This was a meta-analysis of randomized clinical trials (RCTs) conducted according to the Cochrane Handbook for Systematic Reviews of Interventions and the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) (15) and Assessing the methodological quality of systematic reviews (AMSTAR) guidelines (16). The Embase, PubMed, and Cochrane Library databases were searched with no language limitations from inception to October 23, 2021. The search strategy is described in detail in Supplementary Table 1. Following the preliminary screening of titles and abstracts, two independent reviewers assessed related publications. The protocol of this study was published and registered in PROSPERO (CRD42021284819).

The studies were screened according to the PICOS (population, intervention, comparison, outcome, study design) criteria. The selection criteria are detailed in Supplementary Table 2.

We extracted data from the included articles, including investigator characteristics, surgical methods, participant characteristics, and main results. Two authors independently worked on this section. The primary outcomes were postoperative success rates, postoperative complication rates, and postoperative reoperation rates. The secondary outcomes included postoperative work status, arm and neck pain scores, the neck disability index (NDI), and surgery time.

The Cochrane Collaboration risk-of-bias assessment tool (17) was used by two reviewers to evaluate the included studies for potential bias independently. Disagreements between the two investigators were resolved by bringing in a third investigator. The overall risk of bias is calculated and classified as “high risk,” “low risk,” or “unclear risk.” The tool to assess the risk of bias has been described in detail in Supplementary Table 3.

Firstly, a random-effects model was used for pairwise analysis to pool relative risks (RRs) or mean difference (MD) and 95% confidence intervals (CIs) (18). P < 0.05 was considered significant. Forest plots and I² were used to explore sources of heterogeneity (19). Secondly, the network geometry was generated using Stata version 16.0 (Stata Corp). Then a Bayesian network meta-analysis was performed using Markov chain Monte Carlo methods in WinBUGS version 1.4.3 (MRC Biostatistics Unit, Cambridge, United Kingdom) (20) using a random-effects consistency model (21). Each surgical intervention's safest and most effective probability was ranked first, followed by second, third, etc., based on the average difference and the risk ratio. As the stability of the results is crucial for network meta-analysis, we used various methods to assess the inconsistency of the results.

The consistency and inconsistency models are compared, and the inconsistency is initially estimated roughly. The entire network on detailed comparisons (nodes) was tested by node splitting analysis; P < 0.05 manifested a significant inconsistency. The indirect results (network meta-analysis results) were then compared with the pairwise direct results (meta-analysis results) to determine the source of the inconsistency. The intervertebral spacer was used to conduct a sensitivity analysis of anterior cervical discectomy and fusion (ACDF) (Zero-P and the other).

The flow of the selection process and the reasons for exclusion are depicted in Figure 1. These electronic searches yielded 861 potentially relevant studies, of which 67 potentially relevant articles were thoroughly evaluated. Finally, 23 trials (26 records) including 1,844 participants were included in the final analysis (22–44). Ten interventions were performed that had anterior cervical discectomy with autologous bone graft (ABG), anterior cervical discectomy with allograft bone graft plus plating (ABGP), anterior cervical discectomy (ACD), ACDF, anterior cervical discectomy with fusion and additional plating (ACDFP), anterior cervical foraminotomy (ACF), cervical disc replacement (CDR), posterior cervical foraminotomy (PCF), anterior cervical discectomy with polymethylmethacrylate (PMMA) and physiotherapy. Most of these studies (69.6%) were conducted in Europe. The characteristics of the included trials and participants are shown in Supplementary Table 4. Two studies showed high risk for generating the randomization sequence (30, 34). Two studies showed high risk in concealing allocation (30, 34). Ten studies showed high risk in blinding of participants and personnel (26, 29, 30, 32, 34–36, 41). Two studies showed high risk in blinding of outcome assessment (37, 40). One study showed high risk in incomplete outcome data (32). Four studies showed high risk in selective outcome reporting (26, 27, 29, 37). Supplementary Figures 1, 2 summarize the risk of bias assessment.

