Stereotactic Radiation Therapy for Lung Cancer Brain Metastases: Dose Optimization and Prognostic Prediction via Biologically Effective Dose–Based Empirical Dose-Response Modeling and Extreme Gradient Boosting–SHapley Additive Explanations Machine Learning Authors: Fasheng Wu, Juzheng Shi, Weishan He, Yang Tang, Guangmei Deng, Xinlan Lin, Long Kou, Yuansheng Huang, Jia Hu, Hui Zhang, Jinzhou Sheng, Wenting Xi, Yajing Lin, Xunxian Zhu, Xiuping Liang, Ying Zhou, Ying Lu, Fangfang Long, Huayan Jiang, Yuanjun Mo, Shuang Luo, Yingchuan Gao, Yanling Cai, Xuemei Li, Yumei Li, Xiaoping Hu, Xuquan Lu, Yong Huang, Wenchuang Huang, Zhilu Tang, Minyang Tang, Youke Xie, Tongze Cai, Wenya Liu, Huiyi Liu, Peipei Yuan, Ye Chen, Hongbin Liang, Jibing Chen Published: 2026 ABSTRACT AbstractPurposeThis study aimed to integrate BED-based empirical dose-response modeling (logistic regression–derived tumor control probability) with interpretable machine learning (ML) to investigate the complex interactions among biologically effective dose (BED10), tumor volume, neurologic function, and targeted therapy patterns in stereotactic radiosurgery (SRS) for lung cancer brain metastases (BM), thereby optimizing personalized treatment strategies.Methods and MaterialsIn this retrospective study, 449 lung cancer patients with BM treated with CyberKnife between June 2006 and March 2025 were enrolled. Stratified BED-based empirical dose-response models based on maximum tumor diameter (≤2 cm, 2-3 cm, >3 cm) were established to analyze BED10 and local control relationship. The extreme gradient boosting (XGBoost) algorithm was employed to build predictive models for early Karnofsky performance status (KPS) decline and tumor control, incorporating BED10, baseline neurologic status, and targeted therapy patterns. Model interpretability was achieved using SHapley Additive exPlanations (SHAP).ResultsBED-based empirical dose-response modeling revealed significant volume dependence in dose response. Tumors of 2 to 3 cm required the highest dose for 50% control (dose required for 50% tumor control [TCD50]: 62.52 Gy for 1-year; 62.48 Gy for 2-year). The ML model excellently predicted early KPS decline (area under the curve, 0.874). SHAP analysis demonstrated that BED10 effect on neurologic outcomes was modulated by baseline function: lower BED10 (30-60 Gy) increased KPS decline risk in patients with poor baseline neurologic function (grade ≥2), whereas higher BED10 (60-90 Gy) was protective in those with good function (grade <2). Continuous targeted therapy combined with high BED10 showed the most favorable synergistic effect on tumor control.ConclusionsThis study confirms volume-dependent heterogeneity in dose response for BM and reveals a dual, baseline-dependent impact of BED10 on functional outcomes. The integration of ML with BED-based empirical dose-response modeling provides valuable insights for individualized dose prescription and combined modality strategies, supporting continuous targeted therapy with high BED10 as a key approach for optimizing local control. Introduction Brain metastases (BM) represent one of the most prevalent complications in advanced lung cancer, with over 90% of patients who receive diagnoses of primary lung cancer being identified as having BM within 1 year of initial diagnosis.1 Over the past few decades, radiation therapy has become an indispensable component in the comprehensive management of BM. According to the latest guidelines from the National Comprehensive Cancer Network, stereotactic radiosurgery (SRS) serves as the standard treatment modality for intracranial lesions that are inoperable. Its clinical value has been fully validated and is recommended in various clinical practice guidelines.2,3 Among radiation therapy techniques, both SRS and fractionated stereotactic radiosurgery (fSRS) are widely utilized in clinical settings due to their advantages of precise dose delivery and minimal damage to surrounding normal brain tissue. Compared with whole-brain radiation therapy, they pose a lower risk of impairing neurocognitive function.4 Traditional tumor control probability (TCP) models, such as those based on tumor size and BED, provide crucial references for dose prescription.5 However, these models often fail to adequately account for interpatient variability, dynamic clinical factors, and the impact of systemic therapy. With the advancement of precision medicine, targeted therapy has emerged as the cornerstone of systemic treatment for lung cancer patients with BM harboring actionable driver mutations.6 Nevertheless, the interaction between targeted therapy and radiation therapy is complex, and its potential influence on local tumor control has not been fully elucidated by conventional statistical models. Additionally, patients’ baseline functional status (eg, Karnofsky performance status [KPS] score, neurologic function) and its dynamic changes are also key factors affecting treatment efficacy and survival outcomes. However, their contributions in multidimensional predictive models still require further quantification. In recent years, machine learning (ML) algorithms represented by extreme gradient boosting (XGBoost) have provided a new tool for exploring hidden patterns in multidimensional clinical data, thanks to their robust ability to handle nonlinear relationships and complex interaction effects.7 Combining such algorithms with interpretable techniques like SHapley Additive exPlanations (SHAP), it is possible to not only maintain the predictive performance of the model but also reveal the roles of key variables and their interrelationships.8 Therefore, this study aims to integrate traditional BED-based empirical dose-response modeling with ML approaches. On one hand, BED10 dose-response models will be established based on the stratification of maximum tumor diameter of metastatic lesions. On the other hand, the XGBoost algorithm will be employed to investigate the impact of multiple factors, including neurologic function status and targeted therapy patterns, on prognosis. It is anticipated that this integration will provide more precise decision support for personalized comprehensive treatment of lung cancer patients with BM. Methods and Materials Patient characteristics This retrospective cohort study enrolled lung cancer patients with BM who received CyberKnife treatment at Ruikang Hospital Affiliated to Guangxi University of Chinese Medicine between June 2006 and March 2025. The inclusion criteria were as follows: patients with pathologically or radiologically confirmed primary lung cancer (including non-small-cell lung cancer and small-cell lung cancer [SCLC]) with BM who underwent CyberKnife treatment. All cases were required to have pretreatment thin-slice magnetic resonance imaging (MRI) (slice thickness ≤1.5 mm) for accurate measurement of the maximum tumor diameter, complete documentation of radiation therapy parameters (including total dose, number of fractions, and fractional dose, for the purpose of calculating the BED; BED is computed using the linear-quadratic [LQ] model with an α/β value of 10 [BED10]), and at least 6 months of imaging follow-up data to assess local control status. We acknowledge the recognized theoretical limitations of the standard LQ model when applied to the high doses per fraction characteristic of SRS/fSRS, as it may overestimate biological effect. Nonetheless, BED10 (based on the LQ model with α/β = 10) was employed in this study for the following reasons: 1.Firstly, BED10 is the most widely used standardized biological dose metric in current clinical practice and the majority of published literature for comparing different fractionation schemes in stereotactic radiation therapy, facilitating direct comparison of our findings with existing evidence. Firstly, BED10 is the most widely used standardized biological dose metric in current clinical practice and the majority of published literature for comparing different fractionation schemes in stereotactic radiation therapy, facilitating direct comparison of our findings with existing evidence. 