The study was reported according to the Preferred Reporting Items for Systematic Review and Meta-Analysis (PRISMA) (Page et al., 2021). The protocol was informed by the Cochrane Handbook for Systematic Review of Interventions and registered via the International Prospective Register of Systematic Reviews (PROSPERO) (CRD42023417200).¹

We selected relevant studies published before March 31, 2024, by searching Embase, PubMed, Cochrane Library, Web of Science (WOS), and ClinicalTrials.gov. We applied no language restrictions. A combination of Mesh with free text search was applied using the keywords gut microbiota, spinal cord injury, and their associated subject words. The specific retrieval strategies are detailed in Supplementary Table S1.

The same two investigators (Zaowei Zhong and Junqiao Lv) implemented study selection on an independent basis by firstly screening the titles and abstracts, followed by reviewing the full texts of eligible articles. Disagreements, if any, were resolved by consulting a third investigator (Lin Sun). Specifically, the inclusion criteria were: (1) applied an observational design (e.g., case–control study, cross-sectional study, and cohort study); (2) performed gut microbiota analysis with available data on diversity or abundance measures; and (3) included participants with spinal cord injury.

We used the retrieval method in PubMed to search for literature. We conducted a meta-analysis of identified bacteria that potential produce short-chain fatty acids by reviewing literature from establishing the database to April 7, 2024. Retrieval strategy: ((((“produced”)) AND (“bacteria”)) OR (“bacterial fermentation”)) AND (“short-chain fatty acids”)) AND (Review).

The eligible studies are as follows: (1) the samples are from patients with SCI, and 16S rRNA gene sequencing technology to elucidate the relative abundance changes of gut microbiota. (2) original research to compare the composition of gut microbiota between patients with SCI and non-SCI controls. The exclusion criteria were as follows: (1) the comparative study of microbial relative abundance between patients with SCI and non-SCI controls could not be provided; (2) There were no data on gut microbiota relative abundance. The Newcastle Ottawa scale (NOS) was used to assess the quality of case–control studies (Wells et al., 2018).

The primary outcomes were as follows: the relative abundance of bacteria. Two reviewers independently extracted details from the studies included in the meta-analysis. It included the first author, year of publication, location, sample size, sex, days or months from injury, injury level, injury degree, sequencing platform, and relative abundance of bacteria. However, most of the results were presented as a graph rather than an exact report of the raw data they obtained, which meant that the actual numbers had to be estimated from the data extracted from the graph. Some study results’ mean and standard deviation could not be obtained, so we used the sample’s quartile, median, maximum, minimum, and p-value to estimate the mean and standard deviation (Luo et al., 2018; Shi et al., 2023; Shi et al., 2020; Wan et al., 2014). Some study results included two subgroups, which we combined as one and estimated their mean and standard deviation for follow-up analyses (Altman et al., 2000). Numerical values from the graphs were estimated by GetData Graph Digitizer 2.22 software,² and the study will be excluded if data were not presented or obtainable. Any controversies lead to a third reviewer settling the problem by discussion.

Demultiplexed raw DNA sequences from the stools of SCI and non-SCI subjects from different studies were downloaded from the NCBI. The data set of subacute(PRJNA 724686) was only the SCI, and the data set of people from the same region was used as the control(PRJNA 247489, PRJNA792991). The dataset has been analyzed to obtain the mean relative abundance and SD. The raw sequencing data were imported into QIIME2/2024.01 for data processing (Bolyen et al., 2019). Due to the technical variation in the data sets included in these studies (DNA extraction kits, primers, sequencing, and platform), each data set was separately denoised and processed into amplicon sequence variants using DADA2 (Bolyen et al., 2019). The reads were trimmed that the Quartile quality score was <30. we also trimmed 21 nucleotides from the 5′ end of each read. Taxonomic classification of the operational taxonomic units (OTUs) was conducted using the classify-sklearn classification methods based on the SILVA database³ and the q2- feature-classifier plugin.

