3635 - Clinical and Imaging Determinants of Overall Survival in Human Epidermal Growth Factor Receptor 2-Positive Breast Cancer with Brain Metastases
Presenter(s)
Q. Xu1, B. Yagmurlu1, S. Agarwal2, M. Iv3, M. Hayden1, K. C. Horst1, G. Sledge1, M. Pegram4, and H. Itakura1; 1Stanford University School of Medicine, Stanford, CA, 2Stanford Brain Tumor Center, Stanford, CA, 3Department of Neuroimaging and Neurointervention, Stanford University, Stanford, CA, 4Stanford University School of Medicine, Palo Alto, CA
Purpose/Objective(s): Development of brain metastasis (BM) is a devastating clinical event for patients initially presenting with non-metastatic human epidermal growth factor receptor 2 (HER2)-positive breast cancer (BC) because it shortens time to death. The purpose of this study was to identify the key clinical and imaging-based features that predict overall survival (OS) among patients with HER2-positive BC who develop BM. We hypothesized that quantitative imaging (radiomic) features extracted from BM on brain magnetic resonance imaging (MRI) studies enhance prediction of those with the highest risk of early death.
Materials/Methods: Our retrospective study included 272 patients initially diagnosed with non-metastatic HER2-positive BC who subsequently developed BM. We collected survival outcome data and 22 clinical features across four categories: (1) Demographic (sex, age, race, ethnicity, smoking history); (2) BC-related (hormone receptor status, T stage, N stage, Ki-67, tumor grade); (3) BM-related (interval from BC to BM, first metastatic site, intracranial location, number of affected regions); and (4) Treatment-related (trastuzumab-based regimens for BC, first-line targeted agent for BM, local BM treatment: surgery + stereotactic radiosurgery (SRS), surgery + whole brain radiation (WBRT), surgery only, radiation only, or no local treatment). We segmented and extracted 1098 radiomic features from each BM, aggregating multiple lesions per patient. Dividing our cohort into training (n=241) and test (n=31) sets, we built our prediction model of OS on the training set based on: Cox proportional hazards regression with Coxnet (LASSO and Elastic Net regularization); 10-fold cross-validation; and using clinical features with and without radiomic features. Model performance was assessed using the concordance index (C-index) on the held-out test set.
Results: In our cohort of 269 women and 3 men (mean age: 54.5 ± 12.8 years), the clinical-only model performed well (C-index: 0.701 [95% CI: 0.54–0.83]). Three key OS predictors were identified: (1) Surgical resection + SRS was the strongest, significantly improving OS (HR: 0.78 [95% CI: 0.65–0.94]); (2) Trastuzumab-based combination therapy as first-line BM treatment was associated with better survival; and (3) More affected brain regions correlated with worse OS. The addition of radiomic features to the model significantly improved the prediction model performance (C-index: 0.73 [CI: 0.60–0.84]).
Conclusion: A limited number of treatment-related and MRI-based radiomic features strongly predicted OS, indicating a tractable method for discriminating patients warranting differential levels of clinical attention upon BM diagnosis.