2.Secondly, the primary aim of this study was not to compare the validity of different biophysical models but to investigate the complex modulating effects of tumor volume, baseline neurologic function, and targeted therapy patterns on outcomes within a unified and commonly adopted dose framework. Secondly, the primary aim of this study was not to compare the validity of different biophysical models but to investigate the complex modulating effects of tumor volume, baseline neurologic function, and targeted therapy patterns on outcomes within a unified and commonly adopted dose framework. 3.Thirdly, all dose comparisons in this study were performed internally within the BED10 framework, ensuring consistency. BED10 was applied for neurologic toxicity/KPS analysis as KPS decline is a composite endpoint, not isolated normal tissue toxicity. Thirdly, all dose comparisons in this study were performed internally within the BED10 framework, ensuring consistency. BED10 was applied for neurologic toxicity/KPS analysis as KPS decline is a composite endpoint, not isolated normal tissue toxicity. Additionally, clinical variables such as KPS score, primary tumor control status, and use of targeted therapy were required to be recorded. Exclusion criteria included: previous history of brain radiation therapy, combination with other local treatments (eg, surgery, whole-brain radiation therapy, or particle implantation), lack of key imaging or dose parameters, BM of non–lung cancer origin, leptomeningeal metastases, active systemic infection, or expected survival time of less than 3 months. Clinical data were retrospectively extracted from the hospital’s electronic medical record system (HIS). Variables included patient age at the time of CyberKnife treatment, sex, diagnosis, histologic status, gross tumor volume (GTV), planning target volume (PTV), location of BM, primary tumor control status, presence of extracranial metastases, number of BM, KPS score, recursive partitioning analysis, and diagnosis-specific graded prognostic assessment. Age and KPS score were evaluated on the day of CyberKnife treatment. All KPS scores were systematically documented in standardized clinical record fields within the HIS, with clear documentation of the assessing clinician, evaluation date, and key clinical observations (eg, ability to perform daily activities, self-care capacity) supporting the score assignment. KPS assessments were uniformly performed by attending radiation oncologists in strict adherence to the official KPS assessment guidelines—this protocol quantifies functional capacity across 3 core domains: daily activity independence, self-care ability, and disease-related functional limitation. To ensure consistent scoring across the 2006 to 2025 study period, regular intradepartmental case reviews and scoring calibration sessions were conducted, where KPS assignments for representative patient cases were cross-validated among attending radiation oncologists to resolve any discrepancies and maintain interrater and temporal uniformity. The status of intracranial and extracranial diseases as well as the control status of the primary cancer were determined based on the last medical assessment prior to CyberKnife treatment. Prognosis-related variables included local progression-free survival (LPFS), overall survival (OS), and the response of BM on the first follow-up MRI (complete response [CR]/partial response [PR]/stable disease [SD]/progressive disease [PD]). The procedures followed for the purposes of this study were in accordance with the ethical standards of the responsible committee on human experimentation (institutional or regional) or with the Helsinki Declaration (1964, amended in 1975, 1983, 1989, 1996 and 2000) of the World Medical Association. This study was approved by the ethics committee of Ruikang Hospital Affiliated to Guangxi University of Chinese Medicine (ethics approval No.: KY2025-144), and data collection was censored as of March 31, 2025. A total of 449 patients who met the criteria were finally included in the analysis. Radiosurgery technique All radiosurgery treatments were delivered using the CyberKnife SRS system (M6 unit, Accuray Incorporated). Patients were placed in the supine position, and their heads were immobilized using a thermoplastic mask. MRI scans were performed with a thin-slice thickness of ≤1.5 mm. The images were transferred to the radiation therapy planning system to construct 3-dimensional volumetric images. The GTV was defined as the enhancing lesion on MRI by neuro-oncologists or radiation oncologists. GTV and organs at risk (OARs) were delineated based on MRI images and approved by both neuro-oncologists and radiation oncologists. The GTV was expanded to form the PTV with a margin of 0 to 2 mm. Typically, a 0 mm margin was used for metastases with well-defined, noninfiltrative borders, whereas a 1 to 2 mm margin was applied for lesions with irregular or ill-defined borders. OARs delineated for all brain cases included the eyes, optic nerves, optic chiasm, visual pathways (optic nerves and optic chiasm), brainstem, and spinal cord. Planning risk volumes for the spinal cord, brainstem, and visual pathways were created by adding a 2 mm margin to account for uncertainties associated with radiation delivery to the target volume. All OAR doses were constrained within clinical safe limits for central nervous system tissues during plan verification. Treatment was delivered using the robotic CyberKnife system (Accuray Inc), which employs a compact 6 MV linear accelerator mounted on a robotic arm. This enables a nonisocentric, noncoplanar delivery technique, where the patient is irradiated by 100 to 200 individually targeted x-ray beams from a wide range of solid angles, optimizing dose conformality and sparing of adjacent OARs.9 With a plan maximal dose of 100%, doses were prescribed to the 55% to 90% isodose lines. The PTV was covered by at least 95% of the dose in all cases. Before the official treatment, technicians verified each treatment plan by checking target coverage, conformity index, and dose gradient index to evaluate the treatment plan. Follow-up Patients were followed up regularly after radiation therapy (every 3 months within the first year and every 6 months after the first year) to assess tumor control status. Follow-up was conducted using clinical evaluations and MRI and/or CT scans. The response of intracranial lesions to treatment was assessed as CR, PR, SD, or PD. CR was defined as the complete disappearance of all irradiated BM on contrast-enhanced MRI with no new lesions. PR was defined as at least a 30% reduction in the sum of the longest diameters of intracranial lesions, sustained for ≥4 weeks. PD was defined as at least a 20% increase in the sum of the longest diameters of intracranial lesions, with an absolute increase of 5 mm or more. SD was defined as not meeting the criteria for either PD or PR.10 At each follow-up visit, functional status and neurotoxicity were scored. The KPS score was used to evaluate the patient’s overall functional status. Neurologic function was assessed and graded (grade, 0-4) according to the National Cancer Institute Common Terminology Criteria for Adverse Events version 5.0,11 as detailed in Table 1. Intracranial lesion–related toxicities were graded in accordance with the National Cancer Institute Common Terminology Criteria for Adverse Events version 5.0. For all deceased patients, the date of death and cause of death (neurologic or nonneurologic) were recorded. Patients with stable systemic disease and progressive neurologic dysfunction were considered to have died from neurologic causes. Patients with improved or stable neurologic function who died from fatal infections, bleeding, or failure of vital organ systems outside the brain were considered to have died from systemic cancer (nonneurologic causes). For surviving patients, survival time was censored at the date of the last follow-up. Two clinical oncologists evaluated the patients and summarized the follow-up results. Table 1Neurologic functionTable 1 dummy alt textNo symptoms(Grade 0)Minor symptoms, fully active without assistance(Grade 1)Moderate symptoms; fully active but requires assistance(Grade 2)Moderate symptoms; less than fully active, requires assistance(Grade 3)Severe symptoms; totally inactive(Grade 4) Neurologic