This meta-analysis was performed according to the recommendations of the PRISMA [14] and the Cochrane Collaboration [15]. The protocol was registered in PROSPERO (https://www.crd.york.ac.uk/prospero, CeRD42017060238).

MEDLINE, EMBASE, and Cochrane Central were used to search articles without restriction to year of publication. Primary investigation was performed to confirm search keywords and strategies.

We searched MEDLINE, EMBASE, and Cochrane Central using the terms (“tranexamic acid” AND (“bleeding” OR “transfusion”)). Broad search terms had been used to achieve higher sensitivity. The language of the article was limited to English only. The last search was in June 2017. We manually checked the references of review articles to find any additional relevant studies.

Two authors (DL and SL) searched for the articles using the following databases. The selection of data was carried out according to the criteria set out in the protocol. Three investigators (DL, KK, and WYL) performed the selection or exclusion of studies. Primary selection was performed based on the article title and abstract by three investigators. Secondary selection was performed independently using the full-text. Studies for final assessment were selected based on agreement among the investigators. If needed, a third-party investigator (SL) participated and the final decision was made based on the majority.

Studies were included in the meta-analysis if they satisfied the following criteria: (1) patients underwent spine surgery under general anesthesia; (2) systemic intravenous administration of TXA was used during surgery as interventions; (3) results of the control group were reported; (4) outcomes including intraoperative, postoperative, and total blood loss (primary outcomes) were reported; (5) reported outcomes included the amount of allogenic blood transfusions (secondary outcome).

Articles were excluded for the reasons below: (1) oral, intramuscular, or topical TXA administration; (2) did not report appropriate outcomes or outcome measurements as mentioned; (3) underwent epidural anesthesia combined with general anesthesia; (4) non-randomized controlled trials; (5) non-human studies; (6) articles not in English.

Four authors (DL, KK, WYL, and SJC) reviewed the articles independently to assess the risk of bias using the “risk of bias” tool provided in the Review Manager ver. 5.3 software (The Cochrane Collaboration, UK) according to the Cochrane guidelines [16]. If necessary, a third-party review author (SL) was included to resolve disagreements. Quality assessments of the studies were based on Cochrane’s assessment of risk of bias. The following seven domains to assess the risk of bias were used in each trial: random sequence generation, allocation concealment, blinding of participants and personnel, blinding of outcome assessment, incomplete outcome data, selective reporting, and other bias. The methodology for each trial was graded as “high,” “low,” or “unclear,” to reflect a high risk of bias, low risk of bias, or uncertainty of bias, respectively.

The initial database search to find potentially relevant articles was completed by two authors (DL and KK). Three investigators (DL, KK, and WYL) performed the selection or exclusion of studies according to the criteria set in the protocol. Primary selection was performed based on article title and abstract. Secondary selection was done independently using the full-text. Studies for final assessment were selected by the agreement among three investigators. If needed, a third-party investigator (SL) took part and decided if the study should have been included.

The data were collected by three authors (DL, KK, and SJC) and rechecked by another investigator (SL) to ensure the accuracy. References and data for each included study were carefully cross-checked to ensure that no duplicate data were present and that integrity of the meta-analysis was maintained. In case of insufficient data, authors of each article were contacted and asked to provide information and clarification. If the data were available as a median and range (minimum–maximum), the conversion to mean and standard deviation (SD) was performed by the methods proposed by Hozo et al. [17].

The general characteristics, treatment or intervention types, and outcomes were collected for each study. General characteristics included study design, publication year, first author, and subject age. TXA dose regimen and timing, type of surgery, and anesthesia methods were recorded for the included studies. The measured outcomes included primary and secondary outcomes, as previously described. For assessing the reliability of agreement between four raters, we used Fleiss’ kappa analysis.

Meta-analysis on the data, including effects and outcomes of treatments, was performed using the Mantel–Haenszel method and Review Manager software (RevMan version 5.3, The Cochrane Collaboration, UK). For continuous data, such as blood loss and transfusion volume, mean ± SD was used to calculate the weighted mean difference and 95% confidence interval (CI). Heterogeneity was analyzed via the value of I² and the result of the chi-squared test. We regarded a level of 10% significance (P < 0.1) in the chi-square statistic or an I² > 50% for considerable heterogeneity and proceeded to sub-group analysis. If any heterogeneity was observed, the causes of heterogeneity were first analyzed and then subjected to sub-group treatment. We used R software (version 3.4.0, R Foundation for Statistical Computing, Austria) for analysis the publication bias and drawing the funnel plot.

