Data synthesis: A total of four hundred and five potentially relevant publications were found according to our initial search strategy, one hundred and sixteen publications from PubMed/Medline, two hundred and eighty-nine publications from Embase. There were three hundred and one publications after duplicate publications were removed; of these, we excluded two hundred and sixty-eight articles because the title or abstract did not meet the eligibility criteria. Eight systematic reviews, four studies on metastatic breast cancer, and eleven non-neoadjuvant chemotherapy were further excluded. Eventually, nine articles[11–19] (eight studies) were identified as eligible for our analysis, including eight independent studies for OS, six studies for RFS, and two for DFS. The flow chart of the literature search and study selection is shown in Fig. 1.
Study quality: The eligible studies were conducted in Australia, Saudi Arabia, Italy, Peru, the USA and Canada. The number of participants ranged from 58-1101 per study for 4521 patients across the ten studies. Detailed baseline characters of each eligible study were listed in table 1. Table 2 showed the HR results from each eligible study.
Data synthesis: The combined HR for OS was 1.52 (95% CI 1.29–1.78; P = 0.000) by fixed-effects model. No statistically significant heterogeneity was found (P = 0.114; I2 = 39.8%), and this difference was statistically significant (Z = 4.31; p = 0.000), Fig. 2.
The pooled HR for RFS/DFS was 1.47 (95%CI: 1.27–1.71,I2 = 61.9%,Fig. 3a) by random-effects model, with significant heterogeneity. When the study by T.L. Sutton et al.  which contributed substantial heterogeneity was excluded, low heterogeneity was found (P = 0.142, I2 = 37,6%), the pooled HR was 1.41 (95% CI:1.22–1.64), the data was statistically significant (Z = 3.77, p = 0.000), Fig. 3b.
Sensitivity analysis was used to assess the root of heterogeneity. As shown in Fig. 4, the individual data set had no significant influence on the OS and RFS/DFS, demonstrating the reliability and stability of the results in our meta-analysis.
Asymmetrical funnel plot showed in Fig. 5a and P = 0.002 < 0.05 for Egger's test demonstrated that our studies existed publication bias for OS. Statistical tests also showed publication bias for RFS/DFS was found in our study ( p < 0.05 for Egger's test and p = 0.003 for Begg's test), Fig. 5b. Then, trim-and-fill method was conducted to adjust funnel plots[20, 21]. Three missing studies were added in analysis of RFS/DFS and four studies in OS, Fig. 6. The recalculated results were still significant for RFS/DFS (HR = 1.33, 95% CI = 1.04–1.72; random-effects model; p < 0.01) and OS (HR = 1.38, 95%CI = 1.07–1.78, random-effects model; p < 0.01), indicating the conclusions in our meta-analysis were stable and reliable.