Pathological complete response (pCR) is an established surrogate marker for survival in breast cancer (BC) patients treated with neoadjuvant chemotherapy. Prediction of pCR based on clinical information available at biopsy, particularly the biomarkers estrogen receptor (ER), progesterone receptor (PgR) and Ki-67 expression, might assist in the identification of patients who benefit of preoperative chemotherapy. Although biomarker assessment is mostly reported as the percentage of positively stained cells with values from 0 to 100%, cut-off points are used to classify patients into groups. Aim of this study was to examine established cut-off points and to develop a prediction tool estimating a patient’s pCR likelihood obtained from clinical predictors and these biomarkers as assessed during clinical routine or categorically with established or newfound thresholds.
This study included all HER2-negative BC patients from one German institution treated with neoadjuvant chemotherapy from 2002 to 2017, having complete observations (N = 829). Various logistic regression models for predicting pCR which differ from each other by the usage of the biomarkers were set-up: (M1) continuous biomarkers from 0 to 100% as assessed during clinical routine, (M2) categorical (positive/negative) biomarkers with established thresholds, (M3) categorical biomarkers with newfound thresholds. Prediction accuracy (e.g., AUC) was assessed using cross-validation.
A total of 163 (19.7%) patients achieved a pCR. The optimal cut-off points for ER, PgR and Ki-67 were 30%, 10% and 35%, respectively. The prediction model M1 with continuous biomarkers was more precise (cross-validated AUC, 0.863) than the prediction models M2 and M3 (both 0.854). The most accurate model M1 had a sensitivity of 0.80 at a specificity of 0.78. Beside these biomarkers, a patient’s likelihood of achieving a pCR depended on age at diagnosis, clinical tumor stage and grading.
Using biomarkers as continuous variables yielded more precise predictions than when used categorically. Therapy decisions should base on predicted pCR probabilities obtained from multivariable prediction models rather than single biomarker values.
Erlangen University Hospital, Erlangen.
Has not received any funding.
A. Hartmann: Personal fees: AstraZeneca, BMS, MSD; Grants: Biontech, Nanostring; Personal fees and non-financial support: Roche, Sysmex; Grants: Janssen, Cepheid outside the submitted work. P.A. Fasching: Grants and personal fees: Novartis; Personal fees: Pfizer, Roche, Teva, Celgene; Grants: Biontech, outside the submitted work. All other authors have declared no conflicts of interest.