Comparison of therapies in lumbar degenerative disc disease: a network meta-analysis of randomized controlled trials

A comprehensive strategy was developed to search the PubMed and Cochrane Central Register of Controlled Trials (CENTRAL) databases to identify relevant RCTs published until December 2015 (Supplementary Appendix 1). Bibliographic and grey literature searches were also conducted. Only English language articles were reviewed.

RCTs that included patients with discogenic low back pain due to single-level lumbar DDD, who were unresponsive to conservative therapy, were considered if they compared a TDR device (Charité, ProDisc-L, Maverick, Kineflex-L, Flexicore, activL) with other TDR devices, fusion (anterior, posterior, or circumferential) or conservative care (rehabilitation, exercise). Studies must have reported at least one of the following outcomes at 2-year follow-up: Oswestry Disability Index (ODI) success (≥15-point improvement), mean back pain score (Visual Analog Scale [VAS] or Numeric Rating Scale [NRS]), patient satisfaction, employment status, reoperations (device failures resulting in removal, revision, reoperation or supplemental fixation) and device-related serious adverse events (SAEs). Studies were excluded if they had >50% loss-to-follow-up, were single-arm or observational studies, reported outcomes for a subset of the population captured in the full study (i.e., double reporting) or presented data for treatment arms in an aggregated form. Studies were selected by two independent reviewers, with differences resolved by discussion resulting in consensus or by a third party.

Data on study and patient baseline characteristics, treatments, loss to follow-up and key outcomes pertinent to this analysis were extracted by a single reviewer and checked for accuracy by a second reviewer; disagreements were resolved by consensus or by a third party.

Some data were extracted by digitizing data (TechDig v2.0 digitizing software) and mean differences were calculated from baseline and follow-up data provided. For continuous outcomes, a normal distribution was assumed. No data imputation was required.

The Cochrane Collaboration's tool for assessing risk of bias was used to evaluate study quality. The tool assesses the risks of selection bias (sequence generation, allocation concealment), performance bias (blinding of participants and personnel), detection bias (blinding of outcome assessment), attrition bias (incomplete outcome data), reporting bias (selective outcome reporting) and other bias [26]. Two independent reviewers scored studies for low, unclear or high risk of bias, with discrepancies resolved by a third party.

An NMA permits comparison of treatments across a network of trials, where different treatment strategies are employed within the same or very similar patient population. The NMA approach allows any two treatments within the network of evidence to be compared, even when a direct comparison from a trial is not available. Each treatment was considered as an individual node in the network, with networks constructed using NetMetaXL (Figure 1) [27] and analyses performed using WinBUGS Software (version 1.4.3, MRC Biostatistics Unit, Cambridge, UK). For outcomes such as device-related SAEs where a network could not be constructed because of insufficient data, an NMA was not conducted. Analyses were conducted using Bayesian fixed effects models that were fit to the data with a binomial likelihood. The models and priors used were consistent with recommendations made by the NICE Decision Support Unit [28]. Results of pairwise comparisons were reported as odds ratios (ORs) and mean differences (MDs). ORs were summarized using the median and 95% credible interval (CrI). ORs less than or greater than 1 favor one of the compared lumbar DDD therapies over the other, whereas ORs equal to 1 indicate equivalency. Mean differences were summarized using the mean and 95% CrI; MDs less than or greater than 0 favor one of the compared lumbar DDD treatments over the other, whereas MDs equal to 0 indicate equivalency. In brief, CrIs can be interpreted as the Bayesian equivalent to confidence intervals. Credible intervals excluding 1 for ORs or 0 for MDs indicate statistical significance. The probability that each treatment was the most efficacious was calculated by counting the number of times each treatment had the highest OR or MD in the model [29]. A single numerical summary of the probability that a treatment is best, second best and so on, known as the surface under the cumulative ranking curve area (SUCRA), was also calculated for each treatment. More detailed information regarding the statistical methods is included in Supplementary Appendix 3.

Because an NMA requires that studies are sufficiently similar in order to pool results, available study and patient characteristics were assessed to ensure similarity and to investigate the potential effect of clinical heterogeneity on ORs or MDs and 95% CrIs. Another key assumption behind NMA is that the analyzed network is consistent; that is, there is no conflict between direct and indirect evidence within a closed loop in the network, where the closed loop is comprised of different studies [30]. Evidence networks in our NMA consisted of single-study connections and contained a single closed loop based on data from 1 trial; therefore, we were unable to perform an inconsistency analysis [30]. Additionally, the results from our NMA were qualitatively compared with the results from direct comparisons generated from traditional meta-analyses.

