Developing and validating a predictive model on 10 years survival of breast cancer in New Zealand

Abstract

Background: We developed and validated a predictive model for breast cancer specific 10-year survival in women with breast cancer in New Zealand. Methods: We used the data collected through the Breast Cancer Registries from 2000 to 2014 in Auckland and Waikato regions of New Zealand. Using a multivariate Cox survival model, we studied women with invasive breast cancer, and assessed risk factors based on their clinical importance and statistical significance on mortality. We developed the predictive model from the Auckland database, and compared predicted 10 years survival against observed 10 years survival. Then we applied the model to the Waikato database, and compared predicated 10 years survival versus observed survival, as a test of validation and transportability. Using the Cox model, we first obtained probability of survival at baseline by setting values of risk factors at zero, and then we powered the 10 years baseline survival by the exponential of linear combination of risk factors included in the model, and provided the predicted 10 years survival for each patient in Auckland. Using the Kaplan Meir method, we obtained observed 10 years survivals and compared them with predicted 10 years survivals by decile. The decile was determined based on 10 years survival probability among patients in the database. Findings: The risk factors used in the model were age, tumour grade, tumour size, number of positive lymph nodes, metastatic disease at diagnosis, ER, PR and HER2 receptors, histology of cancer, and patient ethnicity. Lymphovascular invasion was also assessed. Some 9358 observations which included 865 breast cancer deaths, were used in the Cox model on the Auckland database. Overall validity of the model was measured by the C statistic which was high, 0.84. In Auckland, predicted 10 year survivals in decile groups ranged from 0.32 to 0.99. The predicted survivals were in most groups within 95% CI of the observed survival for the respective decile. Applying the model to Waikato gave similar results. Conclusion: These results based on application to an independent data set show the stability, validity, and potential clinical value of the model. We plan to further define and test predictive models, and compare the results with those of other models which have been developed overseas and have not yet been tested in New Zealand. We will then explore potential clinical applications. Acknowledgement: This work was supported by the Auckland Medical Research Foundation, and the Auckland and Waikato breast cancer registries.