Two authors independently searched the database of Pubmed, Embase, and Cochrane Library using the PICOS (Population, Intervention, Comparison, Outcome, and Study design) methods by the time of October 22, 2021. The entry terms included “general anesthesia” OR “general anesthesia” AND “local anesthesia” OR “local anesthesia” OR “regional anesthesia” OR “regional anesthesia” OR “spinal anesthesia” OR “spinal anesthesia” “epidural anesthesia” OR “epidural anesthesia” OR “neuraxial anesthesia” OR “neuraxial anesthesia” AND “delirium” OR “confusion” OR “agitation” OR “acute confusional state” OR “acute confusional syndrome,” and the search field was “title and abstract.” Since we sought to investigate all studies about the effect of GA vs. RA on POD incidence in adult patients undergoing surgery, we did not constrain the search terms for study designs.

Two authors were independently responsible for the screening process for titles and abstracts, while another two authors conducted the screening process for full text. The inclusion criteria were: (1) participants aged 18 years or older, (2) patients undergoing surgery under general and regional or local anesthesia, and (3) articles reporting the effect of GA vs. RA on POD incidence. The exclusion criteria were: (1) duplicate articles, (2) participants younger than 18 years old, (3) review or meta-analysis, (4) articles published as an abstract, letter, case report, basic research, editorial, note, method, or protocol, (5) articles presented in a non-English language; (6) studies without a specific number of patients with and/or without delirium, and (7) studies of all patients receiving GA or RA.

Two authors independently assessed the quality of included studies. For retrospective and prospective observational trials, the risk of bias was assessed using the Newcastle–Ottawa Quality Assessment Scale (NOS), which comprises the following three domains: selection, comparability, and outcome for cohort studies (13). There were four stars in the selection domain, two stars in the comparability domain, and three stars in the exposure domain. Trials with seven or more cumulative stars were considered to be of high quality, those with six stars of moderate quality, and those with <6 stars of low quality (13). For RCTs, risk of bias was assessed using the Cochrane Collaboration Risk of Bias Assessment tool, which included the following seven items: random sequence generation, allocation concealment, blinding of participants and personnel, blinding of outcome assessment, incomplete outcome data, selective reporting, and others (bias due to vested financial interest and academic bias) (14). If a trial was found to have one or more of the items associated with a high or unclear risk of bias, this trial was classified as high risk. If the two authors disagreed on their assessment, they consulted the third or fourth author. Eventually, we reached a consensus (14).

Two authors were responsible for extracting the following information: (1) authors, (2) publication year, (3) study designs, (4) country of publication, (5) total number of participants in each study, (6) percentage of male, (7) mean age of all the participants, (8) procedures that the participants underwent, (9) the volatile anesthetic in patients underwent GA, (10) the anesthetic method of RA, (11) number of patients with and without POD, (12) methods of POD assessment, and (13) the follow-up time. Another three authors were responsible for adjusting data discrepancies.

The sole aim of this meta-analysis was to determine whether different anesthesia methods were associated with POD incidence in adult patients undergoing surgery.

Review Manager version (RevMan) 5.3 (Cochrane collaboration, Oxford, UK) was used to perform statistical analyses. We assessed the heterogeneity of included studies using the values of I² and the Mantel–Haenszel chi-square test (p-value for heterogeneity). The values of I² <40%, I² = 40–60%, and I² > 60% indicated low, moderate, and high heterogeneity, respectively (15). If we identified I² > 50% or a p-value for heterogeneity <0.1, we used a random-effect model to analyze the data. Conversely, if we identified I² <50% or a p-value for heterogeneity ≥ 0.1, we used a fixed-effect model to analyze the data (16). The dichotomous outcomes were presented as odds ratios (OR) with 95% confidence intervals (CI). The statistical tests were two-sided, and overall effects with a p-value for effect < 0.05 were considered to exhibit significant differences.

We conducted a sensitivity analysis to address high heterogeneity (I² > 50%) through the methods of one-by-one article removal. Lastly, we performed the subgroup analyses according to study designs (retrospective and prospective), male percentage (≥50 and <50%), age gaps (≥80, 70–80, 60–70, and <60 years), and anesthesia methods (neuraxial anesthesia and PNB) in RA group to observe if these risk factors could affect the outcome.

Figure 1 shows the PRISMA flow chart for our screening process. We obtained 120 trials from Pubmed, 405 from Embase, and 89 from Cochrane Library. We removed 367 duplicate trials and excluded 221 trials at the title-and-abstract review stage based on our exclusion criteria. We excluded five trials at the full-text review stage, including three without a specific number of patients with POD, two with all included patients who underwent general anesthesia. Eventually, our search strategy yielded 21 relevant trials with a total of 1,702,151 patients (Figure 1) (17–37).