function Statistical analysis and modeling Survival analysis was performed using the Kaplan-Meier method to calculate OS rates and LPFS rates. For half-year local control binary outcomes, patients who died/were lost to follow-up before assessment were coded as 0 (control/censored), with preassessment events treated as noninformative for binary labeling. The 1-year and 2-year local tumor control was analyzed as a binary endpoint. Patients who were alive, had available follow-up imaging at or beyond the respective time point, and had unequivocal radiographic evidence (per RANO-BM criteria)10 were classified as having either controlled or uncontrolled disease. Patients who died or were lost to follow-up before reaching the 1-year or 2-year assessment time point were excluded from the corresponding binary outcome analysis for that specific time point, as their true tumor status at that future time was unknown. This approach assumes that the reason for censoring (death or loss to follow-up) is independent of the likelihood of local failure, which is a conservative simplification to avoid misclassification bias. With a median follow-up of 15.2 months, the 1-year local control analysis included 446 patients and the 2-year local control analysis included 449 patients, with the above censoring criteria consistently applied to both time-point analyses. BED-based empirical dose-response modeling was stratified based on the maximum tumor diameter (≤2 cm, 2-3 cm, >3 cm). Logistic regression models were used to fit the relationship between BED10 and 1-year and 2-year tumor control, and dose required for 50% tumor control (TCD50) was calculated to compare differences between groups. ML and interpretability analysis To explore complex predictors and interaction effects that are difficult to capture by traditional statistical models, this study adopted a machine learning–based modeling strategy to improve the reliability and clinical interpretability of prediction results. Model construction and data set partitioning Three independent interpretable ML models were constructed for the 3 core clinical endpoints to verify different research hypotheses as described below. TCP model The TCP model focused on predicting 1-year and 2-year local control rates. By integrating multidimensional clinical features, this model aimed to verify and expand the conclusions of traditional dose-response models. Only baseline KPS was used as the KPS-related predictor, and no follow-up KPS data were included. Functional preservation model This model took the decline in KPS score within 3 months after radiation therapy as the prediction target. “KPS decline” was defined as a difference of ≥10 points between the baseline KPS score and the KPS score 3 months after radiation therapy (binary outcome: 1 = early KPS decline occurred, 0 = no early KPS decline occurred). For this endpoint, patients with missing baseline/3-month KPS data were excluded upfront, and those who died before the 3-month assessment were coded as 0 (no decline), with preassessment death treated as noninformative for binary labeling. Meanwhile, baseline neurologic function grade (≥2 points defined as “poor baseline neurological function,” <2 points defined as “good baseline neurological function”) was included as a key stratification variable. The goal was to identify factors related to functional safety, such as BED10, baseline neurologic function status, and their interaction effects. Only baseline KPS and baseline neurologic function grade were used as predictors. The 3-month KPS was only used to define the outcome and was not included as an input feature. Comprehensive prognostic model This model targeted the prediction of 1-year local control rates. It incorporated treatment parameters, tumor characteristics, patient baseline status, and systemic targeted therapy patterns to achieve a comprehensive prognostic assessment. In addition to all baseline features, including baseline KPS, 3-month and 6-month KPS were added as early dynamic temporal features, which were all available before the 1-year endpoint. Six-month tumor control status (assessed per RANO-BM criteria) was also included as a midtreatment dynamic feature, measured strictly prior to the 1-year local control endpoint to avoid information leakage. All models were constructed using the XGBoost algorithm, which has significant advantages in handling mixed-type features, accommodating missing values, and capturing nonlinear relationships and high-order interaction effects between variables. Thus, it is highly suitable for analyzing complex clinical data. The entire data set was randomly divided into a training set and an independent test set at an 8:2 ratio. Stratified sampling was performed based on the target variable during the partitioning process to ensure consistent distribution of outcome events (eg, local failure, KPS decline) in the 2 data sets, thereby reducing selection bias. Hyperparameter optimization of the models was conducted on the training set using 5-fold cross-validation combined with grid search to balance model performance and interpretability. The finally determined optimized parameters included n_estimators = 100, max_depth = 1, and learning_rate = 0.1. For data sets with class imbalance (eg, a small number of cases with KPS decline), the scale_pos_weight parameter was adjusted according to the inverse ratio of event samples to nonevent samples in the training set to reduce the prediction bias of the model toward the majority class samples. Antioverfitting strategy Given the moderate sample size (n = 449), a multifaceted strategy was employed to ensure model generalizability and mitigate overfitting. First, model complexity was strictly constrained by using decision stumps (max_depth = 1). Second, randomness was introduced via row (subsample = 0.8) and feature (colsample_bytree = 0.8) subsampling. Third, an independent test set (20%) was created via stratified splitting and was used exclusively for final performance evaluation. Fourth, a conservative set of hyperparameters was chosen (n_estimators = 100, learning_rate = 0.1). Finally, the stability of all performance metrics (eg, area under the curve [AUC]) was assessed by calculating their 95% CIs through 1000 bootstrap iterations on the test set. Feature engineering and preprocessing Handling of dosimetric data for patients with multiple metastases The BED10 was calculated for each individual treated lesion. For patient-level analyses (eg, prediction of early KPS decline or OS), a composite patient-level BED10 was defined. This was calculated as the maximum BED10 value among all irradiated lesions for a given patient, representing the peak dose exposure relevant to radiation-induced neurologic toxicity risk. All ML models used this patient-level BED10. In contrast, analyses of local tumor control were performed at the individual lesion level, pairing each lesion’s own BED10 with its outcome. For integration of lesion-level (tumor diameter, GTV, BED10, local control, LPFS) and patient-level endpoints (KPS, neurologic function): lesion-specific LPFS was defined as time from CyberKnife treatment to local progression, with local control outcomes determined per RANO-BM criteria; a composite patient-level BED10 (maximum BED10 among all lesions) was used for patient-centric outcomes (eg, KPS decline, OS), as peak dose correlates with radiation-induced neurologic toxicity; a patient was classified as “locally controlled” only if all lesions achieved CR/PR/SD; lesion features (maximum diameter, lesion count, GTV) and patient characteristics (baseline KPS, neurologic function, targeted therapy) were jointly incorporated into XGBoost models, with interaction terms (eg, BED10 × baseline neurologic function) quantifying modulatory effects. Input features of the models were systematically selected from the perspective of clinical relevance, covering 4 core dimensions, and additional interaction terms were constructed to verify prior hypotheses as detailed below. Treatment parameters The treatment parameters were BED10, total prescription dose, fractionation mode (SRS vs fSRS), and prescription isodose line. Tumor characteristics The tumor characteristics were maximum