Two authors systematically searched the bibliographic databases, including PubMed, Embase, and the Cochrane Library, starting from their inception to January 31, 2025, with no limits for language and geographical region. The search strategies used a combination of the following search terms (1): carbapenem-resistant or carbapenemase-producing or carbapenem-nonsusceptible or multidrug-resistant gram-negative bacteria or extensively drug-resistant gram-negative bacilli; (2) Colistin or Colimycin or Colisticin or Polymyxin E or Colistin Sulfate or Sulfate or Colistin or Totazina or Coly Mycin. The search strategies are depicted in Supplementary Table S1.

The inclusion criteria were as follows: (1) study designs limited to observational studies or randomized controlled trials (RCTs); (2) participants with microbiologically confirmed CR-GNB infections; (3) intervention measures included COL combination therapy and monotherapy with intravenous administration; and (4) reported one of the following endpoints: 28-day all-cause mortality, in-hospital all-cause mortality, clinical improvement rate, microbiological eradication, length of stay (LOS) in the ICU, total LOS, nephrotoxicity and neurotoxicity. No exclusion criteria were set for the dose of COL.

The criteria for exclusion were as follows: (1) animal experiments, in vitro studies, pediatric research, editorial letters, comments, guidelines, conference abstracts, systematic reviews, meta-analyses; (2) studies with incomplete outcome data or noncomparable outcome metrics; and (3) studies enrolling fewer than 10 patients.

Two authors independently assessed the relevant studies according to inclusion/exclusion criteria, and negotiated with a third party to resolve any disagreements. The study data were independently extracted by two reviewers in a standardized established data format, including the following study characteristics: first author’s name, type of study design, publication year, country, sex and age of patients, Acute Physiology and Chronic Health Evaluation (APACHE II) score and Sequential Organ Failure Assessment (SOFA), sample size, type of pathogen, co-administration of other antibiotics, 28-day all-cause mortality, in-hospital all-cause mortality, microbiological eradication rate, clinical improvement rate, length of stay (LOS), nephrotoxicity, and neurotoxicity. Any disagreements were resolved through discussion and consultation. The reviewers attempted to establish contact with the authors via email in cases where insufficient data were available.

CR-GNB are referred to as gram-negative bacteria (GNB) identified from clinical specimen cultures that demonstrate resistance to imipenem, meropenem, and ertapenem, as indicated by antimicrobial susceptibility testing results. The primary outcome was 28-day all-cause mortality, while secondary outcomes included in-hospital all-cause mortality, clinical improvement rate, microbiological eradication rate, incidence of nephrotoxicity and neurotoxicity, ICU length of stay (LOS), and total hospital LOS. The all-cause mortality referred to the all-cause hospital mortality. Clinical improvement was defined as the resolution of infection-related signs or symptoms without recurrence or survival during the follow-up period. Microbiological eradication was defined as the absence of baseline pathogens in cultures obtained during follow-up. Nephrotoxicity was defined as a marked increase in serum creatinine level or an obvious decrease in glomerular filtration rate, which prompted renal replacement therapy and was diagnosed based on the classification of risk, injury, failure, loss, and end-stage kidney disease criteria (RIFLE) (Hartzell et al., 2009). Neurotoxicity is defined as the occurrence of symptoms such as dizziness, muscle weakness, facial and peripheral paraesthesia, visual disturbances, vertigo, confusion, hallucinations, seizures, ataxia, partial deafness, and neuromuscular blockade during the administration of polymyxins (Molina et al., 2009). The critically ill patient is defined as having an Acute Physiology and Chronic Health Evaluation II score greater than 15. To distinguish publications from the same first author published in the same year, alphabetical suffixes (e.g., “Author2025a”, “Author2025b”) were added.

Quality assessment was performed independently by two investigators, using the ROBINS-I for observational studies and the ROB 2.0 for RCTs (Sterne et al., 2016; Sterne et al., 2019). The classification of the overall risk of bias for each included study was as follows: low, if there was low risk of bias in all domains, unclear, if there was unclear risk of bias in one or more domains without any judgment of high risk of bias, and high, if there was high risk of bias in one or more domains (Alhazzani et al., 2018). Discrepancies were resolved by a third investigator after a joint re-evaluation of the original studies was conducted by the previous reviewers.

