The Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) (5) for NMA guidelines were followed in the study. Literature search was conducted in PubMed, Web of Science, Embase, and Cochrane Library, the publication time was set from the establishment of the database to October 31, 2021. Using the following Medical Subject Headings (MeSH) terms and keywords: “delayed encephalopathy after acute carbon monoxide poisoning,” “carbon monoxide intoxication,” “acupuncture,” “mesenchymal stem cells (MSCs),” “mental cord blood mononuclear cells (CB-MNCs),” “nalmefene,” “n-butylphthalide (NBP),” “kinnado,” “Ginaton,” “Xingzhi Yinao granules (XZYN),” “glucocorticoids,” “dexamethasone (DXM),” “oxiracetam,” “translational direct current stimulation (tDCS),” “Citicoline,” “hyperbaric oxygen,” and other relevant conceptual keywords.

Diagnostic Criteria for DEACMP (7). Any of the following clinical abnormalities observed 2–60 days after the recovery of the consciousness disorder caused by acute CO poisoning: (1) disturbances of mental state and/or consciousness (such as dementia or delirium), (2) extrapyramidal lesions (such as Parkinson's syndrome), and (3) pyramidal damage and focal cortical dysfunctions.

The inclusion criteria of the present study was set as follows: (1) Randomized controlled trials (RCTs); (2) Patients diagnosed with DEACMP; (3) Interventions included commonly used drugs for the treatment of DEACMP; and (4) At least one of the following outcomes was included: mini-mental state examination (MMSE), the national institutes of health stroke scale (NIHSS), the activity of daily living (ADL), Montreal cognitive assessment (MoCA), Barthel index (BI), P300 latency, P300 amplitude, malondialdehyde (MDA), and adverse reactions were measured with more than one outcome as their endpoints.

Exclusion criteria of the present study was: (1) Animal experiments; (2) Inconsistent intervention measures; (3) Duplicate articles; (4) Diagnostic or efficacy evaluation criteria were not mentioned; (5) There were obvious statistical errors in the literature; and (6) Research whose full text cannot be obtained.

According to the predetermined screening criteria, two researchers independently screened the papers, extracted data, and conducted cross-checks. Any disagreements among the researchers were discussed or arbitrated by a third party. Then articles were imported into Endnote, duplicates were removed, and preliminary screening was done by reading the titles and abstracts. If an article satisfied the requirements for inclusion, the reviewers carried out a full-text read. Finally, the first author, year of publication, sample size, age and course of the disease, intervention measures, outcome indicators, and course of treatment were mostly extracted using an Excel data extraction table.

The risk of bias (ROB) of the included studies was assessed using the Cochrane Risk of Bias tool. These studies were estimated from seven domains of ROB, defined as having a high, low, or unclear ROB, and the seven domains included “random sequence generation,” “allocation concealment,” “blinding of participants and personnel,” “blinding of outcome assessment,” “incomplete outcome data,” “selective reporting,” and “other bias.” The judgment of ROB was carried out by two authors separately in Review Manager (Version 5.3).

The statistical method of NMA was based on the frequency framework, and all outcome indicators were analyzed by the random effect model. For continuous variables, the standardized mean difference (SMD) was used as the effect size; for dichotomous variables, the odds ratio (OR) was as the effect size and the corresponding 95% confidence interval (CI) was calculated. Stata 16.0 software was applied to select the frequencyological framework random-effects model for the NMA, and the data were pre-processed using the network group command, and inconsistency tests were to assess the consistency of the results of direct and indirect comparisons. Efficacy ranking was then performed and the figure of the surface under the cumulative ranking curve (SUCRA) was plotted. A SUCRA of 100% indicated that the intervention was absolutely effective, whereas a SUCRA of 0 pointed out absolutely ineffective (8). If a closed loop existed, an inconsistency test was performed to determine the degree of agreement between the results of direct and indirect comparisons. Finally, funnel plots were drawn to verify whether there were small sample effects in the network and whether there was publication bias in the included studies.

