Abstract
OBJECTIVE: The PREDICT breast prognostic and treatment benefit model has undergone several revisions since its release. The latest version (v3.1) was developed using a data set of 35,474 cases diagnosed between 2000 and 2017 in a single region of England. PREDICT breast provides predicted outcomes at 5, 10 and 15 years, but most clinicians use the 10-year outcomes for decision making. The purpose of this study was to reparameterize the model using a larger data set from across the UK and to compare the performance of v4.0 with that of v3.1. RESULTS: There were 172,208 eligible cases randomly split 50:50 into model development and validation data sets. Cox proportional hazards models were derived for estrogen receptor negative and estrogen receptor positive cancer for breast cancer specific mortality with a third model for non-breast cancer mortality. In cases with at least 5 years follow-up and censored at 10 years, the model was well-calibrated with a less than 5% difference between observed and predicted breast cancer deaths. Model discrimination was also good with AUCs in the validation data of 0.735 and 0.794 for ER negative and ER positive cases respectively. Calibration and discrimination were slightly improved compared to PREDICT breast v3.1.