhibitors (ICIs), chemotherapy, radiotherapy, and ICIs combined with chemotherapy in NSCLC brain metastases, more considerable survival benefits are conferred by immuno-combination therapies (e.g., ICIs plus chemotherapy) compared with conventional single-agent chemotherapy or radiotherapy alone. These studies indicate that chemotherapy, as a key component of multimodal treatment regimens, may exert synergistic effects when combined with immunotherapy, anti-angiogenic therapy, or targeted therapy. It is well known that the efficacy of single-agent chemotherapy is significantly limited by the blood-brain barrier and the blood-tumor barrier (22,28,29). Differences in barrier-penetrating abilities exist among various chemotherapeutic agents, and although some drugs are allowed by the blood-tumor barrier to enter brain metastatic lesions, the overall intracranial drug concentration is maintained below the therapeutic threshold (30). It is shown by in vitro and clinical studies that platinum-based chemotherapy is effective against primary lung lesions, but its efficacy against brain metastases is limited, partly owing to altered intrinsic drug resistance in brain metastatic cells (31). Furthermore, cisplatin resistance is suggested to enhance the brain metastatic potential of NSCLC cells through molecular mechanisms, indicating that chemotherapy resistance is not merely an efficacy issue but is also associated with the risk of disease progression (32). The efficacy of single-agent chemotherapy may be further attenuated by the immunosuppressive microenvironment of brain metastases (e.g., increased PD-1⁺ T cells and decreased immune cells) (33,34). With the research and development of ICIs and central nervous system-penetrating targeted agents, the treatment paradigm for LCBM has been gradually shifted toward combined regimens with high barrier-penetrating ability. Clinical decision-making should be guided by molecular subtyping, barrier status, drug penetrability, and a patient’s overall condition, in order to avoid isolated reliance on chemotherapy. It was found in this study that patients over 70 years old were more prone to early death. A retrospective study conducted by Shandong University includes 81 patients with LCBM undergoing Gamma knife treatment. The study discovers that the risk of death among patients aged ≥ 70 is significantly higher than that among younger patients [hazard ratio (HR) =2.666, 95% confidence interval (CI): 1.257–5.657, P=0.0106] (35). Similarly, a study in Japan also shows that in NSCLC patients with brain metastasis, the median OS undergoing SRS in patients aged ≥75 years is considerably shorter than that in patients aged 65–74 years (9.0 vs. 13.2 months) (36). This might be related to the tumor’s differences in biological behaviors (such as proliferation rate and invasiveness) or treatment responses in older patients. Previous studies have found that in the central region of the tumor at T3/T4 stage, hypoxic cells have stronger resistance to radiation, and the risk of recurrence in the local tumor is higher after radiotherapy (37,38), which is similar to what has been found in this study: The rate of early death is positively correlated with T stage. In conclusion, those factors affect the early mortality among patients suffering from LCBM, and this study’s findings are expected to provide more valuable references for clinical decision-making.

Different from previous ML studies, this study not only provided a global explanation for the LightGBM model with satisfactory performance, but also explored the model’s prediction for individual selected observations through local explanations, thereby generating different visualized results. We can intuitively estimate the impact of any single determinant on the OS of each patient, and the results may vary. Therefore, this study provided a more comprehensive and direct explanation for the early mortality among patients with LCBM after radiotherapy, both for the entire study population and specific patients.

Nevertheless, several limitations of this study should be acknowledged. First, this study was performed as a retrospective analysis based on the SEER database, and potential selection bias and information bias were inevitably introduced. Data collection in this database was not conducted according to the pre-specified hypotheses of this study, and heterogeneity in clinical treatment regimens, as well as uneven distribution or missingness of some key variables, might have influenced the stability of the statistical results to some extent. Although baseline confounding factors between groups could be partially balanced by propensity score matching and other statistical methods, the bias inherent in retrospective studies could not be fully eliminated. Second, several key clinical indicators for prognosis assessment in LC were not included in the SEER database, which limited the refinement and accuracy of the conclusions of this study to some degree. Specifically, patients’ Karnofsky Performance Status (KPS) and Eastern Cooperative Oncology Group (ECOG) performance status scores were not recorded. There was a lack of information on the number, size, and intracranial tumor burden of brain metastases. Molecular pathological features of driver genes such as EGFR, ALK, and ROS1 were not available. Details of targeted therapy, immunotherapy, and radiotherapy modalities (including SRS and WBRT), as well as radiation doses and fractionation schemes, were not documented. The aforementioned indicators are currently recognized as independent prognostic factors for LC and are closely associated with tumor progression and patient survival. In subsequent studies, these clinical variables could be further integrated to effectively optimize the discrimination and calibration of the prediction model and enhance its overall prognostic performance.

Conclusions

Seven ML models were constructed in this research to forecast the risk of early death among patients suffering from LCBM after radiotherapy. Their predictive performance was satisfactory, and the LightGBM model demonstrated satisfactory predictive accuracy. The interpretability analyses were performed to provide global and local explanations for the models, highlighting the accuracy and intuitiveness of the ML-based prediction models. Despite the fact that clinical diseases are complex and heterogeneous, interpretable ML models may provide valuable and important guidance for clinicians to make informed decisions in advance.

Supplementary

The article’s supplementary files as

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. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.