DSAI-13 PREDICTING CENTRAL NERVOUS SYSTEM METASTASIS ACROSS SOLID TUMORS: A MACHINE LEARNING APPROACH

Abstract

Brain metastases (BM) represent a major source of morbidity across solid tumors, yet robust clinical tools to accurately predict neurotropism are currently lacking. In this study, we analyzed 25,775 patients from the MSK-MET dataset to develop a robust predictive model for BM. An elastic net regularized logistic regression model (glmnet) was trained on primary tumor samples utilizing a 90:10 train-test split (n = 14,070 and n = 1,562). Clinical and genomic features, including cancer type, copy number alterations, and mutations filtered for oncogenic variants via OncoKB, were used for modeling. Feature selection was conducted utilizing SHapley Additive exPlanations (SHAP) importance, restricting the final predictive signature to the top 50 features. The probability threshold was established via the Youden index on the training data and fixed. In the primary tumor test cohort, our model achieved an Area Under the Receiver Operating Characteristic Curve (AUC) of 0.73, with a sensitivity of 64.9% and a specificity of 72.4% for predicting BM risk. To confirm model generalizability, we evaluated the model on an independent cohort of 10,143 metastatic samples, where it maintained robust performance with an AUC of 0.74, a sensitivity of 69.3%, and a specificity of 68.9%. SHAP analysis identified cancer type and recurrent genomic alterations as key drivers of BM risk. Enrichment analysis revealed several of these predictive features map to CNS-relevant biological programs like WikiPathway Glioblastoma Signaling Pathway WP2261 (RB1, CCND1, CDKN2B, EGFR, PIK3R1, ERBB2, and TP53). These findings indicate that BM risk might be inherently encoded within primary tumor genomics, driving partial organotropism and potential tumor pre-adaptation to the brain microenvironment. These findings demonstrate that integrating genomic profiling with clinical features can effectively predict BM and may inform surveillance strategies for earlier detection of BM. Cancer-specific predictive models are currently under development to further enhance and personalize clinical utility.