Abstract PS2-02-28: Prognostic Value of a Machine Learning Tool for Recurrence Risk Stratification in HR+/HER2- Early Breast Cancer: A Brazilian Cohort Analysis

Abstract Purpose Genomic assays are unavailable in Brazil’s public health system and are not covered in the private sector, limiting risk stratification in early stage HR+/HER2- breast cancer. Previous studies have developed machine learning tools to predict the 21-gene recurrence score directly from quantitative immunohistochemistry results and clinical factors. We evaluated this approach in a Brazilian cohort to determine the ability to predict prognosis and guide chemotherapy decisions in a resource-constrained setting without genomic testing. Patients and methods We retrospectively analyzed HR+/HER2- breast cancer patients (stage I-III, ≤3 positive nodes) treated at the Instituto Nacional de Cancer, in Brazil (INCA) from 2016 - 2018. Age, grade, histologic subtype, and quantitative ER, PR, and Ki-67 expression were used to predict the 21-gene recurrence score with the previously published machine learning tool. Patients were separated into low-risk group and a high-risk group using a predefined threshold that achieves 90% sensitivity for recurrence score 25. Disease free interval (DFI) and overall survival (OS) were estimated by Kaplan-Meier curves; hazard ratios were derived from Cox models to determine the prognostic difference of the low / high risk groups. Results We identified a sample of 299 cases for this analysis, with an average age of 59. 92 (31%) were PR negative, 212 (70%) had tumor sizer ≥ 2 cm, 163 (55%) had Ki-67 expression of 20% or higher, and 181 (61%) received chemotherapy. Median follow-up was 73 months (95% CI, 71-75). Using the 90% sensitivity threshold, 15.4% of patients were low risk. Recurrence rates were 2.2% in low-risk versus 19.4% in high-risk patients. Five-years DFI at 90% sensitivity was significantly better for low-risk versus high-risk (97.4% vs. 81.8%; long-rank p=0.0054). No significant differences in 5-year OS were observed. In multivariate analysis, stage III disease and Ki-67 ≥ 20% were also independent predictors of recurrence. Although chemotherapy was not associated with a reduction in disease free interval in either group, the excellent outcomes in the low risk cohort with (recurrence rate 6.3%) or without (recurrent rate 0%) chemotherapy, suggest this subgroup can safely forgo such treatment without genomic testing. Conclusion This machine learning prognostic tool demonstrated strong prognostic performance in a Brazilian cohort of HR+/HER2- early stage breast cancer patients, identifying a low-risk subgroup with excellent DFI and no apparent benefit from chemotherapy, and is available for use online (rsncdb.cri.uchicago.edu). These findings support its use for therapy de-escalation when genomic assays are unavailable. Further prospective studies are required to validate these findings. Citation Format: R. Moreira, D. Huo, A. Pearson, I. Small, O. Olopade, J. Bines, F. Howard. Prognostic Value of a Machine Learning Tool for Recurrence Risk Stratification in HR+/HER2- Early Breast Cancer: A Brazilian Cohort Analysis abstract. In: Proceedings of the San Antonio Breast Cancer Symposium 2025; 2025 Dec 9-12; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(4 Suppl):Abstract nr PS2-02-28.