The standardized mean difference (SMD) was used as the effect indicator because the outcome indicators were all continuous variables, and point estimates and 95% confidence intervals (CI) were given for each effect measure. Statistical results were presented with a forest map. We ran I² testing to assess the magnitude of the heterogeneity among the studies included. Suppose the heterogeneity test result was I² ≤ 50%, meta-analysis was performed using a fixed effects model; if the heterogeneity test result was I² > 50%, me-ta-analysis was performed using a random effects model, and the sources of heterogeneity needed to be further analyzed. After excluding obvious clinical and methodological heterogeneity, a random effects model was used for the meta-analysis (Higgins et al., 2003). Sensitivity analyses were conducted by omitting each study in turn and then rerunning the meta-analysis and assessing the differences between the results and the actual combined results.

We assessed the possibility of publication bias with a funnel plot. Egger’s tests were used to assess funnel plot asymmetry, and no significant publication bias was defined as a p-value >0.1. The effect of publication bias on the results of the meta-analysis was assessed using the trim-fill method (Duval and Tweedie, 2000). Egger’s tests and trim-fill method were performed with R language version 4.2.2,⁴ All graphical presentations in this study were performed with GraphPad Prism version 9,⁵ and Review Manager version 5.3 (Cochrane Collaboration).⁶

Three hundred seventy-eight studies were retrieved from PubMed, Embase, WOS, CNKI, and Cochrane Library; 164 studies marked as ineligible by automation tools were excluded, and the remaining 174 articles were evaluated. After carefully reading the titles and abstracts, 196 articles were excluded because they needed to meet the inclusion criteria. Three articles were excluded as they did not set a health control group. After reading the full texts carefully and comparing the selection criteria, 11 studies were finally included. All studies included were published between 2016 and 2024, yielding 720 individual fecal samples for microbiome analysis. Seven studies were conducted in China, one in Turkey, one in United States, one from Israel, and one in Italy. Seven studies included male and female participants, while two studies included male only (Figure 1).

All studies included in the meta-analyses were compared between patients with SCI and non-SCI controls, who were adjusted with age, and all participants without any chronic conditions. Additionally, body mass index (BMI) was matched in two studies, dietary habits were matched in four studies by all patients, and non-SCI individuals were provided standard hospital food before sample collection. Seven studies excluded the participants treated with antibiotics within 1 to 3 months before stool collection; two studies used antibiotics for inflammation in the acute stage. Five studies excluded the participants who used probiotics before enrollment. The ASIA neurological function scale of patients with SCI ranged from A to D in four studies and A, C, and D in one study: three included patients with only an A score. The rest of the studies did not specify the ASIA scores.

All studies used 16S rRNA sequencing to evaluate gut microbiota samples. Seven studies measured the V3-V4 regions, and four measured the V4 region. As for Sequencing platforms used, Illumina was used in 10 studies and microbial ecology in one study. Bacterial and archaeal rRNA databases for taxonomic assignments of sequence data in studies included are SILVA or GreenGenes. More details are shown in Table 1. The quality of case–control studies is shown in Table 2.

We searched for literature and summarized the bacterial genera that potential produce SCFAs (Dalile et al., 2019; Dicks, 2022; Vacca et al., 2020). Summarized in Table 3.

We analyzed the PRJNA724686 dataset, a sample collected from 21 to 36 days after SCI. We identify the data as subacute. The injury time and sampling time of the study population included in the meta-analyses range from 2 months to 10 years, and we consider it the chronic phase.

In patients with SCI during the subacute phase, the relative abundance of Firmicutes and Lactobacillus decreases. In contrast, the relative abundance of Bacteroidetes, Bacteroides, Blautia, Clostridales, Faecalbacterium, Ruminococcus, Coprococcus, Lachnospira, Dorea, Prevotella, Roseburia, Atopobium, and Bifidobacterium increases. There is no significant difference in Eubacteria (Figure 2).

