The present systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA 2020) guidelines (9). The study methods, inclusion criteria, and analysis plan were defined a priori before the literature search and data extraction.

MEDLINE/PubMed (pubmed.ncbi.nlm.nih.gov/), Embase (https://www.embase.com/), Cumulative Index to Nursing & Allied Health Literature (about.ebsco.com/products/research-databases/cinahl-database), the Cochrane Library (https://www.cochranelibrary.com/), PsycINFO (https://www.apa.org/pubs/databases/psycinfo), Web of Science (https://mjl.clarivate.com/home), and Scopus (https://www.scopus.com/sources) were searched for relevant literature from inception until May 2025 without regard to date or language. Search terms (Data S1) for ‘cardiac surgery’ [such as coronary artery bypass grafting (CABG), valve, or aortic surgery], ‘postoperative cognitive dysfunction’, and ‘Mini-Mental State Examination’ combined keywords and MeSH/Emtree terms. The reference lists of relevant reviews, dissertations, and conference proceedings were also searched.

Titles, abstracts, and full texts were independently screened by two reviewers against the predefined inclusion criteria. Inter-rater agreement during the screening process was assessed using Cohen's κ statistic after the initial title-abstract screening stage to quantify reviewer consistency. Any discrepancies were resolved through discussion or consultation with a third reviewer.

The inclusion criteria for the studies were as follows: i) Adult patients (age ≥18 years) undergoing cardiac surgery (CABG, valve, or aortic surgery); ii) cognitive function assessed using the MMSE both before and after surgery; iii) sufficient data reported (mean and SD or change scores) to calculate the standardized mean difference (SMD) and iv) original clinical studies (randomized trials or observational cohort studies).

Studies that did not use the MMSE for cognitive assessment, case reports, reviews, pediatric populations, or non-cardiac procedures (thoracic surgery, cardiac transplantation, or transcatheter valve replacement) were excluded. No restriction on publication year was applied.

Because included studies reported postoperative MMSE assessments at different follow-up intervals, data were extracted from the longest postoperative follow-up reported in each study when multiple time points were available. When studies reported several clinically distinct follow-up windows (early and late postoperative assessments), these were considered separate comparisons in the meta-analysis provided that independent summary statistics were available.

For subgroup analyses, patients were categorized as having POCD or not (non-POCD) according to the definitions reported in the original studies. In the perioperative literature, POCD is generally defined as a decline in cognitive performance relative to the patient's preoperative baseline measured using neuropsychological tests (2). However, there is currently no universally standardized diagnostic threshold, and studies typically employ different statistical approaches to determine cognitive decline (decreases of ≥1-2 SD from baseline performance, percentage decline in test scores, or composite indices derived from multiple cognitive tests) (1). POCD classification was therefore based on the criteria used in the original publications, which typically relied on postoperative deterioration in cognitive test performance, assessed using the MMSE or broader neuropsychological test batteries, relative to baseline values. Because diagnostic thresholds, testing batteries, and timing of postoperative assessments vary across studies, study-specific definitions were used rather than attempting to retrospectively standardize the classification using a single cutoff, which would not have been feasible with the available summary data (9).

Data were independently extracted by two investigators using a pre-piloted standardized form. Extracted variables included study design (prospective or retrospective; single-center or multicenter), country and setting, sample size, patient demographic characteristics (age and sex distribution), type of cardiac surgery, and timing of MMSE assessment. For each eligible comparison, the number of participants and the mean MMSE scores at baseline and postoperative time points or the reported change in MMSE when available were extracted.

When information was missing or unclear, attempts were made to contact the study authors for clarification. Any discrepancies in data extraction were resolved through discussion until consensus was reached. In addition, key study characteristics, including differences in surgical procedures, follow-up intervals, and study design, were recorded to facilitate interpretation of heterogeneity across studies.

A total of two reviewers independently evaluated methodological quality using the Critical Appraisal Skills Programme (CASP) checklist for cohort studies (8). The CASP tool includes 12 items evaluating study validity, methodological rigor, and relevance across domains such as selection bias, measurement of outcomes, confounding control, and applicability. Each item was scored as ‘yes’, ‘no’, or ‘unclear’. Studies were categorized according to the number of criteria fulfilled as follows: ≥10, good quality; 7-9, fair quality, and ≤6, low quality. Disagreements were resolved through discussion until a consensus was reached.

All meta-analytic computations were conducted using Comprehensive Meta-Analysis software, version 4 (Biostat, Inc.). The effect size for postoperative change in cognitive performance was estimated using SMD (Hedges' g) with corresponding 95% confidence intervals (CIs). Effect sizes were calculated from the reported preoperative and postoperative mean MMSE for each cohort. This approach standardizes the mean difference using the pooled SD of the two measurements and therefore does not require explicit imputation of the within-subject association between pre- and postoperative scores. Because the primary studies did not consistently report change-score variances or pre-post association, no fixed correlation coefficient was assumed in the analysis. Where only medians with ranges or interquartile ranges were reported, means and SD were estimated using established conversion formulas, and SD was back-calculated from reported standard errors or CI when necessary (17).

