Data from NSCLC patients with BM who underwent SRS therapy at The Second Affiliated Hospital of Xuzhou Medical University between February 2016 and November 2022 were reviewed. This study was approved by the Institutional Review Board of the Second Affiliated Hospital of Xuzhou Medical University and conducted in accordance with relevant guidelines/regulations and the Declaration of Helsinki. This study is a retrospective study, and the used data collected as part of the participants’ routine care. Written informed consent for participation was waived for this study in accordance with the national legislation and the institutional requirements.

Inclusion criteria: (1) Pathologically confirmed NSCLC; (2) No more than four BM; (3) Exclusive SRS therapy; (4) KPS ≥ 70; (5) Complete pre-therapy and follow-up MRI, including T1 weighted imaging (T1WI), T2 weighted imaging (T2WI), diffusion weighted imaging (DWI), and contrast enhanced T1 weighted imaging (CE-T1WI); (6) Complete records of clinical, pathological, and treatment information. Exclusion criteria: (1) Pathologically confirmed combined with another primary cancer; (2) Prior craniocerebral surgery or WBRT before SRS therapy; (3) Concurrent chemotherapy or immune checkpoint inhibitor therapy within one month before or after SRS; (4) Cystic or cranial metastases.

Patients information included: (1) Clinical information: sex, age, KPS (before SRS therapy), Score Index for Radiosurgery (SIR), neutrophil-to-lymphocyte ratio (NLR) (a key systemic inflammatory marker measured from peripheral blood before surgery), CEA (normal or abnormal), CA125 (normal or abnormal), Cyfra-211 (normal or abnormal), pathological type [squamous cell carcinoma (SCC), adenocarcinoma, adenosquamous carcinoma (AC), large cell carcinoma (LCC), or others], Ki-67 (≤ 30% or > 30%), epidermal growth factor receptor (EGFR) (mutation or non-mutation), anaplastic lymphoma kinase (ALK) (mutation or non-mutation), mediastinal lymph node metastases before surgery (yes or no), postoperative chemotherapy (yes or no), primary tumor control (yes or no), interval between NSCLC confirmation and presence of BM (≤ 3 months or > 3 months), neurological symptoms (yes or no), number of BM (1 or 2-4), extracranial metastasis (yes or no); (2) Radiological information of BM: location, volume of brain metastases (measured on CE-T1WI rather than the total tumor burden), maximum diameter, edema index (EI), T1WI signal intensity, T2WI signal intensity, DWI signal intensity, and enhanced pattern.

For patients with multiple brain metastases, the sum of the volumes of all measurable metastases (up to a maximum of 4, per inclusion criteria) was calculated and used for analysis. Volumetric measurements were performed manually by contouring the border of each metastasis on every slice of the axial CE-T1WI sequences using the PACS (Picture Archiving and Communication System) software, which automatically calculates the volume based on the contoured area and slice thickness.

Patients underwent SRS therapy using the Leksell Gamma Knife® Perfexion™ (Elekta, Norcross, GA, USA). All patients received the prescribed dose in a single fraction.

The prescription dose was individualized based on the tumor volume and location, following our institutional protocol. The median margin dose was 18 Gy (range 16–20 Gy), and the specific distribution was as follows: 92 patients (19.8%) received 16 Gy, 301 patients (64.9%) received 18 Gy, and 71 patients (15.3%) received 20 Gy. Key dosimetric parameters, including the prescription isodose line (median 50%, range 50%-70%) and the maximum dose, were recorded. A preliminary analysis found no significant correlation between these dosimetric parameters and OS or treatment response in our cohort, likely because the dose variation was protocol-driven and relatively narrow. Therefore, to maintain model parsimony and clinical applicability, they were not included as candidate variables in the final predictive models.

Pre-therapy MRI served as the baseline for assessing treatment response, with follow-up MRI performed every 3 months post-SRS. All of the MRI examinations were conducted using a 1.5T MRI system (uMR 588 1.5T; United Imaging Healthcare, Shanghai, China) with an 8-channel phased-array head coil. MRI sequences included T1WI, T2WI, DWI, and CE-T1WI.

Treatment responses were evaluated based on the pre-therapy MRI and the MRI approximately 3 months post-SRS, using the Response Assessment in Neuro-Oncology Brain Metastases (RANO-BM) criteria (12).Patients were categorized into complete response (CR), partial response (PR), stable disease (SD), and progressive disease (PD). OS was defined as the time from BM diagnosis to the last follow-up or to death, with a follow-up deadline of July 30, 2024.

