The initial literature search identified 497 articles. Following the removal of duplicates and primary screening, 43 articles were assessed as full text for eligibility in the meta-analysis records. Finally, we included a total of 23 articles [2,3,4, 6, 9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27] selected by eliminating articles with incomplete information and articles that used different classification (Fig. 1).

Prisma flow diagram of studies included

The systematic review and meta-analysis include twenty-three studies published between 1999 and 2023. Six studies were conducted on cadaveric samples [9, 15, 16, 23, 27, 28], and seventeen were performed intraoperatively [2,3,4, 6, 10,11,12,13,14, 17,18,19,20,21,22, 24,25,26]. The studies exhibit a broad geographic distribution, with 10 studies from Asia [3, 4, 9,10,11,12, 16, 18, 21, 28], 6 from Europe [6, 14, 19, 20, 25, 26], 4 from South America [2, 15, 22, 24], and one each from North America [23], Africa [17], and Australia [27]. Among the twenty-three studies, 1,636 patients were included, and 1,883 axillae were evaluated (494 from cadaveric dissections and 1,389 from intraoperative dissections): 765 from Asians, 570 from Europeans, 318 from South Americans, 200 from North Americans, 30 from Africans, and 28 from Australians (Table 1).

The AQUA tool probes for potential risk of bias in 5 studies domains (objectives and subject characteristics, study design, methodology characterization, descriptive anatomy and reporting of results) (Henry, B.M. et al. 2017 [8]). The risk of bias within each domain is normally categorized as “Low”, “High” or “Unclear”. Twenty-two of the studies included showed low risk in domain one (Objectives and Subject characteristics), ten studies showed high risk of bias in domain three (Methodology characterization), mainly because there is an important reduction of possibility of studying anatomy during an intervention. A summary of the assessment of quality and risk of bias by the AQUA tool is displayed in the Fig. 2.

Assessment of quality and risk of bias by the AQUA tool

For the primary outcome—prevalence of the ICBN—pooled prevalence estimates (PPES) and their 95% confidence intervals are reported using MetaXL software (V. 5.3). We used a DerSimonian and Larid random effects model with a double arcsin transformation, normalized prevalence, and a 0.5 continuity correction. Heterogeneity was investigated through the I2 statistics, Cochrane’s Q statistic, and a visual analysis of forest plots and funnel plots. In addition, we examined evaluation type (cadaveric dissection or intraoperative dissection) and geographic region of the first author’s affiliation (Africa, Americas, Asia, Oceania, or Europe) as factors. A leave-one-out sensitivity analysis was conducted to examine the effect of outlying studies. Funnel plots and Doi plots were used to investigate possible sources of bias—including publication bias. The leave-one-out sensitively analyses yielded PPEs from 99.3% with the Andersen 2014 [19] study excluded to 99.1% with the Loukas et al. 2016 [23] study excluded.

There was great variety in the outcome categories that authors used in the secondary outcomes: this is the reason why a single multicategory pooled prevalence estimate was not possible for any secondary outcomes. Furthermore, it was not feasible to report on the PPEs for tens of secondary outcomes if each of the outcome categories were reported as individual binary outcomes. Therefore, we descriptively report the secondary outcomes using only raw, marginal proportions.