Outcomes were rated according to the Grading of Recommendations, Assessment, Development, and Evaluations (GRADE) Framework (Guyatt et al., 2008). Certainty of evidence could be considered as ‘very low’, ‘low’, ‘moderate’, or ‘high’ depending on the number of downgrades attributed to each of the five topics: (1) risk of bias, (2) imprecision, (3) inconsistency, (4) indirectness, and (5) publication bias. Risk of bias was rated based on the outline in the section ‘Risk of bias assessment’. Impression was deemed to be present if outcomes were calculated from only a few studies with small sample sizes, or if decision-making would differ when the lower and upper confidence limits were considered as the real effect. Publication bias was determined by assessing funnel plots. Indirectness was deemed if the study did not use a placebo or control as a comparator, whereas inconsistency was determined according to the heterogeneity measures (I² or tau²) (Guyatt et al., 2008).

All statistical analyses were conducted using a random-effects model via Review Manager (RevMan) version 5.4 (The Nordic Cochrane Centre, Copenhagen, Denmark). Differences were expressed as odds ratios (ORs) with 95% confidence intervals (CIs) for dichotomous outcomes and as mean differences (MDs) with 95% CIs for continuous outcomes. Heterogeneity was assessed using the inconsistency index (I2) and the Q statistic. A P-value of less than 0.10 for the Q statistic was considered significant (Higgins and Thompson, 2002). To evaluate the impact of potential outlier studies on the stability of effect estimates, sensitivity analyses were conducted using the leave-one-out method (Deeks et al., 2024). Subgroup analyses were conducted for specific categories, including study design (RCTs and observational studies), study setting (multicenter vs. single-center), pathogen subtype (CRAB-infected patients only), antibiotic regimen (COL + meropenem, COL+ tigecycline, COL + rifampicin), and baseline severity (critically ill and stable patients). Publication bias was assessed using funnel plots and Egger’s test, conducted with R version 4.4.2 (R Foundation for Statistical Computing, Vienna, Austria). The presence of publication bias was inferred when both indicators yielded statistically significant results. A two-tailed P-value less than 0.05 was considered statistically significant (Page MJ and Sterne, 2024).

Systematic searches across PubMed (n=1082), Embase (n=4759), and the Cochrane Library (n=159) initially identified 6000 records, with 4928 studies retained for screening after duplicate removal. A total of 4782 publications unrelated to the research topic were excluded by screening the titles and abstracts. Subsequently, the full text of the remained 146 articles was thoroughly examined, and 120 were excluded for not meeting the inclusion criteria. Finally, 26 studies were included in the analysis (Falagas et al., 2006; Aydemir et al., 2013; Durante-Mangoni et al., 2013; Batirel et al., 2014; Kalin et al., 2014; Sirijatuphat and Thamlikitkul, 2014; Rigatto et al., 2015; Yilmaz et al., 2015; Ghafur et al., 2016; Parchem et al., 2016; Ghafur et al., 2017; Abdelsalam et al., 2018; Amat et al., 2018; Makris et al., 2018; Paul et al., 2018; Park et al., 2019; Shi et al., 2019; Katip and Uitrakul, 2020; Katip et al., 2020; Park et al., 2020; Katip and Oberdorfer, 2021; Chang et al., 2022; Hao et al., 2022; Sirijatuphat et al., 2022; Kaye et al., 2023; Katip et al., 2024). The analysis included a total of 3964 patients (2135 received COL combination therapy and 1829 received COL monotherapy). Figure 1 presents the literature selection flowchart. Table 1 and Table 2 summarize the baseline characteristics of the included RCTs and observational studies, respectively.

The risk of bias assessment for the 8 RCTs via the ROB 2.0 tool (Figure 2) revealed high risk in 5 studies and moderate risk in 3 studies, with primary biases stemming from deficiencies in double-blinding implementation and inadequate allocation concealment. For the 18 observational studies evaluated with the ROBINS-I tool (Figure 3), 9 studies demonstrated a moderate risk of bias and 9 exhibited a serious risk of bias, predominantly attributed to uncontrolled confounding factors, heterogeneity in intervention protocols, and selection bias.