diameter of metastatic lesions, GTV, anatomic location, number of BM, pathological subtype. Patient baseline characteristics Patient baseline characteristics were age at radiation therapy, baseline KPS score, baseline neurologic function grade, diagnosis-specific graded prognostic assessment. Systemic therapy Targeted therapy patterns were categorized as “no targeted therapy,” “adjuvant targeted therapy” (initiated after radiation therapy), and “continuous targeted therapy” (initiated before radiation therapy and maintained after radiation therapy). Interaction terms To explore potential effect modification, 3 key interaction features were constructed: dose-tumor volume interaction (BED10 × maximum diameter group of metastatic lesions), BED10 × poor baseline neurologic function, and BED10 × targeted therapy. Different models used different KPS time points: the TCP model and functional preservation model only used baseline KPS, while the comprehensive prognostic model additionally included 3-month and 6-month KPS as early dynamic features. All feature engineering was performed after data splitting to ensure temporal validity. To ensure model stability, standardized preprocessing was performed on the features. Categorical variables were processed using one-hot encoding to avoid introducing ordinal assumptions. Continuous variables were standardized to a mean of 0 and a standard deviation of 1 using the StandardScaler tool from the scikit-learn library, preventing features with large scales from dominating the model learning process. Model validation and performance evaluation To ensure the generalization ability of the models, the performance of the models was strictly evaluated only on the independent test set, and appropriate evaluation metrics were selected based on the nature of different endpoints as detailed below. For binary classification tasks, core evaluation metrics included accuracy, precision, recall, and F1-score. Meanwhile, confusion matrices were used to visualize the number of true positives, false positives, true negatives, and false negatives. The area under the receiver operating characteristic curve (AUC-ROC) was used to quantify the discriminative ability of the models. To assess the stability of this metric, 1000 bootstrap resamplings were performed on the test set, and the 95% CI of the AUC was calculated to correct for the impact of sampling errors. In addition, to verify the consistency between the predicted probabilities of the models and the actual event rates, calibration curves were plotted for each model, and the overall calibration error was quantified using the Brier score. This ensured that the predicted results of the models not only had good discriminative ability but also possessed quantitative reliability. Model interpretability analysis To address the “black box” issue of ML models and explore clinically applicable conclusions, the SHAP framework was used for posthoc interpretability analysis of the models as detailed below. Global feature importance The mean absolute SHAP value (mean [|SHAP value|]) of each feature across all samples in the test set was calculated. This metric reflects the average contribution of a feature to the model’s prediction results, based on which features can be ranked according to their overall impact. SHAP dependence plots These plots visualized the relationship between continuous features (eg, BED10) and their corresponding SHAP values. Samples were colored according to the categorical variables of interest to identify nonlinear trends and effect modification. Partial dependence plots Partial dependence plots were drawn for features with top importance rankings as a supplement to the SHAP dependence plots. These plots could demonstrate the marginal effect of a single feature on the predicted probability of the endpoint event while controlling for the influence of other features. All statistical analyses, model training, and interpretability evaluations were implemented in the Python 3.9 environment using the following libraries: scikit-learn (for data preprocessing and calculation of performance metrics), XGBoost (for model construction), SHAP (for interpretability analysis), and statsmodels (for supplementary traditional statistical verification). Results Patient and treatment characteristics Between 2006 and 2024, a total of 449 lung cancer patients with BM who received CyberKnife treatment were enrolled. The baseline demographic and clinical characteristics of the patients are detailed in Table 2. Histologically, non-SCLC was the predominant subtype (85.5%, 384/449), including adenocarcinoma (77.7%, 349/449) and squamous cell carcinoma (7.7%, 35/449). This was followed by SCLC (8.9%, 40/449) and other histologic types (5.5%, 25/449). As indicated in Table 1, the number of BM lesions per patient ranged from 1 to 10, with a total of 1126 BM lesions analyzed. Table 2Baseline demographic and clinical characteristics of the study populationTable 2 dummy alt textCharacteristicsValueMedian age in years at CK, (IQR)59 (52-66)Sex, n (%) Men294 (65.4%) Women155 (34.5%)KPS, n (%) 90-10038 (8.4%) 70-80402 (89.5%) <709 (2%)Targeted therapy, n (%) Definitive targeted therapy100 (22.2%) Adjuvant targeted therapy148 (33%) No targeted therapy201 (44.7%)Histology, n (%) Non-small-cell lung cancer384 (85.5%) Adenocarcinoma349 (77.8%) Squamous cell carcinoma35 (7.7%) Small-cell lung cancer40 (8.9%) Others25 (5.5%)Number of brain metastases, n (%) 198 (21.8%) 2-4150 (33.4%) 5-1083 (18.5%) >10118 (26.2%)Primary tumor status, n (%) Stable222 (49.4%) Active182 (40.5%) Unknown45 (10.0%)Extracranial metastases, n (%) Yes323 (71.9%) No126 (28%)Control of extracranial diseases, n (%) Stable177 (39.4%) Active171 (38.0%) Unknown27 (6.0%)RPA class, n (%) I86 (19.1%) II217 (48.3%) III146 (32.5%)DS-GPA, n (%) 0-129 (6.5%) 1.5-2236 (52.6%) 2.5-3148 (33%) 3.5-436 (8.0%)Neurologic function classification before CK, n (%) Grade 071 (15.8%) Grade 1201 (44.8%) Grade 2151 (33.6%) Grade 326 (5.8%) Grade 40 (0%)Median maximum diameter of brain metastases, cm (IQR)2.3 (1.5-3.4)Abbreviations: CK = CyberKnife; DS-GPA = diagnosis-specific graded prognostic assessment; IQR = interquartile range; KPS = Karnofsky performance status; RPA = recursive partitioning analysis.Data are presented as median (IQR) for continuous variables and n (%) for categorical variables. Baseline demographic and clinical characteristics of the study population Abbreviations: CK = CyberKnife; DS-GPA = diagnosis-specific graded prognostic assessment; IQR = interquartile range; KPS = Karnofsky performance status; RPA = recursive partitioning analysis. Data are presented as median (IQR) for continuous variables and n (%) for categorical variables. The total prescription dose across all patients ranged from 12 to 40 Gy. For single-fraction SRS, the prescription dose ranged from 12 to 24 Gy (mean, 17.2 Gy; median, 18 Gy). For fSRS, the total prescription dose ranged from 20 to 40 Gy delivered in 2 to 5 fractions (mean, 29.8 Gy; median, 30 Gy). The calculated BED10 varied between 26.4 Gy and 93.3 Gy. Detailed data on KPS scores and neurologic function grades at baseline and key follow-up time points, including the number of evaluable patients at each time point, are available in Table E1. Survival and follow-up The median follow-up duration was 15.2 months (range, 1.6-133.9 months), and the median follow-up duration for surviving patients was 20.9 months (range, 2.9-133.9 months). Kaplan-Meier analysis revealed that the median OS of all patients from the initiation of radiation therapy was 28.9 months (95% CI, 20.9-36.9), with 1-year and 2-year OS rates of 72% and 53.8%, respectively. The median LPFS was 8.8 months (95% CI, 7.9-9.7), with 1-year and 2-year LPFS rates of 40.1% and 17.8%, respectively. Dose-response relationship and volume dependence of TCP BED-based empirical dose-response models for 1-year and 2-year follow-ups were constructed based on stratification by the maximum tumor diameter. In the 1-year BED-based empirical dose-response model, significant differences were observed in the dose-response curves among tumors of different diameters as detailed below. For lesions with a maximum diameter ≤2 cm, TCP increased