Identifying Prognostic Factors in Brain Metastasis Patients Using MRI Morphological Features: A Machine Learning and Survival Analysis Approach

Background/Objectives: Our study aims to identify potential new MRI features of brain metastases (BMs) that could be further used in overall survival (OS) assessment. Methods: A total of 109 patients with BMs were included. Kaplan–Meier analysis, the log-rank test, and Cox Regression were implemented in the survival analysis. The first ten significant features were incorporated into four distinct machine learning (ML) algorithms to predict six-month survival. Results: Survival analysis revealed that multiple brain lesions and synchronous presentation were associated with a poor prognostic value (HR > 1; p = 0.01, p = 0.02). Other features demonstrated a protective effect on OS including the absence of extracranial lesions (HR < 1, p = 0.04) and the presence of solid enhancement (HR < 1, p < 0.05). In this observational cohort, treatment was associated with longer OS—including surgery, gamma knife radiosurgery, whole brain radiation therapy, and chemotherapy—compared to best supportive care (HR < 1, p < 0.005); these treatment-related hazard ratios are not interpreted causally. The shallow Neural Networks model was the top-performing ML model, achieving an AUC of 0.93 (CI = 0.89–0.97). According to the Shapley Additive Explanations analysis, the solid enhancement type had a positive impact on OS, whereas a higher number of lesions, larger volumes and a cystic morphology were associated with negative outcomes. Conclusions: Our results confirm that including morphological MRI features of BMs in the prediction of OS significantly contributes to the enhancement of ML algorithms’ prediction and discriminatory capacity.