The current network meta-analysis was performed following the Cochrane handbook for reviewers of systematic reviews (Higgins and Green, 2011), and the pooled results were reported according to the preferred reporting items for systematic reviews and meta-analysis (PRISMA) for network meta-analysis (PRISMA-NMA) (Hutton et al., 2015; Page et al., 2021). This network meta-analysis was exempted from ethical approval and informed consent because it was undertaken with previously published data.

We defined inclusion criteria following previous meta-analyses (Shah et al., 2020; Quinzi et al., 2021; Wadhwa et al., 2021): 1) patients were confirmed with PPDFF based on recognized criteria, 2) comparison of at least two of ORIF with periarticular locked plates, retrograde IMN, and DFR was available, 3) reported at least one of the data on major complications and revision rate, 4) randomized controlled trials (RCTs), or retrospective or prospective study designs, and 5) full text published in English.

The following exclusion criteria were also defined: 1) adequate sample size (<5 patients), 2) did not report outcomes of interest, 3) duplicate studies with inadequate information or poor methodological quality, and 4) animal study, review, correspondence, letter to the editor, and conference abstract.

We obtained six meta-analyses (Ebraheim et al., 2015; Shin et al., 2017; Liu et al., 2019; Shah et al., 2020; Quinzi et al., 2021; Wadhwa et al., 2021) through searching PubMed by two reviewers using periprosthetic distal femoral fracture and synonyms, and two (Quinzi et al., 2021; Wadhwa et al., 2021) out of the six meta-analyses published in 2021. Then we conducted an updated literature search in PubMed, Embase, and the Cochrane Central Register of Controlled Trials (CENTRAL) from January 2020 until July 2021 in order to identify additional potential studies. We also manually checked references of all published meta-analyses and included studies to identify additional studies. Details of the literature search strategies are documented in Supplementary Table S1. A third senior reviewer was invited to solve any disagreements during the literature retrieval.

Two independent reviewers conducted the study selection. First, all identified potentially eligible records were imported into EndNote for literature database development. Second, all duplicate records were labeled and then removed into the trash file. Third, ineligible records were removed after screening titles and abstracts for eligibility. Fourth, full-text was checked for evaluating eligibility. Fifth, full-texts of all studies included in previous meta-analyses were downloaded and then verified for eligibility. A third senior reviewer was invited to solve any disagreements. The details of identifying and selecting eligible studies are displayed in Figure 1.

In this network meta-analysis, we evaluated two outcomes including major complication rates over the follow-up period and all-cause revision rates. Major complication rates were defined as the incidence of deep infection, nonunion, symptomatic malunion, hardware failure, aseptic loosening of arthroplasty, periprosthetic fracture, and extensor mechanism complications (Quinzi et al., 2021).

Two independent reviewers utilized the standard data extraction sheet to extract the following information: basic characteristics of the study (the first author’s name, publication year, country of the first author, study design, comparison, and methodological information), essential characteristics of patients [sample size (%male), mean age (years), mean Charlson Comorbidity Index (CCI) score, mean American Society of Anesthesiologists (ASA) score, and mean time of follow-up], and outcomes. A third senior reviewer was asked to solve any disagreements.

We used Stata 14.0 (Stata Corp LP, College Station, Texas, United States) to create an evidence structure of three surgical treatments on the basis of accumulated sample size and the number of eligible studies. The accumulated sample size was used to weigh the size of the circle representing treatment, and the accumulated numbers of eligible studies were used to weigh the width of the line representing the number of studies.

We simultaneously performed direct and network meta-analysis, which was conducted with the aggregate data drug information system (ADDIS) software (Groningen, the Netherlands, www.drugis.org). For the direct meta-analysis, a random-effects model which considers the variations across studies in the real world was used. Estimates were reported in the odds ratio (OR) with a 95% confidence interval (CI) because two outcomes were categorical variables. Statistical heterogeneity was evaluated by using the chi-square test (Cochrane Q) (Bowden et al., 2011) and an I² statistic (Higgins and Thompson, 2002). Substantial statistical heterogeneity was considered if the p-value was less than 0.1 and the I² value was equal to or more than 50.0% (Higgins et al., 2003).

