The literature search was conducted to identify eligible studies for this systematic review and meta-analysis. Two researchers (S.P. and B.H.) independently searched the following electronic databases: Medline, EMBASE, Cochrane Controlled Trial Register, Web of Science, and Google Scholar for relevant studies published in English. Articles published between September 2005 and December 2020 in peer-reviewed journals were included. The primary search was conducted on January 28, 2021; however, an additional search was conducted on February 28, 2021 to include more recent studies. In addition, the studies referenced in the selected articles were searched manually.

The search strategy was as follows: (“Video assisted thoracoscopy surgery” or VATS) and [(“Thoracic paravertebral block” or TPVB) or (“paravertebral block” or PVB) or (“Serratus plane block” or “Serratus anterior plane block” or “Serratus interfascial plane block” or SPB) or (“Erector spinae plane block” or ESPB) or (“Intercostal nerve block” or INB)].

Studies were considered eligible if they were randomized controlled trials (RCTs) published in English that reported postoperative pain scores or total postoperative opioid consumption in both the experimental and control groups as outcomes. Non-RCTs (quasi-experimental designs), abstracts, conference proceedings, unpublished gray literature, and review studies were excluded. Among the regional analgesia techniques, continuous blocks via catheterization were also excluded.

We performed six steps to select the studies. First, two researchers (S.P. and B.H.) imported the titles and abstracts of the articles identified in the searches into reference management software (EndNote 20, Clarivate^TM) and performed a preliminary review. Second, duplicate papers were identified and eliminated using the reference management software. Third, two researchers (S.P. and B.H.) independently reviewed the imported studies. We excluded all of the imported studies that did not clearly meet the inclusion criteria (due to the study design, participants, types of intervention, and comparison groups). Fourth, they also independently screened the titles, abstracts, and methodology sections of the studies that appeared to meet the inclusion criteria. Fifth, we retrieved the full texts of the papers that met all the inclusion criteria for data extraction and linked multiple reports of the same study. Finally, the studies included in the final selection were confirmed and coded for analysis by two researchers (B.H. and Y.J.). These coding sheets were independently checked for accuracy by researchers who were not involved in the review process. If there were any differences between the codes provided by the two reviewers, the discrepancies were resolved by consulting a third independent reviewer (C.O.).

The information from the included articles was independently extracted by two reviewers (B.H. and Y.J.), and each selected article was reviewed twice by both reviewers. To determine the outcomes of individual studies, pain scores and opioid consumption were determined for each group and recorded as the means and standard deviations (SDs). Medians and interquartile ranges (IQRs), as approximations of the mean and SDs, were determined using the estimation method proposed by Wan et al. [18]. When outcome data were available only as graph, a virtual ruler was used to extract the value by matching the interval between the basic unit of the plot and the ruler. The effect sizes and standard errors were calculated. Additional data, including the location, sample size, characteristics of individual study populations, and the intervention design, were extracted using a predetermined data extraction table.

The primary outcome was cumulative opioid consumption during the first 24 h postoperative period. All opioids were converted to equianalgesic intravenous (IV) morphine doses (IV morphine 1 mg = IV fentanyl 10 µg = IV sufentanil 2 µg = IV tramadol 10 mg). The secondary outcome was pain scores assessed during three different periods in the first 24 h, namely, the early (0–6 h), middle (6–18 h), and late (18–24 h) periods. For studies that included several time points within each time period, pain scores close to 1 h for early, close to 12 h for middle, and close to 24 h for late were used. In the one study that timetable was expressed as an interval (ex. 6 a.m. to 2 p.m.), a similar period expected to include the interval was used. Pain scores that were assessed using visual analogue scales (VAS) were converted to a 0–10 analogue scale to allow for statistical evaluation.

A random-effects NMA within a frequentist framework was performed using the R software version 4.0.3 (R Foundation for Statistical Computing, Austria) and the “netmeta” package for frequentist NMA [19,20]. A network plot was constructed to evaluate the direct and indirect comparisons of network structures, including studies. Heterogeneity was evaluated using I² statistics. The Q statistic based on the full design-by-treatment interaction random-effects model was calculated to evaluate the global inconsistency [21]. We also evaluated the local inconsistencies between the direct and indirect effects using the net splitting technique. If the P value of the net splitting was < 0.05, we presumed there was a significant disagreement (inconsistency) between the direct and indirect estimates. We visualized the net split results using forest plots and direct evidence plots, which showed the percentage of direct and indirect evidence used for each estimated comparison. A mean path length > 2 indicated that the comparison estimate should be interpreted with caution. Additionally, a net heat plot was constructed to determine the importance of each comparison and the inconsistency of the design. Network league tables and forest plots were produced to show details of the results of the comparisons between the interventions. Outcomes are presented as mean differences with a 95% CI. To rank the analgesic interventions in order, we reported the P score, which measures the level of certainty that an intervention is better than the competing interventions [22]. In this study, the P score ranged from 0 to 1, with 1 indicating that the treatment option was statistically best and 0, the worst. Potential publication bias was assessed using comparison-adjusted funnel plots and Egger’s test. The confidence for every outcome was rated according to the grading of recommendations assessment, development, and evaluation (GRADE) system with the support of the CINeMA (Confidence in Network Meta-Analysis, https://cinema.ispm.unibe.ch/) web application (Institute of Social and Preventative Medicine, University of Bern, Switzerland) [23]. This is based on a methodological framework that considers six domains: within-study bias, reporting bias, indirectness, imprecision, heterogeneity, and incoherence [24]. The minimal clinically important difference was set at 1 out of 10 for postoperative pain and 10 mg for IV morphine-equivalent consumption.

