Abstract
BACKGROUND: Melanoma has the third highest rate of brain metastases among all cancers and is associated with poor long-term survival. This study aimed to develop machine learning models to predict early death in melanoma brain metastasis (MBM) patients to guide clinical decision-making. METHODS: We analyzed MBM patients from the SEER database and Xinjiang Medical University. Patients were randomly divided into training and testing cohorts (7:3 ratio). Seven machine learning models were developed and validated using cross-validation, ROC analysis, decision curve analysis, and calibration curves to predict cancer-specific early death (CSED) and all-cause early death (ACED) within 3 months of diagnosis. RESULTS: Among 1,547 MBM patients, 531 (34.3%) experienced CSED, and 554 (35.8%) experienced ACED. Key predictive factors included age, treatment modalities (radiation, chemotherapy, surgery), tumor characteristics (ulceration), and extracranial metastases (bone, liver). XGBoost achieved the best performance for ACED prediction (AUC=0.776), while logistic regression performed best for CSED prediction (AUC=0.694). External validation confirmed model reliability with comparable performance. CONCLUSION: These machine learning models demonstrate strong predictive performance and may assist clinicians in early risk stratification and treatment planning for MBM patients. The models provide objective risk assessment tools that could improve patient counseling and guide aggressive versus palliative care decisions.