Related Experiment Videos
Development and external validation of an XGBoost-based prognostic model for small cell lung cancer with brain
Pei Jiao1, Chen Chen2, Koudong Zhang2
1Department of Medical Oncology, Yancheng Dafeng People's Hospital, Yancheng, China.
Background:
Patients with small cell lung cancer (SCLC) and brain metastases (BM) have poor and heterogeneous survival outcomes, while existing prognostic tools have shown limited discriminatory performance. We aimed to develop an extreme gradient boosting (XGBoost) based prognostic model and perform preliminary external validation.
Methods:
Patients with small cell lung cancer and brain metastases (SCLCBM) diagnosed between 2010 and 2020 were identified from the Surveillance, Epidemiology, and End Results (SEER) database for model development and internal testing. External validation was performed in an independent single-center Chinese cohort. Variables screened by univariable Cox regression and retained in multivariable Cox regression analyses were used as predictors in the XGBoost models. Grid search was used for hyperparameter tuning, and candidate parameter combinations were evaluated using 10-fold cross-validation. Model performance was assessed using the areas under the curves (AUCs) and confusion matrices, and SHapley Additive exPlanations (SHAP) analysis was used to interpret model predictions.
Results:
The study included 7,692 patients from the SEER database and 128 patients in the external validation cohort. The final models incorporated demographic, tumor-related, treatment-related, and metastatic factors, including age, sex, T and N stages, chemotherapy, radiotherapy, surgery, and extracranial metastases. AUCs for 3-, 6-, 12-, and 18-month overall survival (OS) were 0.840, 0.801, 0.738, and 0.759 in the internal testing cohort and 0.779, 0.790, 0.739, and 0.763 in the external cohort, respectively. SHAP analysis consistently ranked chemotherapy as the most influential predictor.
Conclusions:
The XGBoost models showed moderate discrimination across the evaluated prediction horizons in both the internal testing cohort and a small single-center external cohort. Further validation in larger, prospective, multicenter cohorts is required before clinical use.