Prognostic value of automated KI67 scoring in breast cancer: a centralised evaluation of 8088 patients from 10 study groups

BACKGROUND: The value of KI67 in breast cancer prognostication has been questioned due to concerns on the analytical validity of visual KI67 assessment and methodological limitations of published studies. Here, we investigate the prognostic value of automated KI67 scoring in a large, multicentre study, and compare this with pathologists’ visual scores available in a subset of patients. METHODS: We utilised 143 tissue microarrays containing 15,313 tumour tissue cores from 8088 breast cancer patients in 10 collaborating studies. A total of 1401 deaths occurred during a median follow-up of 7.5 years. Centralised KI67 assessment was performed using an automated scoring protocol. The relationship of KI67 levels with 10-year breast cancer specific survival (BCSS) was investigated using Kaplan–Meier survival curves and Cox proportional hazard regression models adjusted for known prognostic factors. RESULTS: Patients in the highest quartile of KI67 (>12 % positive KI67 cells) had a worse 10-year BCSS than patients in the lower three quartiles. This association was statistically significant for ER-positive patients (hazard ratio (HR) (95 % CI) at baseline = 1.96 (1.31–2.93); P = 0.001) but not for ER-negative patients (1.23 (0.86–1.77); P = 0.248) (P-heterogeneity = 0.064). In spite of differences in characteristics of the study populations, the estimates of HR were consistent across all studies (P-heterogeneity = 0.941 for ER-positive and P-heterogeneity = 0.866 for ER-negative). Among ER-positive cancers, KI67 was associated with worse prognosis in both node-negative (2.47 (1.16–5.27)) and node-positive (1.74 (1.05–2.86)) tumours (P-heterogeneity = 0.671). Further classification according to ER, PR and HER2 showed statistically significant associations with prognosis among hormone receptor-positive patients regardless of HER2 status (P-heterogeneity = 0.270) and among triple-negative patients (1.70 (1.02–2.84)). Model fit parameters were similar for visual and automated measures of KI67 in a subset of 2440 patients with information from both sources. CONCLUSIONS: Findings from this large-scale multicentre analysis with centrally generated automated KI67 scores show strong evidence in support of a prognostic value for automated KI67 scoring in breast cancer. Given the advantages of automated scoring in terms of its potential for standardisation, reproducibility and throughput, automated methods appear to be promising alternatives to visual scoring for KI67 assessment. ELECTRONIC SUPPLEMENTARY MATERIAL: The online version of this article (doi:10.1186/s13058-016-0765-6) contains supplementary material, which is available to authorized users.