The two-dimensional (2D) chemical structure of the drug was downloaded in Spatial Data Format (SDF) from the PubChem database (https://pubchem.ncbi.nlm.nih.gov/). Then the SDF file was submitted to Pharmmapper (http://www.lilab-ecust.cn/pharmmapper/) database for target prediction. All targets were transformed into recognized gene symbols using the UniProt database (9) (http://www.Uniprot.org/), and complete protein names were corresponded to the gene abbreviations.

Excel was used to prepare the network and type files of drug and gene symbols, which were then imported into Cytoscape 3.9.2 to construct the drug-targets network.

The GeneCards database (10) (https://www.genecards.org) and the Oline Mendeline Inheritance in Man (OMIM) database (11) (http://www.omim.org) were mined for potential targets related to DEACMP using the keyword “delayed encephalopathy after acute carbon monoxide poisoning.” The DEACMP-related targets were then obtained by integrating the targets from the two disease databases and removing duplicate values.

The obtained drug targets and disease targets were intersected by Venny 2.1.0 (https://bioinfogp.cnb.csic.es/tools/venny/) software to obtain potential targets. Then the potential targets were submitted to the STRING11.0 database (12) (https://string-db.org) to construct a PPI network model, with the organism species set to “Homo sapiens” and all settings set to default. The CytoNCA (13) plugin in Cytoscape 3.9.2 was used to analyze the attribute values of each node in the intersection network. All nodes with degree values more than twice the median degree value were chosen as key targets. The PPI network was further analyzed by the Molecular complex detection (MCODE) (14) plugin in Cytoscape 3.9.2 to identify potential protein functional modules.

Gene ontology (GO) enrichment and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis were performed by the Metascape database (15) (http://metascape.org/gp/index.html) to examine the biological function and potential mechanisms of the potential targets.

Molecular docking was performed among the drug and key target proteins. The three-dimensional (3D) structures of the key target proteins were downloaded from the Protein Data Bank (PDB) database (16) (https://www.rcsb.org). Then the key target proteins were imported into Molecular Operating Environment (MOE) software for removing water molecules and the original ligands, and molecular docking with the drug was then carried out. The binding activity is evaluated by the lowest binding energy.

A total of 816 articles were retrieved, and 17 articles were finally included after layer-by-layer screening (17–33), of which two articles were three-arm experiments and the other 15 were two-arm experiments. The literature selection process is shown in Figure 1. A total of 1,293 patients were considered eligible for this NMA, including 663 cases in the experimental group and 630 cases in the conventional therapy group. The basic characteristics of the included articles are shown in Table 1.

All 17 studies were RCTs, the main risk of bias in this study came from “random sequence generation (selection bias),” 13 studies reported on the method of generating specific randomized sequences, four of them distributed treatment regimens according to the order of admission, therefore, it is defined as high risk. One study applied allocation concealment, four studies were blinded, all 17 studies reported on the outcome indicators expected to be measured and did not identify early termination of trials (Figure 2).

Network plots of outcome indicators were constructed to exhibit all the available evidence of each treatment. All network relationships were generally centered on Conventional Therapy treatment, the dot size represented the sample size of the intervention, and the line thickness represented the number of RCTs using that treatment (Figure 3).

The closed loops in the reticulation were three-armed studies from the same article and there was no inconsistency.

The league plots of meta-analysis and probability ranking results are displayed in Table 2. A total of 12 studies reported MMSE, including 13 interventions and producing 78 two-by two comparisons, and 50 of the comparisons were statistically significant (P < 0.05, Table 2A); 5 studies reported ADL, including six interventions and producing 15 two-by two comparisons, and 10 of the comparisons were statistically significant (P < 0.05, Table 2B); five studies reported NIHSS, including five interventions and producing 10 two-by two comparisons, and six of the comparisons were statistically significant (P < 0.05, Table 2C); four studies reported BI, including five interventions and producing 10 two-by two comparisons, and 9 of the comparisons were statistically significant (P < 0.05, Table 2D); two studies reported MoCA, including three interventions and producing three two-by two comparisons, and two of the comparisons were statistically significant (P < 0.05, Table 2E); three studies reported P300 latency, including four interventions and producing six two-by two comparisons, and three of the comparisons were statistically significant (P < 0.05), Table 2F); three studies reported P300 amplitude, including four interventions and producing six two-by two comparisons, and three of the comparisons were statistically significant (P < 0.05, Table 2G); two studies reported MDA, including four interventions and producing six two-by two comparisons, and five of the comparisons were statistically significant (P < 0.05, Table 2H). Significant pairwise comparisons were highlighted in TextTitle and underlined.

