Individualized prediction of brain metastasis risk in patients with breast cancer.

BACKGROUND

Breast cancer patients have a high risk of brain and central nervous system metastases (BrM) and therefore have poor outcomes. This study aimed to assess the impact of clinico-genomic factors on BrM risk in breast cancer and develop validated nomograms for risk prediction.

METHODS

Clinico-genomic data on breast cancer patients were obtained from the AACR GENIE Biopharma Collaborative. These data were split into training and testing sets by academic institution. Using the training set, risk of BrM was evaluated through multivariable Fine-Gray subdistribution hazard modeling, implementing LASSO penalization for feature selection. Prediction metrics were assessed internally on the training set through bootstrap-cross validation, and externally on the testing set. Calibration curves, AUC, and Brier scores were calculated to assess prediction and discrimination. A nomogram for 2-, 3.5-, and 5- year risk of BrM was constructed.

RESULTS

Synchronous bone metastases, HER+/HR- breast cancer subtype, and TP53 alterations were selected as top predictors of BrM risk. The final multivariable model showed good discriminate capability at 2- (AUC: 68.7; Brier: 5.3), 3.5- (AUC: 65.9, Brier: 12.9), and 5-years (AUC 63.0, Brier: 17.4). This model was externally validated on the testing set and performed well at 2- (AUC 72.5; Brier: 10.5), 3.5- (AUC: 73.3; Brier 13.6), and 5-years (AUC: 74.9; Brier: 16.5).

CONCLUSIONS

Nomograms that calculate individualized BrM risk probabilities for patients with breast cancer can provide clinical utility informing patients and their health care team members on risk of BrM development. An interactive webtool for individualized BrM risk probabilities can be found here: https://gcioffi.shinyapps.io/BrCA_nomo_brm_risk/ .