Tables 1, 2 present the basic characteristics of the included studies. There were nine retrospective (17, 18, 22, 25, 28–31, 35), and 12 prospective (19–24, 26, 27, 32–34, 36, 37) articles, including seven randomized and non-randomized controlled trials, in the prospective studies (20, 23, 24, 32–34, 37). Male accounted for 50% or more of all patients in 4 trials (17, 22, 23, 25). The mean or median age of all included patients was 80 years or older in five trials (20, 26, 31–33), 70–80 years in seven trials (21, 28, 29, 34–37), 60–70 years in seven trials (17, 18, 23–25, 27, 30), and younger than 60 years in two trials (19, 22). Except that the patients underwent lower limb vascular surgery in two articles (23, 25) and uncertain surgeries in one article (19), the patients underwent orthopedic surgery in the other trials. One trial included the patients undergoing simple urgent surgery (29), and five trials simple planned surgery (19, 21, 24, 27, 30). Three articles did not provide information on preoperative cognitive or neuropsychological assessment (23, 28, 35). The patients in the RA group received the sole PNB in only two trials (17, 29), and underwent PNB or neuraxial anesthesia in another two trials (25, 35), while neuraxial anesthesia was the sole anesthesia in RA patients in the other 17 enrolled trials. The follow-up time was hospital stay after surgery in 13 trials (18, 22, 25–33, 35, 36), 1 year in one trial (23), 30 days in one trial (17), 1–7 days in one trial (20), 1–5 day in one trial (34), 1–4 days in one trial (37), 1–3 days in two trials (21, 24), and 1 day in one trial (19). The methods of POD identification included antipsychotic drug use (17, 18, 22, 30), Confusion Assessment Method for Intensive Care Unit (CAM-ICU) or CAM (19, 21, 36, 37), DSM-IV criteria (24, 27), NEECHAM confusion scale (25), change in mental state (23), and Pittsburgh Agitation Scale (26). Seven articles did not provide a specific diagnostic method of POD (20, 28, 29, 31–33, 35), and a combination of CAM with DSM-IV criteria (34).

We used NOS to assess the risk of bias in observational studies (retrospective and prospective), and all 14 trials obtained seven stars or more, indicating high quality (Supplementary Table 2) (17–19, 21, 22, 25–31, 35, 36). We used the Cochrane Collaboration Risk of Bias Assessment tool to assess the risk of bias in RCTs. Due to the considerable procedural difference in GA and neuraxial anesthesia or PNB, it is extremely difficult to conduct blindness in participants. Therefore, the performance bias was high risk in all included RCTs and non-RCTs. The included studies clearly assessed random sequence generation (three studies-42.9%), allocation concealment (four studies-57.1%), blinding of participants (0%), blinding of outcome assessment (seven studies-100%), incomplete outcome data (seven studies-100%), and selective outcome reporting (seven studies-100%), and the other bias (three-42.9%) (Supplementary Figures 1, 2) (20, 23, 24, 32–34, 37).

The pooled result using random-effects model with OR demonstrated significant difference in POD incidence between patients with GA and RA (OR = 1.15, 95% CI: [1.02, 1.31], I² = 83%, p for effect = 0.02) (Figure 2). I² = 81% and the funnel plots demonstrated the considerable heterogeneity of included trials (Figure 3A). The sensitivity analysis was performed to solve the high heterogeneity by the method of one-by-one literature removal and found that six trials were the main sources of heterogeneity (Figure 3B) (18, 19, 24, 30, 35, 36). We conducted post-hoc meta-analysis for the remaining articles using a fixed-effects model with OR, and the pooled result was not consistent with that prior to sensitivity analysis (OR = 0.95, 95% CI: [0.83, 1.08], I² = 13%, p for effect = 0.44) (Figure 4). Additionally, we excluded the articles without the information on preoperative cognitive or neuropsychological assessment and did not obtain the statistical difference in POD incidence between patients with GA and RA (OR = 1.12, 95% CI: [1.00, 1.25], I² = 80%, p for effect = 0.05) (Figure 5).

The subgroup analyses according to study designs, male percentage (≥50 and <50%), mean (or median) age gaps (≥80, 70–80, 60–70, and <60 years), and anesthesia methods in RA group (neuraxial anesthesia and PNB) demonstrated the significant difference in retrospective articles (OR = 1.23, 95% CI: [1.08, 1.39], p for effect = 0.001) (Figure 6), male percentage <50% (OR = 1.26, 95% CI: [1.09, 1.46], p for effect = 0.002) (Figure 7), age gap between 60 and 70 years (OR = 1.20, 95% CI: [1.07, 1.35], p for effect = 0.002) (Figure 8), and neuraxial anesthesia (OR = 1.15, 95% CI: [1.01, 1.31], p for effect = 0.03) (Figure 9). However, we did not obtain statistical difference in the subgroups in prospective studies (OR = 0.91, 95% CI: [0.55, 1.49], p for effect = 0.70) (Figure 6), male percentage ≥ 50% (OR = 0.96, 95% CI: [0.82, 1.12], p for effect = 0.60) (Figure 7), age gaps ≥ 80 years (OR = 0.98, 95% CI: [0.61, 1.57], p for effect = 0.93), 70–80 years (OR = 1.93, 95% CI: [0.66, 5.60], p for effect = 0.23), <60 years (OR = 0.67, 95% CI: [0.26, 1.71], p for effect = 0.40) (Figure 8), and PNB group (OR = 1.64, 95% CI: [0.70, 3.87], p for effect = 0.26) (Figure 9).