Table 1 shows the main characteristics of the included studies. All 11 included studies were conducted from 2006 to 2024, with sample sizes ranging from 14 to 214 participants. Eight studies were RCTs (Aydemir et al., 2013; Durante-Mangoni et al., 2013; Sirijatuphat and Thamlikitkul, 2014; Abdelsalam et al., 2018; Makris et al., 2018; Paul et al., 2018; Sirijatuphat et al., 2022; Kaye et al., 2023), while 18 studies were observational (Falagas et al., 2006; Batirel et al., 2014; Kalin et al., 2014; Rigatto et al., 2015; Yilmaz et al., 2015; Ghafur et al., 2016; Parchem et al., 2016; Ghafur et al., 2017; Amat et al., 2018; Park et al., 2019; Shi et al., 2019; Katip and Uitrakul, 2020; Katip et al., 2020; Park et al., 2020; Katip and Oberdorfer, 2021; Chang et al., 2022; Hao et al., 2022; Katip et al., 2024).Seven studies were multicenter (Durante-Mangoni et al., 2013; Rigatto et al., 2015; Abdelsalam et al., 2018; Makris et al., 2018; Paul et al., 2018; Chang et al., 2022; Hao et al., 2022; Kaye et al., 2023) while the others were single-center (Falagas et al., 2006; Aydemir et al., 2013; Kalin et al., 2014; Sirijatuphat and Thamlikitkul, 2014; Yilmaz et al., 2015; Ghafur et al., 2016; Parchem et al., 2016; Ghafur et al., 2017; Park et al., 2019; Shi et al., 2019; Katip and Uitrakul, 2020; Katip et al., 2020; Park et al., 2020; Katip and Oberdorfer, 2021; Sirijatuphat et al., 2022; Katip et al., 2024). The study population had a mean age of 61.4 ± 15.8 years, with a male predominance (57.6%). The research spanned multiple countries, such as China, the United States, Greece, Italy, Turkey, Thailand, South Korea, Brazil, and India. Fourteen studies reported exclusively on CRAB infections. Two studies focused solely on carbapenem-resistant Enterobacteriaceae (CRE). Common infection sites include the respiratory tract, bloodstream, abdominal cavity, and urinary tract. The combination regimens encompassed carbapenems (e.g., meropenem, imipenem), aminoglycosides (e.g., amikacin, gentamicin), β-lactam/β-lactamase inhibitor combinations (e.g., piperacillin/tazobactam, ceftazidime/avibactam, cefoperazone/sulbactam), fluoroquinolones (e.g., levofloxacin, sitafloxacin), and other antibiotics (e.g., tigecycline, rifampicin). Notably, seven studies utilized triple or higher-order COL combination protocols (e.g., COL plus meropenem plus tigecycline). Baseline data indicated greater disease severity and Charlson Comorbidity Index (CCI) scores in the combination therapy group than in the monotherapy group. The detailed distributions of infection sites, disease severity, and the CCI are provided in the supplementary file (Supplementary file, Table S2).

All twenty studies (n=3193) were accessible to compare the 28-day all-cause mortality rate (Durante-Mangoni et al., 2013; Batirel et al., 2014; Sirijatuphat and Thamlikitkul, 2014; Rigatto et al., 2015; Yilmaz et al., 2015; Ghafur et al., 2016; Parchem et al., 2016; Ghafur et al., 2017; Amat et al., 2018; Makris et al., 2018; Paul et al., 2018; Park et al., 2019; Shi et al., 2019; Katip and Uitrakul, 2020; Park et al., 2020; Katip and Oberdorfer, 2021; Hao et al., 2022; Sirijatuphat et al., 2022; Kaye et al., 2023; Katip et al., 2024). The 28-day all-cause mortality rate was 38% in the COL combination therapy group and 42.6% in the COL monotherapy group. There was no significant difference between patients treated with colistin and other antibiotics (RR 0.94, 95% CI 0.85–1.05, I² = 25%, p = 0.30) (Figure 4). The combination therapies included dual therapies and a small number of triple or quadruple therapies. Figure 5 shows a funnel plot of the 28-day all-cause mortality rate, which shows no evidence of publication bias. In addition, the results of Egger’s test indicated a low risk of publication bias (p = 0.5156). Sensitivity analysis via the leave-one-out method confirmed the robustness of the findings. Subgroup analyses stratified by study design, study setting, pathogen subtype (CRAB or CRE), antibiotic regimen, and baseline severity were conducted. The results revealed no significant difference in 28-day all-cause mortality between the two groups across all subgroups (Table 3) (Supplementary file, Figure S1).

Evidence quality for 28-day mortality and microbiological clearance rate was rated as moderate, primarily because the included studies carried a moderate risk of bias. Evidence quality for in-hospital mortality, clinical improvement rate, total hospital stay duration, ICU stay duration, nephrotoxicity, and neurotoxicity was rated as low, mainly due to a high risk of bias and imprecision in the included studies.