Sensitivity analyses were conducted for the following: variations in efficacy of anterior lumbar interbody fusion (ALIF) instrumentation [31], where efficacy was adjusted using a simulated trial of 100 patients to consider assumptions regarding the difference between using recombinant human bone morphogenetic protein 2 (rhBMP-2) and iliac crest bone graft in ALIF (i.e., 10% increase in efficacy in favor of rhBMP-2) [14,17]; outcome definitions, where studies that used different definitions or measures for outcomes of interest were excluded [15,17]; follow-up rate, where studies with a ≥20% loss to follow-up were excluded [12]; exclusion of studies that included patients with 1 or 2 affected spinal segments [12]; and exclusion of studies with a high risk of bias [12,17].

From the literature search, 2458 potentially relevant records were identified, of which 2278 records were excluded during title and abstract review. Of the 180 eligible, full-text articles assessed for inclusion, 6 publications were included in the analysis [12,14,15,17,23,24] (Figure 1). Three trials compared a TDR device with fusion [14,15,17], two compared a TDR device with another TDR device [23,24] and one compared a TDR device with conservative care [12] (Table 1). Only one study included three treatment arms [23]. Charité and ProDisc-L were included in three trials each (each as an investigational device in one and as a control device in two others), whereas Maverick, Kineflex-L and activL were included in one trial each. ALIF was included as a comparator in two trials, circumferential fusion in one trial and rehabilitation in one trial. Five studies, which were corporately sponsored, were regulated under the US FDA Investigational Device Exemption (IDE) program; one study was not involved in the FDA's IDE program and was not corporately sponsored.

Sample sizes for included trials ranged from 173 to 577 participants (Table 1). Follow-up for all studies was 2 years; 1.8–20% loss to follow-up was observed at 2 years. Only one study included patients with lumbar DDD at 1 and 2 levels [12]. Most patients were treated at the L5/S1 level. In all studies, patients were unresponsive or had insufficient improvement to non-surgical therapy. At baseline, the ODI scores were 42–63% and mean back pain scores were 69–80 across studies.

The risk of bias was similar across studies (Supplementary Appendix 2). Methods for randomized sequence generation were reported in all studies: all studies used block randomization, with three studies reporting use of block sizes of six [14,15,17]. Among patients randomized to the control arm in one study, surgeons were given the choice of implanting ProDisc-L or Charité [23]. Central allocation was reported as the method used in five studies [12,14,15,17,24], with sites notified of allocation using sealed envelopes in three studies [14,17,24], by telephone in one study [15] and through website access in one study [12]. Surgeons and/or staff were not blinded for preparatory purposes in three studies [14,15,23], were blinded until informed consent was received in one study [17] and were blinded until shortly after randomization in one study [12]. Participants were blinded to their randomization group until after surgery in four studies [14,15,23,24]. Intention-to-treat (ITT) analyses were conducted in five studies [12,14,15,23,24]. Analyses in one study were not conducted by ITT, however, this study was deemed to have low risk for selection bias because analyses were approached conservatively: the one study participant who was randomized to receive the investigational device received the control treatment and had successful outcomes postoperatively; therefore, this patient was included within the control group [17].

All studies were included in the evidence network for ODI success, with all eight treatment comparators considered (Figure 2A). Outcomes data from the individual studies are presented in Supplementary Appendix 4A. When compared with all treatment comparators, activL was associated with the highest ORs of ODI success (Supplementary Appendix 5A). The ORs of achieving ODI success were significantly in favor of activL when compared with circumferential fusion (OR: 2.58; 95% CrI: 1.13–5.83), ALIF (OR: 2.57; 95% CrI: 1.08–6.05) and rehabilitation (OR: 3.87; 95% CrI: 1.64–9.13) (Figure 3A, Supplementary Appendix 5A). Overall, activL had the highest probability of being the best treatment among all therapies (72%; Figure 4) and had the highest overall SUCRA (94%; Supplementary Appendix 6).