In the meta-analyses of patients with chronic SCI, the number of Firmicutes decreased in the SCI group, while there was no significant difference in the number of Bacteroidetes. The average abundance proportion of Bifidobacterium, Bacteroides, Blautia, and Eubacterium in SCI patients is higher, with statistical differences. In contrast, the abundance of Clostridiales, Ruminococcus, Feacalbacterium, Coprococcus, and Lachnospira in SCI patients is lower, achieving statistical differences. There was no significant difference in Dorea, Prevotella, Roseburia, Atopobium, and Lactobacillus (Figures 3–6).

We use funnel plots and Egger’s test to evaluate publication bias (Figures 7, 8; Table 4). We observed a phenomenon in the meta-analyses of the genera Bacteroidetes, Bluatia, and Bifidobacterium, where there was one study in their funnel plots, and the effectiveness of the study was outside the dashed range of the funnel plots. We consider that the results of these studies beyond the dashed range may have significant biases or anomalies, which may be due to methodological issues, data quality considerations, or other potential factors (Supplementary Figures S1, S2). We paid special attention to this study and tested its impact on the overall results in subsequent sensitivity analysis. Considering their significant bias in the meta-analyses of bacterial abundance, we have decided to exclude them. Then, a meta-analysis was conducted again, and it was found that heterogeneity was significantly reduced (Figures 3, 4) After Egger’s test, there are public biases in Bacteroides and Prevotella. The trim-fill shows that bias does not affect the meta-analyses result. The rest of the bacteria were not found in public bias. After applying the trim-fill method, the meta of Bacteroides changed direction and showed no significant difference. The meta-analysis of Prevotella did not show significant differences after using the trim-fill method, and there is no evidence of publication bias (Table 4; Supplementary Figure S3).

We performed sensitivity analyses by omitting each study, and the results were stable, except Bacteroides and Eubacterium were unstable. At the same time, we also summarized the studies that significantly impacted heterogeneity in the meta-analyses of the relative abundance of various bacteria (Table 5). The study by Zhang et al. (2018) significantly impacts the p-value results of Bacteroides, which may be due to the higher relative abundance of genus Bacteroides in the population with ASIA grade A SCI in the chronic phase compared to other grades. ASIA grade may be a source of heterogeneity (Zhang et al., 2018).

The meta-analysis of Eubacterium showed no significant difference after excluding studies by Yu et al. (2021) which may be related to their study only including the causes of thoracic spinal cord injury. In Blautia meta-analysis, the study by Kong et al. showed the opposite direction compared to others (Kong et al., 2023). After exclusion, it was found that there was a significant difference in the meta-analysis results and a significant reduction in heterogeneity. After comparing it with other studies, it was found that the study of Kong et al. population used antibiotics. Antibiotics may be a source of heterogeneity and bias, and antibiotics may significantly influence the relative abundance of Blautia. After conducting a sensitivity analysis in Ruminococcus, significant heterogeneity was observed in the study by Kong et al., but it did not alter the results of the meta-analysis. After incorporating study of Lin et al. into the meta-analysis of Bifidobacterium, it was found that there was a significant increase in heterogeneity and a change in the results of the meta-analysis, with a significant increase in p-value. We found bias in the funnel plot and compared it with the results of other studies. We found that research of Lin et al. direction was the opposite, and the different lengths of injury time may be the reason (Lin et al., 2020). Research of Yu et al. contributed to all heterogeneity in the meta-analysis of Faecalbacterium, and we speculate that the source of heterogeneity is the damaged segment. For the meta-analysis of Prevotella, studies from Yu et al. and Lin et al. significantly influenced heterogeneity, and the use of antibiotics or damaged segments may affect the relative abundance of Prevotella. Clostridiales did not observe significant heterogeneity (Table 4; Figures 3–6).