An actionable machine learning–driven clinicogenomic model as a predictor of brain metastasis risk in breast cancer.

Abstract

106 Background: Brain metastasis (BM) is a frequent site of disease progression for patients living with metastatic breast cancer (MBC). Guidelines do not recommend routine MRI brain surveillance in asymptomatic patients. Consequently, patients with MBC who develop BM often present with extensive disease, leading to lasting neurological damage or death. Methods: This study included MBC patients without known BM at presentation who underwent genomic sequencing of a non-BM specimen with MSK-IMPACT, a custom tumor-normal next-generation sequencing assay, within one year of M1 diagnosis. We developed an ensemble time-dependent LASSO machine learning (ML) model with BM-free survival (BMFS) as the primary endpoint, integrating baseline clinical, pathologic, and genomic features for risk stratification, using a cross-validation framework. Benchmarking was conducted using a time-dependent neural network designed to model competing risks (DeepHit), and further validation was performed using an independent clinical trial dataset. Results: 1594 MBC patients were divided into a training set (n=1118) and a test set (n=476), with 320 events over a median follow-up of 39.7 months. The ensemble ML model identified distinct clinicogenomic features associated with shorter BMFS, including receptor subtype, ER/PR percent positivity, menopausal status, metastatic burden, metastatic site distribution, disease-free interval, and alterations in TP53 , ERBB2 , and RB1 . The model stratified patients into low-, intermediate-, and high-risk groups (training C-index: 0.690; test C-index: 0.696). In the test cohort, 24-month BMFS was 68%, 89%, and 98% in high, intermediate, and low risk groups (HR 19.2, p<0.001 high vs. low risk; HR 6.5, p<0.001 intermediate vs. low risk), with model predictions retaining robust predictive ability beyond 24 months (time-dependent AUC at 10 years of 0.79). These results were confirmed using DeepHit, a competing-risk-specific neural network (training C-index 0.71; test C-index 0.61). The model similarly identified high-risk patients within a single-arm phase II clinical trial dataset utilizing MRI screening in patients with MBC. Conclusions: We developed an actionable ML-driven clinicogenomic model that accurately identifies MBC patients at high risk of developing BM. Biologically plausible and readily available features defined a high-risk patient category with a >30% risk of developing BM within 2 years and would likely benefit from MRI screening. The results will be prospectively validated in BRAINSTORM (Breast Cancer Radiologic Assessment and Intervention for Neurological Surveillance, Tracking, and Optimized Risk Management), a phase II randomized clinical trial of intensified MRI surveillance versus standard symptom-based screening in high-risk MBC patients.