Abstract
BACKGROUND
Breast cancer (BC) is a leading cause of mortality and the most frequent malignancy in women; most deaths are due to metastatic disease, expressly brain metastases (BM). Currently, there is no biomarker or a prediction model to accurately predict BM.
OBJECTIVE
To generate a BM prediction model from variables acquired at BC diagnosis.
METHODS
A retrospective cohort of BC women diagnosed from 2009 to 2020 at a single center was divided into training (TD) and validation datasets (VD). The TD was used to generate a multivariable prediction model. The modeĹs performance was measured applying the area under the curve (AUC), C-statistic, and the Akaike information criteria (AIC).
RESULTS
5,009 patients were divided into a TD (n 3339) and a VD (n 1670). In the TD, the model with the best performance (lowest AIC) was built with the following variables: Age, estrogen receptor status, tumor size, axillar adenopathy, AJCC anatomic clinical stage, Ki-67, and the Scarf-Bloom-Richardson score. This model had an AIC of 1241 and an AUC of 0.793 (95%CI 0.761 – 0.825) p <0.0001 in the TD. A 10-fold cross-validation showed good stability of the model. In the VD, the model had an AUC = 0.812 (IC95% 0.774 – 0.850) P < 0.0001 and an AIC = 644. Finally, we present with an online APP and an online calculator for its clinical use.
CONCLUSION
In a retrospective cohort of women with breast cancer, a prediction model built with clinical and pathological variables at diagnosis displayed a robust performance to measure the individual odds of BM. The model is currently considered for external validation in other institutions and countries.