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

Background: Lung cancer is the leading cause of cancer-related death worldwide, and brain metastasis (LCBM) is among its most lethal complications with limited prognostic tools. This study aimed to build a machine learning model using the extreme gradient boosting (XGBoost) algorithm to predict the prognosis of lung cancer brain metastasis (LCBM) patients more accurately. Methods: Data from 15,216 patients with synchronous LCBM in the Surveillance, Epidemiology, and End Results (SEER) database diagnosed between 2010 and 2014 were used for model development and internal testing, while a temporally distinct cohort of 3,181 patients diagnosed in 2015 was used for temporal validation. The XGBoost algorithm was trained to predict 6-month, 1-, 2-, and 3-year overall survival, incorporating 13 clinical variables identified via multivariate Cox regression. Model performance was comprehensively evaluated via area under the curve (AUC) of the receiver operating characteristic (ROC) curve, confusion matrix, calibration curves, and decision curve analysis. Pairwise AUC comparisons between models were performed using the DeLong test with Holm correction for multiple comparisons. Propensity score matching (PSM) and Kaplan-Meier analysis were conducted to explore the association between surgical treatment and survival. Results: The XGBoost model achieved AUC values ranging from 0.805 to 0.822 and 0.796 to 0.816 in the training and test sets, respectively, and 0.780 to 0.822 in temporal validation. DeLong test confirmed statistically significant differences in AUC between XGBoost and all four competing models across all follow-up time points (all adjusted P<0.05). The XGBoost model also demonstrated superior calibration and higher clinical net benefit compared to traditional algorithms. Chemotherapy was the variable with the highest predictive importance, while the relative importance of age at diagnosis, histologic type, and radiotherapy varied across prognostic time horizons-patterns that should be interpreted as hypothesis-generating. After PSM and Kaplan-Meier analysis, surgery was associated with improved overall survival in the overall LCBM cohort, with exploratory analyses suggesting potential variation in this association by sex, race, histologic type, lymph node (N) stage, and the presence of extracranial metastases. Conclusions: The XGBoost model demonstrated good discriminative performance, excellent calibration, and superior clinical net benefit for survival prediction in synchronous LCBM patients using routine clinical variables. Chemotherapy showed the highest predictive importance, though treatment-related variables are subject to confounding by indication. Exploratory analyses suggest heterogeneous associations between surgery and survival across clinical subgroups, warranting prospective validation. These findings provide robust evidence supporting the clinical utility of the XGBoost model as a promising prognostic tool for LCBM patients.