When primary studies reported >1 eligible comparison (distinct postoperative follow-up time points or separate patient strata), each comparison was treated as an independent effect size when it represented a clinically distinct contrast with separate summary statistics (means and SD). This approach allows the meta-analysis to incorporate all available evidence while preserving the specific clinical context of each comparison.

To account for anticipated between-study heterogeneity, a random-effects model was prespecified. Subgroup analyses were conducted using a mixed-effects approach, in which random-effects models were applied within each subgroup, while differences between subgroups were assessed using a fixed-effect Q_between statistic. Between-study variance (τ²) was estimated using the DerSimonian-Laird method. Sensitivity analyses were performed using the Paule-Mandel, Sidik-Jonkman, and restricted maximum-likelihood estimators, and the Hartung-Knapp-Sidik-Jonkman adjustment was applied to the pooled variance. Statistical heterogeneity was quantified using Cochran's Q statistic (P<0.10), I², τ², and τ. In addition to 95% CI, 95% prediction interval (PI) was calculated for each pooled SMD to estimate the range within which the true effect of a comparable future study was likely to lie.

Robustness of the findings was assessed through several sensitivity analyses, including leave-one-out influence analysis, Baujat plots to identify influential comparisons, comparison with a fixed-effect model, exclusion of studies at higher risk of bias, and reanalysis using alternative effect-size metrics (raw MMSE point change). Potential publication bias and small-study effects were evaluated by visual inspection of funnel plots of SMDs against their standard errors, Egger's weighted regression test, and Begg's rank-correlation test (two-tailed P<0.1) and Duval and Tweedie's trim-and-fill method when asymmetry was detected to estimate the number of potentially missing studies and generate an adjusted pooled effect size. In addition, Rosenthal's and Orwin's fail-safe N statistics were calculated to estimate how many null studies would be required to overturn the overall result.

P<0.05 was considered to indicate a statistically significant difference, and all tests were two-tailed unless otherwise specified (P<0.10 for Cochran's Q and funnel-plot asymmetry tests). CI, PI, and test statistics were based on the t-distribution with k-1 degrees of freedom (df).

The systematic database search identified 1,564 records. Following removal of 312 duplicates, 1,252 titles and abstracts were screened, of which 1,160 were excluded for not meeting the criteria (non-cardiac surgery, pediatric populations or absence of a cognitive outcome; Fig. 1). In total, 72 full-text articles were assessed for eligibility and 64 were excluded for the following reasons: No MMSE data (n=21), review, editorial or conference abstract (n=13), ineligible study design (no original data; n=12), mixed surgical cohorts without extractable cardiac subgroup data (n=10) and duplicate or overlapping populations (n=8). In total, eight studies met the eligibility criteria and were included in the qualitative synthesis (10-16,18), yielding 14 independent comparisons (distinct patient groups, time points, or surgical strata).

All eight studies (10-16,18) were prospective single-center cohorts published between 2005 and 2024; four were conducted in Japan (10-13), with one each in India (16), Sweden (14), China (18), and Malaysia (15). A total of 867 patients (sample size range, 28-280) undergoing on-pump CABG or valve surgery were analyzed. Mean or median age was between 60 and 65 in the Japanese (10-13) and Swedish (14) cohorts and ~50 years in the Chinese valve cohort (18); the overall proportion of male patients was ~70%, but was lower (39%) in the study by Zhang et al (18). All studies assessed MMSE scores pre-operatively and at ≥1 early postoperative time point [within the first postoperative week or at hospital discharge in six studies, 2 weeks in Maekawa et al (13), and 6 weeks in Yazit et al (15)]. Several cohorts also included longer follow-up assessments up to 6 months (Table I) (10). Postoperative cognitive assessment spanned the very early phase (≤1 week), an intermediate window (2 weeks), and a later window (6 weeks-6 months), yielding 14 independent comparisons included in the meta-analysis. Baseline MMSE scores ranged from 26 to 29 points, indicating generally preserved preoperative cognitive function across the populations.

Cardiac surgery was associated with a significant decline in MMSE (Hedges g=-0.60, 95% CI -0.85 to -0.35, P<0.001; Fig. 2). For clinical interpretability, the pooled standardized effect size was translated into an approximate change in raw MMSE points. Using the typical baseline SD reported across the included cohorts (~3 MMSE points), the pooled effect size (g=-0.60) corresponded to an estimated decline of 1-2 points on the 30-point MMSE scale. Between-study heterogeneity was notable (τ²=0.178, τ=0.422; Q=116.4, df=13, P<0.001; I²=88.8%). The 95% PI (-1.56 to 0.36) spanned both negative and slightly positive values, indicating that while most cohorts are expected to show cognitive decline, the inclusion of values above zero suggests that small improvements remain possible in some populations. A fixed-effect model yielded a smaller but significant estimate (g=-0.35, 95% CI -0.43 to -0.27, P<0.001).