Continuous variables such as age, volume of brain metastases, maximum diameter, and EI were dichotomized to facilitate their integration into the decision-tree model. The optimal cut-off points for dichotomizing continuous variables were determined by maximizing the separation of survival curves using the maximally selected log-rank statistics (via the ‘surv_cutpoint’ function from the ‘survminer’ package in R). This approach was chosen over median split or ROC analysis because it is specifically suited for identifying prognostically relevant thresholds in time-to-event data. We acknowledge the exploratory nature of this method. Therefore, a sensitivity analysis was conducted using the original continuous variables to confirm the robustness of the prognostic factors identified in the Cox model.

Patients were randomly divided into the training and validation sets at a ratio of 7:3. Using the training sample, a decision tree model was developed, by C5.0 algorithm. The depth of the decision tree was limited within 5 layers. The relative importance of the predictive factors was calculated by the C5.0 algorithm using the information gain metric. The analysis stream of decision tree development and validation is available in Figure 1, including processes of data input, data filtering, data type determination, model training, and model testing. In the process of model testing, a scatter plot was generated to evaluate the prediction accuracy of the decision tree. After that, univariable Cox analysis was proceeded for all predictive factors, to filter the potentially important prognosis factors, which were contained in the multivariable Cox model.

Graphical and online nomograms were developed using significant factors from the multivariate analysis. The graphical nomogram was constructed based on the final multivariate Cox model. The points assigned to each variable level in the nomogram were derived by scaling the corresponding regression coefficients (β), typically by the largest coefficient among all predictors (13). These points are then summed to obtain a ‘Total Points’ score, which is projected onto the bottom scales to estimate survival probabilities and median survival time.

The online dynamic nomogram is hosted on a secure server, and we commit to maintaining its availability and functionality for at least two years from the date of publication. The code and underlying data schema will be archived to ensure future reproducibility.

Calibration curve and C-index were used to assess model accuracy and discrimination.

All statistical analyses were performed using R version 4.1.3, with a significance level of 0.05.

The key R packages used for analysis included: rms (version 6.7-0) for nomogram construction and calibration curves, survival (version 3.5-7) for Cox regression, time ROC (version 0.4) for time-dependent ROC analysis, and dplyr (version 1.1.4) for data handling.

A total of 464 patients met the inclusion criteria, with 325 and 139 patients in the training and validation sets, respectively. Baseline characteristics are summarized in Table 1. The median OS for the entire cohort was 15.8 months (95% confidence interval [CI]: 14.6 to 17.0 months). The Log-rank test showed that the optimum cut-off points for age, EI, maximum diameter and volume of brain metastases were 48, 4.11, 3.15 cm and 3.10 cm³, respectively (Supplementary Figure S1). The optimal cut-off for age (48 years) was derived statistically to maximize survival curve separation, which is lower than conventional clinical thresholds. Given that the statistically derived cut-off for age was considerably lower than the median age and conventional clinical thresholds, we performed a sensitivity analysis using a clinically relevant cut-off of 60 years. When the multivariate Cox model was refitted with age dichotomized at 60 years, it remained an independent prognostic factor (HR = 1.82, 95% CI: 1.25-2.65, P = 0.002), and the model’s C-index was similar (0.66). This confirms that older age is a robust predictor of poorer OS, though the magnitude of the effect is more clinically plausible when using a conventional age cut-off.

The therapeutic response to SRS, evaluated at 3 months using RANO-BM criteria, is summarized for both cohorts in Supplementary Table S1. In the entire cohort, the objective response rate (ORR, comprising CR and PR) was 68.1% (316/464), while the disease control rate (DCR, comprising CR, PR, and SD) was 89.4% (415/464). The distribution of responses was well-balanced between the training and validation sets (ORR: 67.7% vs. 69.1%; DCR: 89.2% vs. 90.0%).

The importance of predictive factors for treatment response is shown in Figure 2. Volume of brain metastases, SIR, EI, maximum diameter, and NLR (a systemic immune-inflammatory marker) were identified as the most influential predictors based on their relative importance. The decision tree model for treatment response is shown in Figure 3.

To illustrate the clinical utility and interpretability of the decision tree model, we present two representative patient pathways based on its specific splitting criteria: Case Example 1 (Predicted Favorable Responder): Consider a patient with a brain metastasis volume of 2.5 cm³, classified as SIR grade II, and presenting with mild peritumoral edema (EI = 2.5). This patient would be stratified directly down the leftmost branch of the tree to a terminal node associated with a high probability of CR or PR. This prediction identifies a patient who is an excellent candidate for SRS monotherapy with a high anticipated likelihood of local control. Case Example 2 (Predicted Poor Responder): Now, consider a different patient also with a metastasis volume of 2.8 cm³ and SIR grade II, but exhibiting significant peritumoral edema (EI = 5.5). The tree would then further evaluate the lesion’s maximum diameter. If the diameter measures 3.5 cm, and the pre-treatment NLR is elevated at 3.8, the patient would be guided to a terminal node dominated by SD or PD. This stratification alerts the clinician to a substantially higher risk of suboptimal response to SRS. This information could prompt considerations for more intensive post-treatment monitoring, earlier follow-up imaging, or the discussion of complementary or alternative treatment strategies at the outset.

