ENTERPRISE AI ANALYSIS

Clinical Prognostic Factors for OS

Survival analysis identified several significant clinical factors influencing overall survival:

• Multiple Brain Lesions: Patients with multiple secondary brain lesions (50.46%) had a significantly poorer prognosis compared to those with a single BM (49.54%), with an HR of 1.73 (CI: 1.13–2.66, p=0.01).
• Synchronous Presentation: Diagnosis of BMs concurrently with the primary tumor (57.8%) was associated with a worse prognosis (HR=1.68, CI: 1.09–2.60, p=0.02) compared to metachronous diagnosis (42.2%).
• Extracranial Metastases: The presence of other systemic metastases at diagnosis (38.53%) correlated with a poorer prognosis, while their absence (61.47%) showed a protective effect (HR=0.63, CI: 0.41–0.97, p=0.04).
• Treatment Options: Active treatments (surgery, gamma knife radiosurgery (GK), whole-brain radiation therapy (WBRT), chemotherapy (CHT)) were significantly associated with longer OS compared to best supportive care (BSC), with HRs < 1 (e.g., GK HR <0.005, Surgery HR 0.01, WBRT HR 0.03, CHT HR 0.055). These are treatment-conditional associations.

Other clinical characteristics like patient age, sex, primary tumor type, diabetes, smoker status, or leptomeningeal involvement did not show significant differences in survival time in this cohort.

MRI Morphological Features Impacting OS

Analysis of 370 secondary brain lesions identified key MRI features influencing overall survival:

• Internal Hemorrhage: Surprisingly, the presence of internal bleeding (18.28%) was associated with better outcomes (HR=0.59, CI: 0.38–0.92, p=0.02), suggesting a protective effect. This often manifested as microbleed lesions.
• Solid Enhancement Type: Solid contrast intake (with or without central necrosis) showed a better survival rate (HR < 1, p=0.04 for solid; p=0.02 for solid with necrosis) compared to cystic or mixed lesions, which were linked to poorer prognoses.
• Lesion Volume and Number: Greater lesion volumes and a higher number of lesions were consistently associated with poorer estimated prognosis in the ML models (SHAP analysis).
• Restricted Diffusion: The presence of restricted diffusion (51.62%) was associated with worse outcomes (HR=1.59, CI: 0.85–2.95, p=0.13), indicating high tumoral cellularity and aggressive behavior.
• Vascularization: Internal vascularization (25.6%) was associated with a negative impact on OS estimation, though not statistically significant in univariate analysis (HR=0.76, p=0.17).

Other features such as peripheric edema, anatomical localization, and hemispheric laterality did not reach statistical significance in the survival analysis.

AI Model Performance & Feature Importance

Four machine learning algorithms (XGBoost, Neural Networks, K-Nearest Neighbor, Random Forest) were tested for 6-month survival prediction using three data setups:

• Clinical Data Only: Models performed poorly (e.g., XGBoost AUC 0.54, RF AUC 0.54).
• MRI Features Only: Models achieved better, but still moderate, results (e.g., RF AUC 0.73).
• Combined Clinical + MRI Data: This setup significantly boosted model performance and discriminatory capacity. The Shallow Neural Networks model was the top performer, achieving an AUC of 0.93 (CI: 0.89–0.97), with 86% accuracy, 82% precision, 88% specificity, and 83% recall. XGBoost followed with an AUC of 0.90.

SHAP analysis for the Neural Networks model revealed that lesion volume had the greatest negative impact on prognosis (larger volumes = worse outcomes). Other strong negative predictors included a higher number of lesions and cystic morphology. Conversely, solid enhancement type and specific treatment methods (WBRT, GK) were strongly associated with positive outcomes.

The addition of morphological MRI features significantly enhances the predictive power of ML algorithms for OS in BM patients, providing crucial insights for clinical decision-making.