Using the prespecified mixed-effects model, patients who developed POCD (eight comparisons) showed a significant decline in MMSE scores (Hedges' g=-0.893, 95% CI -1.258 to -0.527; P<0.001; 95% PI -2.114 to 0.328), whereas patients without POCD (six comparisons) showed a non-significant change (g=-0.274, 95% CI -0.631 to 0.084; P=0.134; 95% PI, -1.517 to 0.970; Fig. 3). Heterogeneity was high in both subgroups (I²=88-89%), but the between-group test showed a significant difference in effect sizes (Q_between=9.08, df=1, P=0.003). Across all 14 comparisons, the pooled mixed-effects estimate was g=-0.576 (95% CI -0.832 to -0.321; P<0.001; 95% PI -1.539 to 0.386), indicating that the overall postoperative decline was driven primarily by the subgroup with clinical POCD.

Sequential leave-one-out sensitivity analysis showed that omission of any single comparison changed the pooled standardized mean difference by <0.08. In every iteration, the pooled estimate remained negative and significant (all P≤0.002), indicating that the overall result was not driven by any individual comparison (Fig. 4). Taken together, these findings support the robustness of the conclusion that cardiac surgery was associated with a significant postoperative decline in MMSE score.

For the primary random-effects model (14 comparisons), the 95% PI ranged from -1.56 to 0.36 SMD units (with the null value falling within the interval), indicating that most future studies of similar design are likely to observe a postoperative decline in MMSE, although a small improvement remains possible in certain cohorts (Fig. 5). The subgroup analyses supported this pattern. Specifically, cohorts consisting primarily of patients with POCD are expected to report a marked cognitive decline, whereas studies including only patients without POCD may plausibly observe outcomes ranging from a modest decline to a modest improvement.

The range of the PI reflects substantial between-study heterogeneity; its predominantly negative distribution supports the overall conclusion that cognitive performance, as measured by the MMSE, generally declined following on-pump cardiac surgery.

Visual inspection of the funnel plot (Fig. 6) suggested an absence of large studies reporting null effects, which was supported by statistical tests. Egger's regression test yielded an intercept of -3.69 (95% CI -6.81 to -0.58; P=0.024), and Begg and Mazumdar's rank-correlation test indicated significant asymmetry (Kendall's τ=-0.43, z=2.14; two-tailed P=0.033). Despite this asymmetry, Duval and Tweedie's trim-and-fill procedure did not identify any potentially missing studies (k_trimmed=0), and the bias-adjusted random-effects estimate was unchanged compared with the original pooled estimate (Hedges' g=-0.598, 95% CI -0.847 to -0.349).

Rosenthal's classic fail-safe N indicated that 431 additional null studies would be required to increase the pooled P-value to >0.05, exceeding the conventional tolerance threshold (5 x k + 10=80). Together, these analyses suggest that although small-study effects may be present, they were unlikely to materially alter the conclusion that cardiac surgery is associated with a significant postoperative decline in MMSE score.

To explore sources of the large between-study heterogeneity (τ²=0.178; I²=~89%), mixed-effects meta-regression analysis was performed, including five a priori clinical moderators: Mean age, cardiopulmonary bypass (CPB) time, study-level prevalence of hypertension and diabetes mellitus, and male sex. The simultaneous (multivariable) model was significant (Q_model=11.22, df=5, P=0.047) and accounted for part of the between-study heterogeneity, although only CPB time was an independent predictor (Table II). Specifically, each 1 min increase in CPB duration was associated with an additional decline of 0.014 SD units in MMSE scores (β=-0.014±0.006, Z=-2.27, P=0.023), corresponding to ~0.9 SD greater decline for 1 h additional bypass time.

When each moderator was entered separately (univariate models; Table III), CPB time showed the strongest association and explained ~50% of the between-study variance (R²_analog=0.50; τ² decreased from 0.198 to 0.100; Q=18.68, P=0.0001), whereas age explained ~29% of the heterogeneity (Q=7.23, P=0.007). The remaining covariates contributed little to explaining the variability across studies.

Diagnostic plots and jackknife influence analyses indicated that no single comparison exerted a disproportionate influence on the regression estimates (largest studentized residual=2.10; Cook's distance <1 for all points). Furthermore, no problematic multicollinearity was detected among predictors; the highest correlation was |r|=0.80 between age and hypertension, and all variance inflation factors were ≤1.34 (Fig. 7).

In total, 5/8 studies achieved good methodological quality. All of these were prospective single-center cohorts with clearly stated aims, well-described recruitment, and rigorous multivariable adjustment; for example, Kadoi et al (12) and Kadoi et al (11) used consensus neuropsychology batteries and logistic regression. Maekawa et al (13) and Shiraboina et al (16) were graded as fair, chiefly because short follow-up windows (≤2 weeks) or limited reporting made it unclear whether all relevant confounders were handled; nevertheless, both used validated cognitive outcomes (MMSE + extended batteries) and prospective designs. No study was downgraded for imprecision (every cohort reported effect sizes with either CI or full regression output), and the absence of serious loss to follow-up across studies supported internal validity. External generalizability was highest for the larger cohorts [Yazit et al (15), 188 patients; Kadoi and Goto (10), 280 patients], whereas single-surgeon pilots require cautious extrapolation. Overall, the evidence base was methodologically sound, with minor limitations concentrated in the smaller exploratory studies (Table IV).