with the elevation of BED10, indicating that a higher BED10 could improve the 1-year tumor control rate for small tumors; for lesions measuring 2 to 3 cm in diameter, TCP remained high within the range of moderate BED10 but declined with further increases in dose; for lesions >3 cm in diameter, TCP decreased as BED10 increased. The corresponding TCD50 were 57.95 Gy (95% CI, 49.62-66.27) in the ≤2 cm group, 62.52 Gy (95% CI, 58.61-66.43) in the 2 to 3 cm group, and 59.01 Gy (95% CI, 54.12-63.91) in the >3 cm group. Statistically significant differences were noted among the groups (P = .0161) (Fig. 1A). (A) Dose-response curves for 1-year, 2-year tumor control probability (TCP) stratified by maximum tumor diameter. (traditional TCP model, logistic regression). (B) The solid lines represent the fitted logistic curves for each size group (≤2 cm, 2-3 cm, >3 cm), and the individual points show the observed data (traditional TCP model, logistic regression). A similar trend was observed in the 2-year BED-based empirical dose-response model: for the groups with maximum tumor diameters of 2 to 3 cm and >3 cm, TCP decreased with increasing BED10. The decrease was more pronounced in the >3 cm group, suggesting that large tumors exhibit poorer 2-year tumor control outcomes under high BED10. The TCD50 values for the ≤2 cm, 2 to 3 cm, and >3 cm groups were 58.55 Gy (95% CI, 50.98-66.12), 62.48 Gy (95% CI, 58.57-66.39), and 60.11 Gy (95% CI, 56.31-63.92), respectively, with significant intergroup differences (P = .0239) (Fig. 1B). Tumor maximum diameter does not serve as a superior predictor of local control compared with BED10; instead, it modulates the BED10-LC relationship, and the interaction between these 2 factors is a key determinant of treatment outcomes for BM of different volumes. Impact of BED10 on neurologic outcomes and key predictive factors The predictive model demonstrated excellent discriminative performance in the test set, with an AUC-ROC of 0.874 (95% CI, 0.766-0.968) for predicting early KPS decline (Fig. 2A). In the subgroup of patients with poor baseline neurologic function (n = 177), 90 (50.8%) received low BED10 (30 to <60 Gy) and 85 (48.0%) received high BED10 (60-90 Gy) radiation therapy. Based on this robust model, SHAP analysis was employed to interpret the prediction results and clarify the influence of key variables. (A) Receiver operating characteristic (ROC) curve for the predictive model for predicting early Karnofsky performance status decline (≥10-point decrease within 3 months) in lung cancer brain metastases treated with CyberKnife stereotactic radiation therapy. The area under the curve (AUC) = 0.874 (95% CI, 0.766-0.968). The dashed diagonal line represents a random prediction (AUC = 0.5) (extreme gradient boosting [XGBoost]: functional preservation model, early KPS decline). (B) ROC curve for the predictive model for predicting 1-year local tumor control in lung cancer brain metastases treated with CyberKnife stereotactic radiation therapy. AUC = 0.853 (95% CI, 0.769-0.926). The dashed diagonal line represents a random prediction (AUC = 0.5). (XGBoost: tumor control probability model, 1-year local control). SHAP dependence plots analyzing the association between BED10 and neurologic function changes revealed the following: in patients with poor baseline neurologic function (grade ≥2), positive SHAP values were observed within the low BED10 range (approximately 30-60 Gy), suggesting that this dose range may be associated with an increased risk of KPS decline; in patients with good baseline neurologic function (grade <2), negative SHAP values corresponded to high BED10 (approximately 60-90 Gy), indicating a potential protective effect on neurologic function (Fig. 3). Baseline neurologic grade served as a surrogate for normal brain tissue radiation tolerance. SHapley Additive exPlanations (SHAP) dependence plot for BED10, colored by baseline neurologic function. Each point represents a single patient. The position on the y-axis shows the impact of BED10 on the model output (SHAP value) for predicting early Karnofsky performance status decline (extreme gradient boosting [XGBoost]: functional preservation model, early Karnofsky performance status decline). SHAP feature importance ranking showed that BED10 was the most critical feature, exerting the greatest impact on model output. This was followed by the interaction between BED10 and poor baseline neurologic function. Total dose, KPS score, and the interaction between BED10 and KPS were also identified as important predictive factors (Fig. 4A). These findings confirm that BED10 is one of the core factors influencing model predictions of treatment outcomes in lung cancer patients with BM. (A) SHapley Additive exPlanations (SHAP) summary plot of global feature importance for predicting early Karnofsky performance status (KPS) decline. SHAP feature importance for the functional preservation model. All features are baseline variables, and KPS refers to baseline KPS (extreme gradient boosting [XGBoost]: functional preservation model, early KPS decline). (B) SHAP summary plot of global feature importance for predicting the interaction effect between targeted therapy modality and BED10 on prognosis. Features are ordered by their mean absolute SHAP value, which represents the average impact on the model’s prediction output. Key predictors include tumor control status at 6 months, neurologic function scores (Karnofsky performance status–neurological function score [KPS-NFS]) at follow-up, RECIST assessment categories, age, and BED10). Abbreviations: BED10 = biologically effective dose (α/β = 10); DS-GPA = diagnosis-specific graded prognostic assessment; GTV = gross tumor volume; Mo = month; RECIST = Response Evaluation Criteria in Solid Tumors; SRS = stereotactic radiosurgery. Six-month tumor control status = midtreatment predictive feature assessed prior to the 1-year endpoint (extreme gradient boosting [XGBoost]: comprehensive prognostic model, 1-year local control). Interaction between targeted therapy patterns and BED10 on prognosis SHAP feature importance ranking indicated that although BED10 was not the most dominant feature, it still occupied a significant position, highlighting its role in model prediction (Fig. 4B). SHAP dependence plots illustrating the contribution of BED10 to model output, stratified by targeted therapy patterns, exhibited distinct patterns as detailed below. For patients without targeted therapy At relatively low BED10 values, SHAP values were negative or close to 0, indicating that low-dose radiation therapy (without adjuvant targeted therapy) exerted a negative contribution to model output (eg, treatment efficacy) and may fail to control tumors effectively. As BED10 increased, SHAP values became positive, meaning high-dose radiation therapy could make a positive contribution to model output and help improve treatment efficacy in the absence of targeted therapy. For patients receiving adjuvant targeted therapy (initiated after radiation therapy) SHAP values remained relatively stable across the observed BED10 range, with no extreme positive or negative deviations. This suggests that in this treatment pattern, changes in radiation therapy dose have a mild impact on treatment efficacy, and targeted therapy may play a favorable supplementary or synergistic role after radiation therapy. For patients receiving continuous targeted therapy (initiated before radiation therapy and maintained afterward): positive SHAP values were predominant at high BED10 values, indicating a synergistic effect between high-dose radiation therapy and continuous targeted therapy (administered both before and after radiation therapy). This combination could significantly exert a positive contribution to model output (eg, treatment efficacy) and is more conducive to tumor control (Fig. 5). SHapley Additive exPlanations (SHAP) dependence plot for BED10, colored by targeted therapy modality. Each point represents a single patient. The position on the y-axis shows the impact of BED10 on the model output for that patient (SHAP value), and the color indicates the modality of targeted therapy received (extreme gradient boosting [XGBoost]: tumor control probability