Prognostic and predictive performance of PREDICT 2.1, PREDICT v3, and RSClin in node-negative early breast cancer: a TEAM pathology substudy.

[BACKGROUND] Despite widespread use of clinico-pathologic and genomic risk scores in early breast cancer (EBC), questions remain as to whether they are predictive, prognostic, or both. We evaluated risk score performance with actual patient outcomes in a trial dataset.

[METHODS] Discrimination and calibration of PREDICT 2.1 and PREDICT v3 for overall survival (OS) and RSClin for distant metastasis-free survival (DMFS) were compared with actual patient outcomes in 645 postmenopausal, node-negative, hormone-positive patients with EBC from the TEAM pathology substudy. Estimated chemotherapy benefit (low < 3%, moderate 3-5%, high > 5%) was also compared.

[RESULTS] Harrell's C-statistic for OS was 0.6893 and 0.6986 for PREDICT 2.1 and v3, and 0.6603 for DMFS using RSClin. PREDICT 2.1 underestimated and PREDICT v3 overestimated the 10-year OS. RSClin underestimated the 10-year DMFS. Of patients predicted by RSClin to derive a large benefit from chemotherapy, 44% and 99% were estimated to have a small (< 3%) benefit by PREDICT 2.1 and v3, respectively.

[CONCLUSION] All three scores demonstrated moderate discriminatory ability while PREDICT v3 and RSClin had better calibration. However, scores varied widely on expected benefit from chemotherapy, suggesting that RSClin is prognostic but not predictive of chemotherapy benefit. Given the cost of RSClin, further studies are required.

[TRIAL REGISTRATION] The trials are registered with ClinicalTrials.gov , NCT00279448, NCT00032136, and NCT00036270; NTR 267; Ethics Commission Trial 27/2001; and UMIN, C000000057.