Network meta-analysis was performed using the Markov Chain Monte Carlo (MCMC) simulation which is graphically displayed in Supplementary Figure S1. Estimates were reported in OR with a 95% credible interval (CrI). For major complications rates, parameters including four chains, 20,000 tunings and 50,000 simulation iterations, the thinning interval of 10, 10,000 inference samples, and variance scaling factor of 2.5 (Cipriani et al., 2013) were defined to achieve a good convergence. However, for the revision rate, four chains, 20,000 tuning, and 100,000 simulation iterations, the thinning interval of 10, 20,000 inference samples, and variance scaling factor of 2.5 (Cipriani et al., 2013) were defined to achieve a good convergence. The convergence of data was evaluated using the Brooks Gelman-Rubin statistical method, and a good convergence was suggested when a potential proportional reduction factor (PRF) was close to 1 (Brooks and Gelman, 1998; Burger and Schall, 2015). Moreover, a PRF of 1.2 also indicated an acceptable convergence (Brooks and Gelman, 1998; Burger and Schall, 2015).

In this network meta-analysis, a closed loop was available for three surgical treatments, and we, therefore, conducted the split node method to investigate whether the presence of inconsistency or not between direct and indirect effects (Dias et al., 2010; Albert and Makowski, 2019). This test was performed by using Stata 14.0 (Stata Corp LP, College Station, Texas, United States).

We estimated the ranking probabilities of being a certain rank for these three surgical treatments and utilized surface under the cumulative ranking curve (SUCRA) values to generate a ranking hierarchy of surgical treatments (Singh et al., 2015).

We created the comparison-adjusted funnel plot to qualitatively examine the publication bias and small study effects because the accumulated number of eligible studies was more than 10 (Palma Perez and Delgado Rodriguez, 2006). The publication bias test was performed by using Stata 14.0 (Stata Corp LP, College Station, Texas, United States).

A total of 119 records were initially identified after conducting an updated search, and then 28 duplicate records were removed. A total of 90 ineligible studies were excluded after screening titles, abstracts, and full-texts for eligibility due to meta-analysis, unrelated to the topic, and duplicate records, so one study (Gausden et al., 2021) was considered to meet our selection criteria. Moreover, we obtained 16 eligible studies (Platzer et al., 2010; Hou et al., 2012; Aldrian et al., 2013; Horneff et al., 2013; Kiluçoğlu et al., 2013; Gondalia et al., 2014; Meneghini et al., 2014; Leino et al., 2015; Park and Lee, 2016; Matlovich et al., 2017; Ruder et al., 2017; Gan et al., 2018; Hoellwarth et al., 2018; Darrith et al., 2020; García Guirao et al., 2020; Jennison and Yarlagadda, 2020) from published meta-analyses and references. Finally, we included 17 studies (Platzer et al., 2010; Hou et al., 2012; Aldrian et al., 2013; Horneff et al., 2013; Kiluçoğlu et al., 2013; Gondalia et al., 2014; Meneghini et al., 2014; Leino et al., 2015; Park and Lee, 2016; Matlovich et al., 2017; Ruder et al., 2017; Gan et al., 2018; Hoellwarth et al., 2018; Darrith et al., 2020; García Guirao et al., 2020; Jennison and Yarlagadda, 2020; Gausden et al., 2021) in this network meta-analysis. Details of study retrieval and selection are shown in Figure 1.

