Risk factors (RF) for brain metastases (BM) in patients (pts) with metastatic breast cancer (MBC): An analysis of US electronic health records (EHRs).

Abstract

2035 Background: BM are a significant clinical challenge in pts with MBC yet risk stratification for early identification is suboptimal. We leveraged a large, contemporary, real-world database to characterize RF associated with BM to inform strategies for enriched surveillance and early detection. Methods: Selected pts from the nationwide Flatiron Health de-identified EHR-derived database had MBC, had initiated first-line treatment (1L tx) before March 2023 (allowing for > 1 year [yr] of potential follow-up [FU]), and were free of BM at tx initiation. Clinical characteristics were examined as potential RF for BM incidence at any time during FU, using univariate sub-distribution (sd) and cause-specific (cs) hazard ratios. Additional analyses focused on BM risk at 3 yrs from tx initiation and included longitudinal data sequential Cox models with landmarks at every 6 months of FU, and a nested case-control (NCC) design for concurrent BM detection. Analyses were conducted on complete cases, no imputation method was used for missing data, and regression methods using multivariate analyses were used to mitigate confounding. Predictive, modeling-based machine learning (LASSO Cox regression, random survival forests) was conducted using cs hazard ratios to identify potential predictors amongst 90 candidates (data-driven approach using most of the dataset). Analyses were conducted in the overall cohort and stratified by subtype: HER2-negative/hormone receptor-positive (HER2–/HR+), HER2-positive/HR-negative (HER2+/HR–), HER2+/HR+, and triple-negative breast cancer (TNBC). Results: The study included 21,368 female pts initiating 1L tx (n = 14,898 HER2–/HR+, 1006 HER2+/HR–, 3468 HER2+/HR+, 1996 TNBC), with 2,530 BM events. Younger age, HER2+ and TNBC subtypes, and more extensive metastasis (≥2 organ sites, particularly liver, lung, or lymph nodes) were associated with higher BM risk (Table); bone-only metastases conferred a lower risk. sd and cs hazard ratios were largely concordant. NCC analyses identified similar predictors for concurrent BM. LASSO Cox modeling yielded a C-index of 0.74 overall (HER2+ 0.70; HER2–/HR+ 0.73; TNBC 0.62). Similar C-indices were seen with random survival forests. Conclusions: Clinical characteristics, including metastatic distribution and tumor subtype, can help identify pts at higher BM risk within 3 yrs of initiating MBC tx. Competing risks (of BM and death) did not appear to substantially affect results, except for recurrence time and ECOG PS. Although these findings are encouraging, further refinement of predictive models is needed to improve discrimination and guide targeted neuroimaging and early intervention strategies. sd hazard ratio Age, yrs (reference [ref]: <45) 45–55 0.81 56–65 0.62 66+ 0.31 Subtype (ref: HER2–/HR+) HER2+/HR+ 2.45 HER2+/HR– 3.40 TNBC 2.33 Metastatic sites, n (ref: 1) 2–3 1.55 4+ 2.63