The INFLUENCE 3.0 model: time-dependent risks of locoregional recurrence and contralateral breast cancer, now also including patients treated with neoadjuvant systemic therapy

Background: Individual risk prediction of locoregional recurrence (LRR) and contralateral breast cancer (CBC) supports decisions for personalized follow-up planning in patients who have been treated for primary breast cancer. This study aimed to improve a previously developed prediction tool by revising the underlying models and expanding them for patients who received neoadjuvant treatment.

Material and Methods: Data from the Netherlands Cancer Registry was collected of women diagnosed with nonmetastatic breast cancer between 2012 and 2016, treated with surgery. Two modelling approaches, Cox regression and Random Survival Forest were compared to predict LRR and CBC during the first five years since surgery. Separate models were developed for patients treated with neoadjuvant systemic therapy (NST). Missing data was imputed using the mice package in the statistical software program R. The models were internally validated by 100 times bootstrap resampling to assess the discrimination using the Area Under the Curve (AUC) and the calibration using the Integrated Calibration Index.

Results: In total, 49,631 and 10,154 patients were included in the non-NST and NST group, respectively. There were 825 (1.7%) and 296 (2.9%) patients diagnosed with a LRR as a first event in the non-NST and NST group, respectively. CBC as first event was diagnosed in 1025 (2.1%) and 141 (1.4%) in the non-NST and NST group, respectively.

Age, mode of detection, histology, sublocalisation, differentiation grade, pT, pN, hormonal receptor status ± endocrine treatment, HER2 status ± targeted treatment, surgery ± direct reconstruction ± radiation therapy, and chemotherapy were identified as significant predictors for LRR and/or CBC in non-NST patients. The models developed for NST patients included the same variables, but excluding pT and pN status, and including axillary lymph node dissection and presence of pathologic complete response.

In the non-NST cohort, the Cox model was chosen as best performing model with an AUC of 0.77 (95%CI: 0.767–0.773) for prediction of LRR. The random survival forest model performed best for prediction of CBC, with an AUC of 0.68 (95%CI: 0.67–0.68). In the NST cohort, for both outcomes the random survival forest model performed best, with AUCs of 0.77 (95%CI: 0.76–0.78) and 0.73 (95%CI: 0.69–0.76) for LRR and CBC, respectively. Regarding calibration, ICI values were all <0.01 (observed-predicted difference was less than 1%).

Conclusions: This revised and expanded INFLUENCE model showed moderate to good performance in the prediction of LRR and CBC. INFLUENCE 3.0 has been made available as an online tool to enable clinical decision support regarding personalised follow-up strategies, also for patients treated with NST.