model, 1-year local control). Abbreviation: BED10 = biologically effective dose. The ROC curve analysis further supported the above conclusions: the predictive model exhibited good performance in distinguishing different treatment outcomes, with an AUC of 0.8534 (95% CI, 0.7689-0.9256) (Fig. 2B), confirming the model’s high discriminative ability. These results are consistent with the contributions of BED10 and different targeted therapy patterns to prediction outcomes observed in SHAP analysis, demonstrating that the model can effectively integrate multifactor information to achieve reliable prediction of treatment efficacy. The predictive performance of the 2 XGBoost models was assessed using accuracy, precision, recall, and F1-score. For the model predicting early KPS decline (≥10-point decrease within 3 months postradiation therapy), the accuracy was 0.9000, precision 0.9403, recall 0.9265, and F1-score 0.9333. For the model predicting 1-year local tumor control of intracranial lesions, the accuracy was 0.7889, precision 0.9394, recall 0.6458, and F1-score 0.7654. These results demonstrate the favorable discriminative ability of the established models. To clarify the analytical independence of the models and their correspondence with the associated figures, key details are supplemented as follows: the 3 XGBoost models (TCP model, functional preservation model for early KPS decline, and comprehensive prognostic model) were analyzed, trained, and interpreted entirely separately for their distinct clinical endpoints, with independent feature sets and separate SHAP analyses performed on each model’s respective test set (no cross-model feature combination or integration). The precise mapping between all models and figures in this study is confirmed as: (1) Fig. 1 corresponds to the traditional TCP model constructed with logistic regression (an independent statistical analysis, unrelated to the 3 XGBoost models); (2) Fig. 2A and Fig. 3 correspond to the XGBoost functional preservation model (early KPS decline); (3) Fig. 2B and Fig. 5 correspond to the XGBoost TCP model (1-year local control); (4) Fig. 4A corresponds to the XGBoost functional preservation model (early KPS decline), and Fig. 4B corresponds to the XGBoost comprehensive prognostic model (1-year local control). Discussion BM is a key driver of poor prognosis in patients with advanced lung cancer.12 The treatment of BM has long been challenged by the clinical trade-off between “local tumor control” and “neurological function preservation,” and balancing efficacy and safety remains a critical challenge in clinical practice. Based on data from 449 lung cancer patients with BM who received CyberKnife treatment, this study integrated traditional BED-based empirical dose-response modeling with XGBoost-SHAP ML technology to analyze 3 core clinical issues: “tumor volume-dose-response relationship,” “interaction mechanism between BED10 and neurological function,” and “synergistic mode of targeted therapy and radiation therapy.” Traditional BED-based empirical dose-response models for BM mostly rely on a simple correlation between BED10 and GTV, while ignoring the regulatory role of tumor diameter in dose response. Through stratified BED-based empirical dose-response modeling, this study revealed the volume heterogeneity in the relationship between BED10 and TCP, providing a quantitative basis for dose prescription. From the results of TCP curves and the TCD50, the 1-year and 2-year TCD50 values in the 2 to 3 cm diameter group were higher than those in the ≤2 cm and >3 cm groups. For BM with a diameter of ≤2 cm, the small volume implies a relatively small total number of tumor cells and potentially good radiosensitivity; hence, a lower BED10 can achieve satisfactory tumor control. This is consistent with the findings of Redmond et al,13 who concluded that small-volume BM exhibit a more rapid response to radiation therapy and require a lower effective dose. In contrast, for BM with a diameter of >3 cm, despite the large volume, more necrotic and hypoxic regions are prone to form inside the tumor. Tumor cells in these regions are insensitive to radiation therapy, and large-volume tumors have higher requirements for the spatial distribution of radiation therapy dose. Simply increasing BED10 not only fails to effectively improve TCP but may also cause complications due to excessive radiation exposure to normal tissues, thereby adversely affecting tumor control outcomes.14 Similarly, Baliga et al15 also proposed a similar view that the dose-effect relationship of radiation therapy for large-volume BM is more complex, and simply increasing the dose has limited improvement on TCP. The higher TCD50 in the 2 to 3 cm diameter group suggests that BM within this volume range have unique biological behaviors. At this stage, the tumor not only has a certain degree of cell proliferation activity but also has not formed a large number of necrotic and hypoxic regions that affect radiation therapy efficacy. Increasing BED10 can more effectively kill tumor cells, thereby improving TCP. This result supplements the detailed understanding of the relationship between BM volume and radiation therapy dose requirements. Previous studies mostly focused on the rough classification of “small volume” and “large volume,” whereas this study clarifies the particularity of dose requirements for the intermediate volume range of 2 to 3 cm. In addition, the 1-year and 2-year TCP results of this study showed good consistency, indicating that the BED10-TCP relationship in each diameter stratum is stable across different follow-up periods. This further verifies the reliability of formulating radiation therapy dose strategies based on this stratification. In summary, this study clarifies the differences in BED10 requirements for BM of different diameters, providing key references for clinically formulating individualized radiation therapy dose regimens based on BM diameter. BM with a diameter of 2 to 3 cm may require a higher BED10 to optimize TCP. For lesions larger than 3 cm, the observed decrease in local control with increasing BED10 reflects the synergistic effect of tumor diameter and dose: intratumoral hypoxia/necrosis reduces radiosensitivity to high BED10, and excessive dose further impairs local control by increasing normal tissue toxicity, rather than indicating that tumor diameter is a better standalone predictor than BED10. Neurologic function preservation is one of the core goals in the treatment of BM. However, traditional studies mostly analyze the impact of BED10 or KPS score in isolation, ignoring their interaction. With the help of ML and SHAP analysis, this study deeply explores the impact of BED10 and its interaction with other clinical factors on the decline of KPS score in lung cancer patients with BM undergoing radiation therapy. From the SHAP dependence plot of BED10, it can be seen that there is a significant interaction between baseline neurologic function status and BED10 on KPS decline. For patients with poor baseline neurologic function (grade ≥2), a lower range of BED10 (30-60 Gy) is more likely to lead to KPS decline. This phenomenon may be due to the lower tolerance of damaged nerve tissue to radiation therapy—patients with poor baseline neurologic function often have pre-existing neuroinflammation or blood-brain barrier damage, which can amplify the effect of radiation-induced injury.16,17 For patients with poor baseline neurologic function (grade ≥2), our SHAP analysis indicated that BED10 in the range of 30 to 60 Gy10 was associated with an increased risk of early KPS decline. However, it is important to note that very few patients in this subgroup received high BED10 (>60 Gy10), as clinicians tend to avoid dose escalation in this vulnerable population. Consequently, the model lacked sufficient data to reliably estimate the effect of doses exceeding 60 Gy10. Therefore, although we observed an elevated risk in the 30 to 60 Gy10 range, the safety profile of higher doses remains uncertain. Our recommendation to limit BED10 to <60 Gy10 in these patients is based on a conservative approach to avoid potentially unknown risks associated with higher doses, rather than an endorsement