A total of 14 retrospective cohort studies (Platzer et al., 2010; Hou et al., 2012; Horneff et al., 2013; Kiluçoğlu et al., 2013; Gondalia et al., 2014; Meneghini et al., 2014; Leino et al., 2015; Park and Lee, 2016; Matlovich et al., 2017; Ruder et al., 2017; Gan et al., 2018; Hoellwarth et al., 2018; Darrith et al., 2020; Gausden et al., 2021) and 3 case series (Aldrian et al., 2013; García Guirao et al., 2020; Jennison and Yarlagadda, 2020) were included. Three studies were 3-arm designs (Platzer et al., 2010; Darrith et al., 2020; García Guirao et al., 2020), and the remaining 14 studies were 2-arm designs (Hou et al., 2012; Aldrian et al., 2013; Horneff et al., 2013; Kiluçoğlu et al., 2013; Gondalia et al., 2014; Meneghini et al., 2014; Leino et al., 2015; Park and Lee, 2016; Matlovich et al., 2017; Ruder et al., 2017; Gan et al., 2018; Hoellwarth et al., 2018; Jennison and Yarlagadda, 2020; Gausden et al., 2021). The accumulated sample size was 655, 250, and 159 in ORIF, IMN, and DFR groups, respectively. Comparison of ORIF and IMN, ORIF, and DFR, as well as IMN and DFR was available in 12 (Platzer et al., 2010; Hou et al., 2012; Aldrian et al., 2013; Horneff et al., 2013; Kiluçoğlu et al., 2013; Gondalia et al., 2014; Meneghini et al., 2014; Park and Lee, 2016; Matlovich et al., 2017; Darrith et al., 2020; García Guirao et al., 2020; Gausden et al., 2021), 8 (Platzer et al., 2010; Leino et al., 2015; Ruder et al., 2017; Gan et al., 2018; Hoellwarth et al., 2018; Darrith et al., 2020; García Guirao et al., 2020; Jennison and Yarlagadda, 2020), and three studies (Platzer et al., 2010; Darrith et al., 2020; García Guirao et al., 2020), respectively. The evidence structure of the three surgical techniques is shown in Figure 2. The MINORS score of individual study ranged from 14 to 17, with a moderate level for overall methodological quality. The basic characteristics of included studies and patients are summarized in Tables 1, 2.

Node split generated an inconsistency factor (IF) of 1.01 with a p-value of 0.303 for the closed loop of major complications rates and an IF of 1.02 with a p-value of 0.264 for the closed loop of revision rates (Supplementary Figure S2), suggesting consistency between direct and indirect effects.

Direct meta-analyses did not reveal significant difference for comparison between ORIF and IMN (12 studies, OR = 0.74; 95% CI = 0.38 to 1.45; I² = 37.7%), comparison between ORIF and DFR (eight studies, OR = 1.08; 95% CI = 0.59 to 1.99; I² = 0.00%), and comparison between INM and DFR (three studies, OR = 0.61, 95% CI = 0.12 to 3.18, I² = 0.00%). Meanwhile, these results were not significantly changed in the network meta-analysis of comparison between ORIF and INM (OR = 0.78; 95% CrI = 0.39–1.49), comparison between ORIF and DFR (OR = 1.04; 95%CrI = 0.44–2.44), and comparison between INM and DFR (OR = 1.34; 95% CrI = 0.48–3.80). However, the trend between INM and DFR was inversely benefitted to DFR because OR was changed from 0.61 to 1.37. All statistical results are shown in Figure 3.

Direct meta-analyses revealed a marginally favorable result in patients who received ORIF compared with IMN (12 studies, OR = 0.60; 95% CI = 0.35 to 1.03; I² = 0.00%). However, this result received the support of network meta-analysis (OR = 0.60; 95% CrI = 0.31–1.10). For remaining comparisons including the comparison between ORIF and DFR (eight studies, OR = 0.96; 95% CI = 0.32 to 1.10; I² = 0.00%) and comparison between IMN and DFR (three studies, OR = 0.73; 95% CI = 0.15 to 3.53; I² = 0.00%) did not achieve statistically significant level in direct meta-analyses. Similarly, the network meta-analyses of these two comparisons did possess unchanged results. However, the trend between INM and DFR was inversely benefitted to DFR because OR was changed from 0.73 to 1.64 in network meta-analysis. All statistical results are shown in Figure 4.

We generated ranking probabilities of these three surgical treatments in terms of major complication rates and revision rates. According to the results of ranking probabilities, DFR had the highest probability of becoming the best option, ORIF had the probability of ranking at a second option, and IMN at the third option for two outcomes. The plot of rankings of three surgical treatments is shown in Figure 5.

We generated a comparison-adjusted funnel plot for major complications and revision rates, and a symmetric outline indicates the absence of publication bias and small sample size effects (Supplementary Figure S3).