In the initial database search, a total of 3,617 articles were identified from PubMed (n = 880), EMBASE (n = 2,393), and Cochrane Library (n = 344) (Fig. 1). After excluding 660 duplications, three authors independently reviewed and excluded 2,957 irrelevant articles based on the title of the article. We excluded 2,848 articles based on the exclusion criteria. The remaining 109 studies were evaluated for eligibility by abstract, and 74 were excluded for the following reasons: no randomized controlled trial (RCT) (n = 67); not spine surgery (n = 2); epidural anesthesia was combined with general anesthesia (n = 1); no appropriate placebo group (n = 2); and no TXA use as an intervention (n = 2). Finally, we conducted a full-text review of the remaining 35 studies and included 12 papers in the meta-analysis. A total of 23 studies were excluded because of grey literature (n = 23), non-RCT studies (n = 8), studies not using systemic TXA (n = 2), and documents in which the full text could not be obtained (n = 2).

Table 1 shows the characteristics of the 12 studies included in our final analysis. Every study was a randomized controlled trial. There was a total of 875 participants (436 in TXA group, 439 in control group) from these studies, ranging from 4 to 86 years old. The number of participants in each study ranged from 40 to 147. There were no significant differences in patients’ characteristics between the intervention and control groups, such as age, gender, height, weight, American Society of Anesthesiologists grading, and preoperative hemoglobin. Various types of surgeries were performed, including lumbar disc resection, idiopathic scoliosis surgery, posterior spinal fusion surgery, anterior cervical discectomy, and excision of spinal tumors after laminectomy. The extent of spinal surgeries varied, but no study differences were reported between groups in terms of the level of surgery. In all studies, general anesthesia was performed, but the anesthetic methods varied.

TXA was administered by intravenous (IV) in every study, although different dosages (ranging from 10–100 mg/kg) and delivery timings (ranging from bolus dose to 30-min infusion) were used. The loading dose of TXA was administered immediately after the anesthesia induction and maintenance infusion continued throughout the surgery in 10 studies. Maintenance infusion was continued until 5 h after operation in one study [18]. In another study [19], TXA was administered as the second bolus dose 3 h after the first loading dose.

The methodological quality of all included studies was estimated to have less bias-based error or uncertainty (Fig. 2).

Most studies reported detailed information regarding the randomization techniques that were used, such as manual random number selection or computer-generated random number table; Xu et al. [20] did not specify a randomization method. Allocation concealment was high in 4 studies [1,2,20,21] and low in the others. The studies by Taghaddomi et al. [22] and Xu et al. [20] were rated as unclear to blinding of the participants and the outcome assessor. The risk of incomplete outcome data was high in the study by Colomina et al. [21] because the dropout rate was over 20%, and the study by Xu et al. [20] had an unclear evaluation process. Some studies except Colomina et al. [21], Farrokhi et al. [23], Peters et al. [24], Raksakietisak et al. [19], Shi et al. [25], and Verma et al. [26] have no information about in the Web-based study registration site (e.g., www.clinicaltrials.gov). We could not evaluate the selective reporting bias of papers without study records, thus rated them as ‘unclear’. Five of the studies [18,20,22,24,25] did not have a detailed description about sample size calculation, indicating a potential bias. The methodological domain assessment for each study is shown in Fig. 3.

A total of 5 studies (n = 320; 160 TXA group and 160 control group) reported the total volume of transfused allogeneic blood cells. Perioperative IV TXA administration was shown to significantly reduce volume of transfused blood cells by a mean volume of 162.10 ml, ranging from 31.77 to 292.44 ml compared with the results of the control patients (95% CI; P = 0.010, I² = 47%) (Fig. 7).

Although most outcomes were favorable for the intervention, subgroup analysis was planned due to high heterogeneity. Subgroup analysis was performed on intraoperative bleeding (I² = 73%) which has high heterogeneity. The subgroup analysis was designed to follow domains based on heterogeneity data: anesthesia type (total intravenous anesthesia [TIVA] technique or inhalation anesthetics), IV TXA dosing regimen and timing, type and range of spine surgery, and transfusion triggers. However, due to insufficient data and ununified reporting criteria, subgroup analysis was conducted on the TXA dosing regimen. Low dose group (loading 10 mg/kg and maintenance 1 mg/kg/h) includes Neilipovitz et al. [1], Wong et al. [3], Farrokhi et al. [23], Peters et al. [24], and Verma et al. [26]. High dose group (higher dosage than low dose group) includes Sethna et al. [2], Elwatidy et al. [18], Raksakietisak et al. [19], Xu et al. [20], Colomina et al. [21], Taghaddomi et al. [22], and Shi et al. [25]. Subgroup analysis showed that high dose of intravenous TXA significantly reduces intraoperative blood loss by a mean volume of 185.32 ml, ranging from 128.95 to 241.70 ml (95% CI; P < 0.001, I² = 84%). Low dose of intravenous TXA reduces intraoperative blood loss by a mean volume of 189.69 ml, ranging from 53.68 to 410.76 ml (95% CI; P = 0.570; I² = 0%), but not significant results. The funnel plot analysis of publication bias was performed using the collected outcomes (Fig. 8). The asymmetrical property of the resulting plot shows the presence of publication bias.