The OR estimates for ODI success from direct comparisons aligned well with the estimates obtained from the NMA (Supplementary Appendix 8). Model fit statistics indicated reasonable model fit.

The analysis for back pain score included all studies and comparators (Figure 2B). Outcomes data from the individual studies are presented in Supplementary Appendix 4B. When compared with all treatment comparators, activL showed the greatest mean difference from baseline in back pain score (Supplementary Appendix 5B). The mean change for back pain score, from baseline to 2 years, was significantly greater for activL than for Charité (MD -10.42; 95% CrI: -20.07, -0.82), Kineflex-L (MD -11.60; 95% CrI: -22.98, -0.33), and ALIF (MD -16.84; 95% CrI: -29.22, -4.39) (Figure 3B, Supplementary Appendix 5B). Overall, activL had the highest probability of being the best treatment among all therapies (probability best, 61%; Figure 4) and had the highest SUCRA (91%; Supplementary Appendix 6).

The MD estimates for back pain score from direct comparisons aligned well with the estimates obtained from the NMA (Supplementary Appendix 8). Model fit statistics indicated reasonable model fit.

For patient satisfaction, the evidence network included five studies and seven comparators (Figure 2D). Outcomes data from the individual studies are presented in Supplementary Appendix 4A. Overall, activL had the highest ORs for patient satisfaction from all comparisons (Supplementary Appendix 5C). The ORs for patient satisfaction significantly favored activL compared with rehabilitation (OR 3.30; 95% CrI: 1.39–7.84) and ALIF (OR 3.75; 95% CrI: 1.56–8.77) (Figure 3C, Supplementary Appendix 5C). Overall, activL had the highest probability of being the best treatment among all therapies (probability best, 46%; Figure 4) and had the highest SUCRA (86%; Supplementary Appendix 6).

The OR estimates for patient satisfaction from direct comparisons aligned well with the estimates obtained from the NMA (Supplementary Appendix 8). Model fit statistics indicated reasonable model fit.

Four studies were included in the evidence network for employment status and seven comparators were considered (Figure 2D). Outcomes data from the individual studies are presented in Supplementary Appendix 4A. When compared with patients who received other treatments, activL patients had the highest ORs of being employed (Supplementary Appendix 5D). Overall, there were no statistically significant differences between the comparators (Figure 3D; Supplementary Appendix 5D).

The OR estimates for employment status from direct comparisons aligned well with the estimates obtained from the NMA (Supplementary Appendix 8). Model fit statistics indicated reasonable model fit.

The evidence network for reoperation, defined as device failures requiring revision, removal, reoperation, or supplemental fixation, included all studies but one [12] (Figure 2E); seven treatment comparators were considered. Outcomes data from the individual studies are presented in Supplementary Appendix 4A. Overall, ORs for reoperation were in favor of Kineflex-L (Supplementary Appendix 5E). However, there were no statistically significant differences between comparators.

The OR estimates for reoperations from direct comparisons aligned well with the estimates obtained from the NMA (Supplementary Appendix 8). Model fit statistics indicated reasonable model fit.

Results of the primary analysis were robust for the variables tested in the sensitivity analyses (Supplementary Appendix 9). For ODI success, when adjusted for differences in efficacy associated with ALIF based on the technique used, results for ODI success were slightly sensitive to the assumption of better efficacy with rhBMP-2 than with iliac crest bone graft. Results remained in favor of activL when the analysis excluded studies using a modified version of the ODI questionnaire, studies with ≥20% loss to follow-up, studies that included 1- and 2-level surgeries and studies with a high risk of bias. Results for back pain score were sensitive to studies with a high loss to follow-up and studies that included 1- and 2-level surgeries, with MDs significantly in favor of activL. Back pain score results were robust to studies using the NRS pain scale and those with a high risk of bias. Patient satisfaction results were robust to excluding studies with ≥20% loss to follow-up and studies with a high risk of bias. Results from the primary analysis for employment status were robust to exclusion of studies with ≥20% loss to follow-up and those that included 1- and 2-level surgeries. Excluding studies with a high risk of bias resulted in numerically higher ORs for employment status that continued to favor activL; however, statistical significance was not reached. Results for reoperations were slightly sensitive to the use of the FDA's criteria for reoperation success in the Kineflex-L IDE study and to the exclusion of studies with a high risk of bias.