OBJECTIVES:

Breast cancer follow-up aims at the early detection of locoregional recurrences (LRR) and potential second primaries (SP). Predicting a patient’s time-dependent risk for such an adverse event might contribute to improve the diagnostic process.

METHODS:

Patient-, tumor-, and treatment characteristics of 11,129 patients from 2007, 2008 and 2012 were provided by the Netherland’s cancer registry (NCR). Using the incidence dates of LRRs and SPs within the first five years after primary treatment, different prognostic models based on a) a Cox-regression, b) a flexible parametric Royston-Parmer model, c) a random survival forest (RSF) were developed based on a training- and a testing-set. Their quarterly risk-predictions for LRR and SP covering the corresponding, subsequent 24 months during the 5-year follow-up period were compared to the actually observed event-rates via Chi-square test. Discriminative ability was assessed using time-dependent c-statistics.

RESULTS:

Of the observed patients, 6.7% developed an LRR or an SP within five years after primary therapy. The accuracy of the three proposed algorithms is comparable: Neither in the training-, nor in the testing-set, significant differences between the observed event-rates and the mean predicted risks per quintile can be observed. The average time-dependent discriminative ability varies only moderately between the three models, ranging between an AUC of 0.67 and 0.72 in the training-set and 0.65 - 0.66 in the testing-set. Patients belonging to the highest quintile according to their risk-prediction by the Cox-model are likely to have 7.4 times more LRRs or SPs than those belonging to the lowest

CONCLUSIONS:

The model performances vary only slightly between the three proposed methodologies. Being the most transparent , the Cox-regression model seems to be the most suitable option to build the prediction model on, enabling health professionals to further personalize follow-up strategies.