The literature screening process and results are shown in Fig. 1. The screening sequence of the PRISMA 2009 flow diagram identified 21 studies that compared the analgesic efficacy of TPVB, ESPB, SPB, intercostal nerve block (INB) and against control (no block) [25–45] with a total of 1,391 patients included. Table 1 shows the characteristics of the included studies. Table 2 shows the number of included studies and enrolled patients sorted by outcomes.

Individual studies were assessed using the Cochrane Collaboration’s Risk of Bias tool [46] and ranked according to a low/high/unclear grading scale (Fig. 2). The overall quality of the 21 included studies was moderate. Some of the studies showed possible patient selection bias and bias in methodology, with 70% showing an unclear or high risk of bias in performance concealment, 25% in blinding of participants and personnel, and 55% in blinding of the outcome assessment. A comparison-adjusted funnel plot showed evidence of a visually symmetric plot of opioid consumption and pain scores during the three time periods. The results of Egger’s regression test of outcomes also showed no significant publication bias (P > 0.05) (page 14 of each Supplementary Materials 1 to 4). The quality of evidence was rated as very-low-to-low in nature according to the GRADE system (Table 2), and the confidence rating of each comparison using CINeMA is described in the supplement file (Supplementary Material 5).

The results of the I² and Q statistics (based on the full design-by-treatment interaction random-effects model) indicated that a random-effects model may be suitable for revealing any inconsistency or heterogeneity in our network model (Table 2). Additionally, according to the colored background of the net heat plot, the random-effects model appeared to be suitable for our data (pages 10 and 11 of each Supplementary Materials 1 to 4). A direct evidence plot (page 5 to 6 of each Supplementary Materials 1 to 4) and forest plot of the net splitting results (page 11 to 12 of each supplementary files, S1 to S4) were used to evaluate local inconsistency.

Of the included studies, 17 [25–28,30,31,33,35,37–45] RCTs reported opioid consumption and 18 [25,26,28,30–36,38–45], 16 [25,26,28–35,38–44], and 17 [25–27,29,30,33–36,38–45] RCTs reported pain scores for each of the three postoperative time periods (early, middle, and late, respectively). The networks for the TPVB and control were greater than the networks for other blocks, followed by the ESPB and control. As shown in Fig. 3, TPVB had the best analgesic effect on opioid consumption compared with the control (mean difference [MD]: −13.2 mg, 95% CI [−16.2, −10.1]), followed by INB (MD: −9.55 mg, 95% CI [−13.2, −5.9]), ESPB (MD: −8.7 mg, 95% CI [−11.4, −6.1]), and SPB (MD: −5.9 mg, 95% CI [−9.4, −2.5]). In terms of pain scores in the early period, ESPB had the greatest effect compared with the control (MD: −1.6, 95% CI [−2.3, −0.9]), followed by TPVB, INB and SPB. In the middle and late periods, TPVB, ESPB and INB showed superior analgesic effects in reducing pain scores compared to the control, whereas SPB did not have a significant effect. The local inconsistency in ESPB and control was significant in the early and middle periods (Table 2). The two studies by Ciftci et al. [34,35] had effect sizes that tended to be higher than those measured in the other studies. Table 3 shows the network league table that displays the direct comparison and full model results separately.

Table 4 shows the P-scores of analgesic efficacy and the ranking of the five groups. TPVB ranked highest for opioid consumption in the first 24 h (0.996) and for middle- and late-period pain scores (0.793 and 0.831, respectively). However, ESPB ranked first for pain scores in the early period (0.792). INB ranked second for opioid consumption and third for pain scores in all three periods. SPB ranked fourth for all the outcomes.

During proofreading, we noticed that the article which included our analysis had been retracted from its journal [37]. Thus, we performed an additional analysis after removing the retracted article. That retracted paper only had information on the 24 hours opioid consumption, and there was no significant change in effect estimate of opioid consumption, P-scores nor ranking of the Included blocks (Supplementary Material 6).