Figure 4 shows the scatter plot for assessing the consistency between predictive and actual treatment response, for training and validation samples. Preferable consistency between the predictive and actual survival status was presented in the 2 samples. Most of the scatters are located on the diamond line, and the fitted line is closed to the diamond line, demonstrating excellent consistency. The accuracy rates were 81.16% and 80.74% for the training and validation sets, respectively. To further evaluate the model’s performance across all response categories, the macro-averaged F1-score was calculated, yielding values of 0.79 and 0.78 for the training and validation sets, respectively. This indicates that the model maintains a good and consistent balance between precision and recall for each class of therapeutic response.

To evaluate the model’s generalizability, we assessed its discriminative ability (C-index) in key clinical subgroups within the validation set. The model demonstrated consistent performance across most subgroups, including patients with and without EGFR mutations, and with and without extracranial metastases (see Supplementary Table S2).

Univariate analysis identified maximum diameter (HR = 2.331, 95%CI: 1.710-3.178, P < 0.001), EI (HR = 4.681, 95%CI: 3.294-6.653, P < 0.001), age (HR = 1.726, 95%CI: 1.129-2.640, P = 0.012), volume of brain metastases (HR = 6.332, 95%CI: 3.856-10.398, P < 0.001), enhanced pattern (HR = 1.213, 95%CI: 1.047-1.406, P = 0.010) and SIR (grade I vs. grade II: HR = 0.454, 95%CI: 0.288-0.717, P < 0.001; grade I vs. grade III: HR = 0.206, 95%CI: 0.119-0.357, P < 0.001) were identified to be significantly relevant to the OS (P < 0.05).

The results of the univariate and multivariate Cox regression analyses for OS are summarized in Table 2. In univariate analysis, maximum diameter, EI, age, volume of brain metastases, enhanced pattern, and SIR were significantly associated with OS. These significant variables were subsequently incorporated into the multivariate Cox model. After multivariable adjustment, age, volume of brain metastases, EI, and SIR were ultimately identified as independent prognostic factors for OS (Figure 5). Notably, maximum diameter, which was significant in univariate analysis, was not retained in the final multivariable model, likely due to its collinearity with the volume of brain metastasis and its incorporation within the SIR score. A sensitivity analysis was conducted using the original continuous variables for age, EI, and volume in the multivariate Cox model. The results corroborated the primary findings, with all factors retaining their independent prognostic significance (see Supplementary Table S3). The Kaplan-Meier survival curves for these significant factors are shown in Figure 6.

Figure 7 presents the graphical nomogram for predicting 6-, 12-, and 24-month survival and median survival time based on the four independent prognostic factors.

The calibration curves at 6, 12, and 24 months, for training (A-C) and validation (D-F) samples, are shown in Supplementary Figure S2. The calibration curves demonstrated good agreement between predicted and observed survival probabilities, with the fitted lines closely approximating the ideal 45-degree line, indicating satisfactory model calibration. The C-indexes were 0.67 (95%CI: 0.59-0.75) and 0.66 (95%CI: 0.53-0.80) for the training and validation samples, respectively. This indicates a moderate yet consistent discriminative ability of the nomogram in both cohorts. The relatively wide confidence interval in the validation set suggests some instability in model performance on external data, which highlights an area for future refinement and validation in larger cohorts.

To benchmark the performance of our nomogram, we compared its C-index in the validation set against that of the Disease-Specific Graded Prognostic Assessment for lung cancer using molecular markers (Lung-molGPA). The C-index of our nomogram was 0.66 in the validation set, which represents a moderate level of predictive accuracy. Although the C-index of our nomogram (0.66) was comparable to or slightly higher than that of the Lung-molGPA score (0.59) in our cohort, this finding primarily establishes the clinical feasibility and utility of integrating radiological features and systemic inflammation markers for prognosis in this patient population. We acknowledge that there is room for improvement in the model’s discriminative power. The moderate C-index underscores the complex and multifactorial nature of overall survival, which is influenced by unmeasured variables such as subsequent lines of therapy and detailed molecular profiles. Nevertheless, this model provides a solid foundation for future research. We envision that incorporating dynamic changes in inflammatory markers, more comprehensive genomic data, and features of the tumor immune microenvironment could further enhance the model’s performance.

Based on the predictive factors identified by multivariable Cox model and the graphical nomogram, an online dynamic nomogram model for OS was established. The online dynamic nomogram is currently accessible as a demonstration version at: https://helloshinyweb.shinyapps.io/brain_metastasis_from_NSCLC/. Upon acceptance of the manuscript, the final tool will be migrated to a stable institutional server with a long-term commitment to maintenance. A screenshot of the online tool is shown in Figure 8.