of the 30 to 60 Gy10 range as safe. Clinicians should exercise caution even within this range, weighing the potential benefits of tumor control against the risk of functional decline, and consider closer neurologic monitoring or alternative strategies such as lower-dose radiation therapy combined with targeted therapy. Future prospective studies are warranted to establish the optimal dose for this high-risk subgroup. In contrast, for patients with good baseline neurologic function (grade <2), a higher range of BED10 (60-90 Gy) has a protective effect against KPS decline. It is speculated that this is because patients with good neurologic function reserve can tolerate higher doses of radiation therapy, and high-dose radiation therapy can more effectively control the tumor, reducing adverse effects such as nerve compression caused by tumor progression, thereby maintaining KPS.18 This result is consistent with the general understanding in radiation therapy that “patients’ baseline health status affects treatment tolerance and efficacy” and also supplements the details of the “dose-neurological function-KPS” interaction in BM radiation therapy. Furthermore, the SHAP dependence plot reveals a key insight for clinical decision-making: clinicians intuitively avoid prescribing high doses (>70 Gy BED10) to patients with poor baseline neurologic function, which is evidenced by the absence of data points in this region. Our model verifies this intuition, and its quantitative results show that even at moderate doses (<60 Gy), this patient subgroup already faces a significantly increased risk of functional decline (high positive SHAP values). The convergence of clinical practice and model predictions emphasizes the high vulnerability of this patient population and indicates that the risk peak at moderate doses may be a driving factor for avoiding dose escalation in this population in clinical practice. In the SHAP feature importance ranking, BED10 ranks first, followed by the interaction term between BED10 and poor baseline neurologic function. This fully indicates that BED10 itself and its interaction with baseline neurologic function play a core role in predicting KPS decline. At the same time, total dose, KPS score itself, and the interaction term between BED10 and KPS also have relatively important impacts, suggesting that factors such as dynamic changes in KPS are indispensable for accurately predicting KPS decline in integrated modeling. In conclusion, our model clarifies a subtle risk pattern: the main risks for patients with poor neurologic function may be insufficient dose and uncontrolled disease, rather than high-dose toxicity. These findings advocate for a refined treatment strategy, in which dose prescription should be actively individualized, considering not only tumor volume but also, more importantly, the patient’s neurologic vulnerability, thereby moving beyond the “one-size-fits-all” dose limitation model. The development of targeted therapeutic agents such as tyrosine kinase inhibitors has improved the treatment efficacy for lung cancer and subsequent BM.19 Despite the limitations of the blood-brain barrier, targeted therapy has been found to improve the survival of patients with BM. Although targeted therapy has become increasingly common, the optimal timing and application method of combining radiation therapy with targeted therapy remain unclear, especially for lung cancer patients with BM. This study specifically focuses on lung cancer patients with BM and uses XGBoost combined with SHAP to explore the interaction between BED10 and targeted therapy patterns in predicting treatment outcomes. Given the unique clinical challenges and treatment considerations of lung cancer with BM, these findings are particularly relevant to this patient subgroup. From the SHAP feature importance ranking plot, although BED10 is not the most dominant feature, it still occupies a prominent position, indicating its relevance in predicting the outcomes of lung cancer with BM. This is consistent with the following understanding: although dose indicators are important, they interact with other clinical factors specific to lung cancer, such as tumor biological characteristics driven by gene mutations (eg, EGFR, ALK), which affect the therapeutic response to radiation therapy and targeted drugs. The prominence of features such as tumor control status at different time points and physical status combined with neurologic function highlights the multifactorial nature of outcomes in this population, where dynamic tumor changes, patient tolerance, and treatment-related effects converge. The SHAP dependence plots stratified by targeted therapy patterns show different patterns, emphasizing the interaction between BED10 and targeted therapy in lung cancer with BM. For patients without targeted therapy, the relationship between BED10 and SHAP values (representing contributions to model output) shows that lower BED10 values are associated with suboptimal contributions, whereas higher BED10 values have more positive effects. In lung cancer, BM may be invasive, and in the absence of targeted drugs, a higher radiation dose may be required to achieve adequate tumor control, which is consistent with the dose-response principle in radiation therapy.20 However, this must be balanced against the risk of neurotoxicity, which is a major concern for lung cancer patients, as they may also have systemic comorbidities or receive other systemic treatments.21,22 In the case of adjuvant targeted therapy (initiated after radiation therapy), the relatively stable SHAP values across the BED10 range indicate a more balanced interaction. Adjuvant targeted therapy in lung cancer may target driver mutations present in both the primary lung tumor and BM, thereby enhancing the efficacy of radiation therapy. Although there is currently a lack of prospective studies directly comparing low-dose radiation therapy combined with targeted therapy with standard treatment, multiple lines of evidence support the rationality of this hypothesis.23, 24, 25 Targeted therapy initiated after radiation therapy can compensate for potential insufficient local control by eliminating micrometastases. The underlying mechanism may be that radiation therapy disrupts the blood-brain barrier, significantly increasing the delivery concentration of targeted drugs in the central nervous system, thereby more effectively eliminating residual lesions or microlesions.26,27 Therefore, in specific clinical scenarios (eg, patients cannot tolerate high-dose radiation therapy), targeted therapy introduced after radiation therapy may compensate for the insufficient initial radiation therapy dose through the above mechanism. For continuous targeted therapy (administered continuously before and after radiation therapy), SHAP values are mainly positive at higher BED10 values, indicating a synergistic effect. In lung cancer, continuous inhibition of driver pathways (eg, EGFR tyrosine kinase inhibitors for EGFR-mutated lung cancer) combined with higher radiation doses may lead to better tumor control.28 This is consistent with the concept of combined modality therapy, where targeted drugs and radiation therapy act synergistically to improve the local control of BM and potentially affect survival.29 However, it should be noted that this observation is based on model predictions and needs to be verified in prospective clinical trials specifically targeting lung cancer with BM to confirm clinical benefits and evaluate potential toxicities, especially considering the potential toxicity overlap between targeted drugs and radiation therapy. In summary, our study shows that there is an interaction between BED10 and targeted therapy patterns in predicting outcomes in lung cancer patients with BM, and continuous targeted therapy combined with higher BED10 exhibits a particularly favorable interaction. Our model supports a more refined treatment strategy: for patients eligible for continuous targeted therapy, the treatment goal should be to pursue curative high-dose radiation therapy to maximize the synergistic effect and long-term control rate. For patients unable to receive targeted therapy, dose prescription needs to