Eight RCTs with a sum of 562 participants compared the differences in arm pain scores under different interventions (Supplementary Figure 29) (23, 27, 29, 38, 42–44). Eight RCTs (627 participants) compared the differences in neck pain scores under different interventions (Supplementary Figure 30) (23, 27, 35, 38, 40, 42–44). No statistical differences were found in scores for arm and neck pain of different interventions (Figure 7A). The results obtained using the consistency model fit well-with the inconsistency model; Node splitting analyzes did not show significant inconsistency (all P > 0.05; Supplementary Tables 13, 15). Direct results were detailed in Supplementary Figures 31–36. Supplementary Figure 37 shows the direct and indirect results of comparing different interventions. The direct results were prominently consistent with the corresponding indirect results insignificance and tendency. Figure 7B showed the distribution of arm and neck pain probability scores for each intervention arranged in seven possible positions. Scores for arm pain ranging from low to high were as follows: ABGP (probability 71%), ACDFP (15%), ABG (5%), Physiotherapy (5%), CDR (3%), ACD (1%), and ACDF (0%). The probabilities are detailed in Supplementary Table 14. The scores for neck pain ranging from low to high was as follows: ABG (probability 46%), ABGP (45%), ACDFP (4%), ACD (3%), ACDF (1%), CDR (1%), and physiotherapy (1%). The probabilities are detailed in Supplementary Table 16.

Six RCTs (575 participants) compared the differences in neck disability index under different interventions (Supplementary Figure 38) (32, 35, 38, 42–44). No statistical differences were found in the neck disability index of different interventions (Figure 8A). The results obtained using the consistency model fit well-with the inconsistency model. The direct results were detailed in Supplementary Figures 39–41. Supplementary Figure 42 showed the direct and indirect results of comparing different interventions. The direct results were prominently consistent with the corresponding indirect results insignificance and tendency. Figure 8B showed the distribution of the probability of neck disability index for each intervention organized in six possible positions. The neck disability index ranging from low to high was as follows: ABG (probability 71%), ABGP (19%), physiotherapy (7%), ACD (2%), ACDF (1%), and CDR (1%). The probabilities are detailed in Supplementary Table 17.

Nine RCTs with 820 participants compared the differences in surgery time under different interventions (Supplementary Figure 43) (26, 31, 32, 35, 36, 39–41, 44). In terms of surgery time (Figure 8A), ABGP (MD, −68.02; [95% CI, −93.46 to −42.60]), ACD (MD, −22.07; [95% CI, −39.21 to −5.85]), ACDF (MD, −22.55; [95% CI, −34.81 to −9.98]), and ACDFP (MD, −23.34; [95% CI, −42.31 to −2.37]) were with shorter surgery time compared with ABG in the consistency model. ACD (MD, 46.07; [95% CI, 19.12 to 71.53]), ACDF (MD, 45.47; [95% CI, 23.15 to 67.67]), ACDFP (MD, 44.57; [95% CI, 18.69 to 72.78]), CDR (MD, 65.50; [95% CI, 38.50 to 92.42]), and PMMA (MD, 59.43; [95% CI, 33.78 to 84.30]) had a longer surgery time compared with ABGP in the consistency model. ACD (MD, −19.36; [95% CI, −38.28 to −0.91]) and ACDF (MD, −19.88; [95% CI, −35.99 to −3.74]) were with shorter surgery time compared to CDR in the consistency model. Furthermore, ACDF (MD, −13.85; [95% CI, −26.11 to −1.24]) were with shorter surgery time compared with PMMA in the consistency model. The results obtained using the consistency model fit well-with the results using the inconsistency model; Node splitting analyzes did not show significant inconsistency (all P > 0.05; Supplementary Table 18). Direct results were detailed in Supplementary Figures 44–49. Supplementary Figure 42 showed the direct and indirect results of the comparison of different interventions. The direct results were prominently consistent with the corresponding indirect results in significance and tendency. Figure 8B showed the distribution of the probability of surgery time for each intervention organized into seven possible positions. Among the seven surgical interventions, ABGP had the shortest surgery time. The probabilities are detailed in the Supplementary Table 19.

This study is the first network meta-analysis that provides an evidence-based comparative evaluation of all surgical interventions for cervical radiculopathy. We have used an innovative method of comparing indirect results (network meta-analysis results) and pairwise direct results (meta-analysis results) to investigate the source of heterogeneity. Our research does, however, have limitations. The sample sizes in the studies were insufficient, which reduced the reliability of the results. The prognostic indicators were reported at various time points, resulting in heterogeneity. Furthermore, the surgical level of different surgeons may contribute to heterogeneity. In addition, although all RCTs included patients with pure cervical radiculopathy, most of the included studies did not report the localization of the degenerative disease (e.g., central, paracentral, foraminal). This is an important factor in the surgeon's decision-making process, as some surgical techniques have specific contraindications.