Network meta-analysis comparing neoadjuvant chemoradiation, neoadjuvant chemotherapy and upfront surgery in patients with resectable, borderline resectable, and locally advanced pancreatic ductal adenocarcinoma

The Bayesian analysis was conducted and reported in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines. We searched PubMed, Embase (Ovid), Cochrane Library, the American Society of Clinical Oncology and ClinicalTrials.gov database using “neoadjuvant therapies of pancreatic ductal adenocarcinoma” as part of the titles and abstracts with English restrictions. A time-frame from January 1, 2009, which was the date of introducing local anatomic subcategories of pancreatic ductal adenocarcinoma, i.e., RPC, BRPC and LAPC [4, 5], to December 16, 2018, was selected for the database search. In addition, we manually searched related reviews and bibliographies of included trials for additional references.

References were included if: 1. local PDAC; 2. compare two of the three treatment strategies (neoadjuvant chemoradiation, neoadjuvant chemotherapy, upfront surgery) with each other; 3. report enough information to calculate hazard ratios (HRs); 4. unrestricted age, gender, performance status (PS), ethnicity and country.

The references were excluded according to the following criteria: 1. neoadjuvant targeted therapy; 2. neoadjuvant chemoradiation and chemotherapy mixed; 3. posters and abstracts; 4. single-arm studies.

The search strategy in strict accordance with Population Intervention Comparison Outcomes Study (PICOS) design framework included the following domains of Medical Subject Heading (MeSH) terms: ‘Pancreatic Neoplasms’ and ‘Neoadjuvant Therapy’; MeSH and Subheadings were combined with ‘AND’ or ‘OR’.

First, the titles and abstracts of articles were screened. Review articles, case series, case reports, guidelines and conference abstracts were excluded from our study; and then full-text articles that met the inclusion criteria were thoroughly reviewed. Two investigators (Qc H, D W) reviewed the full manuscripts independently and extracted information including patient characteristics, period and type of study, treatment protocols, the sample size and outcomes (median OS, hazard ratio, 95% confidence interval and R0 resection rate) into the electronic database. Disagreements in study and data selection among investigators were resolved by discussion and consensus. The quality and risk of bias of randomized controlled trials (RCTs) were assessed by using Cochrane Collaboration’s tool [12], and the other trials were assessed by Risk if Bias in Non-randomized Studies of Interventions (ROBINS-I) [13].

The outcomes we analysed were median OS and R0 resection rate. R0 was defined as margin negative if tumour cells were present > 1 mm from the any surface. R1 was defined as margin positive if tumour cells were present within 1 mm from the any surface [14, 15]. Results on OS in the Bayesian analysis were expressed as a hazard ratio (HR) with 95% confidence intervals (CI). P < 0.05 was considered as significant level. Heterogeneity was assessed with the I² statistic. I² values less than 25% and greater than 50% were regarded as indicating low and high heterogeneity, respectively [16]. When HRs were not reported, we made estimations from summary statistics with the method described by Tierney et al. in 2007 [17], and Kaplan-Meier curves were digitized using Getdata Graph Digitizer 2.26 (http://​www.​getdata-graph-digitizer.​com). We applied the traditional pairwise meta-analysis between direct comparisons with Stata13 (StataCorp, College Station, TX, USA). The network meta-analysis was conducted with GeMTC version 0.14.3 (http://​drugis.​org/​software/​addis1/​gemtc) and WinBUGS version1.4.3 (MRC Biostatistics Unit, Cambridge, UK) by fixed-effect models. For R0 resection rate, we used GeMTC for network meta-analysis. Parameters for the GeMTC software were selected as: number of chains, 4; tuning iterations, 20,000; simulation iterations, 50,000; thinning interval, 10; inference samples, 10,000; and variance scaling factor, 2.5. Besides, we chose 5000 burn-ins and a thinning interval of 1 for each chain for HRs. The consistency model would be used when there was no significant inconsistency; otherwise, the inconsistency model was applied. We assessed the convergence of the model using the potential scale reduction factor (PSRF) of the Brooks–Gelman–Rubin method [18]; PSRF closer to 1 indicated the better convergence.

We identified 1089 studies from the title and abstract review to start with (Fig. 1). After initial screening, we retrieved the full text of potentially eligible articles for further detailed assessment. With the predeveloped search strategy, 14 eligible publications including three randomized controlled trials were included for meta-analysis, with a total of 1056 patients received at least one of the three treatment strategies (Table 1) [19‐32]. The eligible studies were published during 2010 to 2018. All studies included in our Bayesian analysis have been published as full manuscripts.

There were five head-to-head studies which compared the OS between neoadjuvant chemoradiation and neoadjuvant chemotherapy. Within the five head-to-head studies, only three of them reported the R0 resection rate. In our pairwise meta-analysis, there was no significant difference between neoadjuvant chemoradiation and neoadjuvant chemotherapy in R0 resection rate (OR 1.02, 95%CI 0.45–2.33, I² = 34.6%, Fig. 2a). Neoadjuvant chemoradiation showed superior OS significantly when compared with neoadjuvant chemotherapy (HR 0.8, 95% CI 0.60–0.99, p < 0.001, Fig. 2b).

The network was designed for three treatment comparisons of neoadjuvant chemoradiation, neoadjuvant chemotherapy and upfront surgery. All the potential scale reduction factors (PSRFs) were in the range of 1.00 to 1.01, so the model was proven convergent and stable; the consistency model was adopted. According to the established network, based on the consistency model, neoadjuvant chemoradiation showed significantly higher R0 resection rate over upfront surgery (HR 0.15, 95% CrI 0.02–0.56), whereas neoadjuvant chemotherapy did not provide better efficacy in R0 resection over upfront surgery (HR 0.42, 95% CrI 0.02–4.41) (Fig. 3a). Compared with results from the traditional pairwise meta-analysis, neoadjuvant chemotherapy showed a superior advantage for OS over neoadjuvant chemoradiation (HR 0.72, 95% CrI 0.58–0.88) or upfront surgery (HR 0.85, 95% CrI 0.72–0.99) in network meta-analysis (Fig. 3b).

The rank probabilities, calculated by the network consistency model, ranked the probabilities of rank order of treatment regimen overall outcomes evaluated (neoadjuvant chemoradiation, neoadjuvant chemotherapy and upfront surgery). Regimens with a higher value in the histogram were associated with higher probabilities for better treatment outcomes. The probability distribution of each regimen which was ranked at each of the possible positions was showed by the histogram (Fig. 4a-b). For the R0 resection rate, neoadjuvant chemoradiation had the highest probability of ranking one compared with neoadjuvant chemotherapy or upfront surgery (Fig. 4a, 79% vs 21% vs 0%). For OS, neoadjuvant chemotherapy had the highest probability of ranking one compared with neoadjuvant chemoradiation or upfront surgery (Fig. 4b). Neoadjuvant chemoradiation ranked as highest (98% vs 0% vs 2%) among all the regimens in all R0 resection rates measured without improving the likelihood of OS in local PDAC.

The comparison of postoperative complications among different treatment strategies was carried out. None of the five studies provided data on postoperative complications. The rankings of the three competitive treatment strategies were summarised in terms of postoperative complications (Fig. 5). Regimens with a higher value in the histogram were associated with higher probabilities for worse treatment outcomes. Neoadjuvant chemotherapy was associated with higher rates of postoperative complications (rank worst: 84%), followed by neoadjuvant chemoradiotherapy (13%) and upfront surgery (3%).