better balance the risks of tumor control and normal tissue damage, and their dose-effect curves may have a plateau phase. Based on the core findings of the integrated BED-based empirical dose-response XGBoost-SHAP model in this study, its clinical application can focus on 4 practical scenarios. First, it can be embedded into radiation therapy planning systems such as CyberKnife and Eclipse. After primary hospital physicians input the patient’s tumor diameter and baseline neurologic function, the system can automatically generate personalized BED10 recommendations (eg, 65-70 Gy10 for 2-3 cm lesions and <60 Gy10 for patients with neurologic function grade ≥2), which shortens the time for planning formulation and reduces dose deviations. Second, it guides the combination of targeted therapy and radiation therapy. For patients with EGFR/ALK mutations who have received targeted drugs for ≥2 weeks before radiation therapy, a BED10 of 60 to 90 Gy10 is preferred; for those who need to suspend targeted therapy, a BED10 of 55 to 60 Gy10 can be selected, and targeted drugs can be restarted within 1 week after radiation therapy. Third, it assists in doctor-patient communication and follow-up management. The model generates “prognostic cards” (eg, “2.2 cm BM + continuous targeted therapy, 1-year local control rate of 76.4%”) to improve patient compliance. For high-risk populations (eg, >3 cm BM + BED10 <55 Gy10), the first follow-up is advanced to 2 months, while for low-risk patients, the follow-up interval is extended to 4 months. Fourth, it serves as a tool for quality control and training. Radiation therapy plans are reviewed using TCD50 values (eg, 57.95 Gy10 for lesions ≤2 cm), and visual tools (TCP curves, SHAP plots) are used to help young physicians master dose principles, ultimately promoting the standardization and individualization of radiation therapy for lung cancer BM. To further streamline the clinical translation of our findings, we have summarized the individualized BED10 dose recommendations for distinct subgroups of patients with lung cancer BM in Table 3. This table integrates the 3 core stratification factors analyzed in our study—tumor maximum diameter, baseline neurologic function, and targeted therapy pattern. Table 3Individualized BED10 recommendations for lung cancer brain metastases based on tumor size, baseline neurologic function, and targeted therapy patternsTable 3 dummy alt textStratificationFactorRecommended BED10 (Gy10)Key rationaleTumor maximum diameter≤2 cm55-60TCD50 = 57.95 Gy10; high radiosensitivity2-3 cm60-65TCD50 = 62.52 Gy10; highest TCP need>3 cm55-60TCD50 = 59.01 Gy10; high-dose harmfulBaseline neurologic functionGrade ≥2 (poor)<60Positive SHAP; no high-dose dataGrade <2 (good)60-90Negative SHAP sufficient reserveTargeted therapy patternNo targeted therapy55-60Suboptimal TCP at low-doseAdjuvant targeted therapy (post-RT)55-60Compensates local controlContinuous targeted therapy (pre-RT + post-RT)60-90Strong synergy; AUC = 0.8534Abbreviations: AUC = area under the curve; BED10 = biologically effective dose (α/β = 10); RT = radiation therapy; SHAP = SHapley Additive exPlanations; TCD50 = dose required for 50% tumor control; TCP = tumor control probability. Individualized BED10 recommendations for lung cancer brain metastases based on tumor size, baseline neurologic function, and targeted therapy patterns Abbreviations: AUC = area under the curve; BED10 = biologically effective dose (α/β = 10); RT = radiation therapy; SHAP = SHapley Additive exPlanations; TCD50 = dose required for 50% tumor control; TCP = tumor control probability. The strength of this study lies in the integration of traditional BED-based empirical dose-response modeling (quantifying dose-response relationships) with XGBoost-SHAP technology (capturing nonlinear interaction effects), which overcomes the limitations of a single method and balances statistical rigor with clinical interpretability. However, this study also has some limitations. First, it is a single-center retrospective study with inherent selection bias. Second, molecular-level data are lacking, and information on patients’ driver gene status (eg, EGFR, ALK, ROS1) and related molecular markers was not collected, which may further refine the interaction between BED10 and targeted therapy. Third, the performance of the ML model needs to be externally validated in prospective, multicenter data sets to demonstrate its generalization ability. Fourth, dose calculation in this study used BED10 based on the standard LQ model, a conventional clinical metric. However, the LQ model has inherent limitations in characterizing cell kill at high doses per fraction (>6-8 Gy) typical of SRS/fSRS,30 and modified models like the linear-quadratic-cubic model may offer more accurate descriptions.31 Additionally, normal tissue low-α/β BED (BED2/BED4) was not included as a covariate; future studies could integrate this metric for more precise toxicity analysis and use revised models for dose conversion to refine ultrahigh-dose per fraction dose-response relationships. Fifth, symptomatic radiation necrosis—a key late dose-limiting toxicity—was not analyzed as a standalone endpoint due to its low incidence in this cohort and lack of systematic prospective diagnostic data (eg, standardized multimodal MRI) in the retrospective design. Future studies with extended follow-up and systematic toxicity assessment are needed to specifically explore the relationship between radiation necrosis and key dosimetric parameters. Conclusions In conclusion, our study demonstrates that integrating ML into radiation oncology analysis can reveal complex interaction effects that are difficult to detect with traditional methods, which has important clinical value in optimizing dose prescription, protecting neurologic function, and formulating combined radiation therapy–targeted therapy strategies. First, the dose-response relationship of TCP has significant volume dependence, providing a quantitative basis for stratified dose prescription. It is found that BM with a diameter of 2 to 3 cm may require a higher BED10 to achieve more effective tumor control. Second, the impact of BED10 on neurologic function is dual, regulated by baseline neurologic function, providing guidance for the selection of safe doses. For patients with poor baseline neurologic function (grade ≥2), a treatment regimen with BED10 <60 Gy10 is recommended, and neurologic function monitoring should be strengthened; for patients with good baseline neurologic function (grade <2), they can benefit from a BED10 of 60 to 90 Gy10. Finally, the interaction between BED10 and targeted therapy patterns improves the systemic-local combined treatment strategy, confirming that “continuous targeted therapy combined with high-BED10 radiation therapy” is a potential key strategy for optimizing local control in lung cancer patients with BM. Future research directions include: (1) verifying the dose recommendations of this study through multicenter prospective trials; (2) integrating genomic data to explore the regulatory mechanism of driver gene status (eg, EGFR, ALK) on the interaction between BED10 and targeted therapy; (3) constructing a personalized TCP prediction nomogram integrating tumor size, dose, and targeted therapy status based on the results of this study to assist clinicians in formulating the most reasonable treatment plans. Acknowledgments Disclosures The authors declare no conflicts of interest. Acknowledgment Weishan He performed the statistical analysis. Footnotes Sources of support: This work was supported by the Special Research Project for Hospital-level Advantage Disease Cohort Study of the Second Affiliated Hospital of Guangxi University of Chinese Medicine (Grant No. 2025YJ-34). Research data are stored in an institutional repository and will be shared upon request to the corresponding author. Supplementary material associated with this article can be found in the online version at doi:10.1016/j.adro.2026.102147. Contributor Information Fasheng Wu, Email: 840766676@qq.com. Jibing Chen, Email: jibingchen398@163.com. Supplementary materials supplementary table E1 References Associated Data Supplementary Materials supplementary table E1