The SUCRA results indicated that: the probability of SUCRA for MMSE scores was ranked from highest to lowest: MSCs + NBP, MSCs, CB-MNCs, Nalmefene, Hydrocortisone, Kinnado, tDCS, Acupuncture, DXM, NBP, XZYN, NBP + DXM, Conventional Therapy (where the dosage form of NBP differs between “NBP” and “NBP + DXM,” the former being NBP soft capsules and the latter being NBP injection; the doses of DXM before and after were different, the former was DXM 10 mg, the latter was DXM 5 mg), MSCs + NBP was probably the most effective intervention (Table 2A); the probability of SUCRA in ADL scores was ranked from highest to lowest as follows: NBP + DXM, CB-MNCs, XZYN, Acupuncture, Hydrocortisone, Conventional Therapy, NBP + DXM may be the most effective intervention (Table 2B); the probability of SUCRA for NIHSS scores was ranked from high to low as follows: NBP, tDCS, DXM, Acupuncture, Conventional Therapy, NBP may be the most effective intervention (Table 2C); the probability of SUCRA for BI scores was ranked from high to low as follows: MSCs + NBP, MSCs, tDCS, Acupuncture, Conventional Therapy, MSCs + NBP was probably the most effective intervention (Table 2D); the probability of SUCRA of MoCA scores was ranked from high to low as follows: XZYN, NBP + DXM, Conventional Therapy, XZYN may be the most effective intervention (Table 2E); the probability of SUCRA for P300 latency was ranked from high to low as follows: tDCS, Oxiracetam, XZYN, Conventional Therapy, tDCS may be the most effective intervention (Table 2F); the probability of SUCRA for P300 amplitude was ranked from high to low as follows: tDCS, Oxiracetam, XZYN, Conventional Therapy, tDCS may be the most effective intervention (Table 2G); the results of SUCRA probability for MDA was ranked from high to low as follows: Kinnado + Citicoline, Nalmefene, Citicoline, Oxiracetam, Kinnado + Citicoline may be the most effective intervention (Table 2H).

The results of the inverted funnel plots showed a concentrated distribution of all studies in the interior of the “funnel,” the nodes symmetrically distributed on both sides of the axis, indicating a low likelihood of publication bias (Figure 4).

Adverse reactions were reported in nine studies, while no serious adverse reactions were observed in all included studies. Adverse reactions that occurred during treatment period included hypothermia, dizziness, headache, swelling, pain in the ear, chest tightness, panic, and mild nausea, which recovered on their own without any specific treatment. Since most studies failed to report adverse reactions in a standardized manner, quantitative analysis of adverse reactions was not performed in this study.

This study had some limitations: (1) Some of the included studies did not report baseline data on patients comprehensively, such as the duration of disease and sex ratio of patients; (2) A rigorous exclusion of the literature was carried out to ensure homogeneity between studies, but there was still some clinical heterogeneity in the interventions of the control group in the included studies, such as the medication regimen, dose, frequency, and duration of medication. Meanwhile, sampling error, differences in patient age and gender, differences in underlying medical conditions, and differences in smoking and alcohol history were also sources of heterogeneity; (3) The disease is not frequently-occurring, and the number of RCT articles on this disease in the database is small, the number of included studies was only 17, and the literature of some treatment options had only one article, so the ranking of efficacy was not stable. Further validation with high-quality, multi-center, large-sample RCTs is recommended in the future.