The review was conducted and reported in accordance with the PRISMA 2020 statement [21] and updated the protocol of the reference meta-analysis [19]; it was prospectively registered (PROSPERO CRD420261404939) and additionally followed the MOOSE recommendations for meta-analyses of observational studies. The PICO framework comprised adult (≥18 years) recipients of a liver graft from Maastricht-III controlled donation after circulatory death (population); graft recovery by in situ NRP, whether abdominal (A-NRP) or thoracoabdominal (TA-NRP) (intervention); and two comparators analyzed separately—non-NRP cDCD procurement (super-rapid recovery with static cold storage, with or without ex situ HOPE or NMP) and donation after brain death (comparators). Primary outcomes were ischemic cholangiopathy, primary non-function and recipient death; secondary outcomes were graft loss, early allograft dysfunction (Olthoff criteria [22]), hepatic artery thrombosis, other biliary complications (anastomotic strictures and leaks), length of stay and graft utilization.

Eligible studies were randomized trials or comparative cohort studies (prospective or retrospective) reporting at least one primary outcome for NRP-cDCD versus non-NRP cDCD and/or DBD, in adults, published in English or Spanish, with no start-date restriction. We excluded uncontrolled (uDCD) donation without disaggregable cDCD data, studies without a comparator, purely technical descriptions, animal studies, case reports, and conference abstracts with insufficient data, and—consistent with the reference review—grafts from jurisdictions imposing a mandatory no-touch (stand-off) period exceeding 5 min (for example, Italy, 20 min). When cohorts overlapped, the most complete or most recent publication was retained and the decision documented.

Five databases were searched: MEDLINE (via PubMed), Embase, Scopus, the Cochrane Library (CENTRAL) and the Web of Science Core Collection, complemented by hand-searching the reference lists of recent reviews and of the included studies. The strategy combined three concept blocks with the Boolean AND—(i) normothermic regional perfusion, (ii) circulatory-death donation, and (iii) liver and transplantation—deliberately omitting comparator and outcome blocks to preserve sensitivity, with both resolved at screening; the term “ECMO” was admitted only when co-occurring with circulatory-death or donation terms, to retrieve the foundational NRP literature without the noise of cardiorespiratory-support ECMO. Records were managed in Rayyan, de-duplicated, and screened in duplicate at the title/abstract and full-text levels, with disagreements resolved by an arbiter. The full search string for each database and the per-database record counts are provided in the Supplementary Material.

Two reviewers extracted data independently using a standardized, piloted form that captured study characteristics (design, period, center or registry, country), donor data (age, NRP type and duration, ischemia times), recipient data (age, MELD), and every predefined outcome with its original definition. Extraction was subjected to multichannel verification—triangulation of figures, tables and text; verification of column order against reported percentages (several primary studies list the comparator or DBD arm in the first column); recomputation of every percentage from raw counts; 1:1 reconciliation of sample sizes against the previous meta-analysis; and a second independent extraction. Ischemic cholangiopathy was defined as non-anastomotic biliary strictures in a graft with a patent hepatic artery; “other biliary complications” as anastomotic strictures plus leaks, excluding ischemic cholangiopathy and hepatic artery thrombosis. Risk of bias was assessed independently by two reviewers with ROBINS-I [23], with disagreements resolved by an arbiter and the results displayed as a traffic-light plot; no randomized trial was identified. The certainty of evidence for each outcome was rated with GRADE [24]. As a final quality-control step, an expert clinician performed a final clinical reading of the extracted dataset and of the complete manuscript to confirm the clinical plausibility and internal consistency of all outcomes and conclusions.

The two comparisons were analyzed independently, and abdominal and thoracoabdominal NRP were grouped as “NRP,” as in the reference review. Dichotomous outcomes were pooled as risk ratios (RR) using DerSimonian-Laird random-effects models [25], cross-checked with REML and with Hartung-Knapp; a 0.5 continuity correction was applied only to studies with a zero cell. For rare-event outcomes the Mantel-Haenszel estimator without continuity correction was preferred, with the Peto odds ratio in support, per Cochrane guidance [26]; this mattered for ischemic cholangiopathy versus DBD, where the inverse-variance estimate (RR 1.21) was unstable because of multiple double-zero studies, whereas the Mantel-Haenszel (0.70), Peto (0.70), crude (0.79) and drop-zero (0.91) estimates were concordant and non-significant. Recipient death and graft loss were pooled as RR from counts, too few hazard ratios being available to pool, with individual hazard ratios described narratively. Pooled incidences used a logit-binomial random-effects model, which is more robust than the Freeman-Tukey transformation under high heterogeneity [27]; the latter is reported only as a cross-check. Heterogeneity was summarized with I², τ² and 95% prediction intervals computed with the t-distribution (k−2 degrees of freedom) [28], and continuous outcomes were converted from medians to means by the method of Wan et al. [29]. Prespecified sensitivity (fixed versus random effects, Mantel-Haenszel and Peto, drop-zero, and exclusion of the salvage study and of higher-risk studies) and subgroup (comparator type, region, and use of ex situ machine perfusion) analyses were performed, complemented by leave-one-out and Baujat influence diagnostics. To address potential confounding by ex situ machine perfusion (normothermic machine perfusion [NMP] or hypothermic oxygenated perfusion [HOPE]), an additional analysis restricted the NRP versus non-NRP comparison to grafts managed without ex situ perfusion in either arm; on full-text re-review, one study whose comparator was entirely NMP [11] and one in which dual HOPE was used in both arms [12] were excluded, and a three-arm study contributed through its static-cold-storage arm only, with its NMP arm removed [9]. Publication bias was not formally tested because every analysis included fewer than 10 studies [26]; contour-enhanced funnel plots are provided for completeness (Supplementary Figure S2). Overlapping cohorts were resolved with predefined hierarchical rules applied to each outcome-by-comparison pair: each study competed only within its comparison; national registries served as anchors; the 11 previously included studies were retained; and the new anti-duplication was applied chiefly to the 9 added studies. For GRADE, observational studies began at low certainty and were downgraded for risk of bias, inconsistency (I² > 60%) or imprecision (a confidence interval crossing 1), and upgraded for a large (RR ≤ 0.55) or very large (RR ≤ 0.20) effect, applied cautiously given possible residual confounding. All analyses were reproduced independently in R/metafor and in Python with exact agreement; the complete statistical and anti-duplication decisions, and the per-study data (Supplementary Table S2), are reported in the Supplementary Material.

Of 1,573 records identified, 858 duplicates were removed, 715 were screened, 118 full texts were assessed and 20 studies were included (11 from the previous meta-analysis by Mastrovangelis et al. [19] and 9 published subsequently); 98 reports were excluded with documented reasons (Figure 1) [19]. The included studies comprised 1776 NRP-cDCD liver recipients across the 20 included studies (1,512 in the 16 studies contributing to the meta-analysis) and their non-NRP and DBD comparators across Europe (predominantly Spain) and North America, and ranged from single-center series to national registry cohorts (Table 1). All 20 studies contributed to the qualitative synthesis, whereas 16 contributed to the quantitative synthesis (meta-analysis); the remaining four were retained descriptively but excluded from pooling—two US registry analyses in which NRP could not be isolated or incidence was unreported [14, 15], one Cambridge cohort overlapping an included study [32], and one technique-comparison study [31]. Throughout, non-NRP cDCD comparators comprised the conventional pathway of super-rapid recovery followed by static cold storage (SRR/SCS) unless a study used ex situ machine perfusion—normothermic (NMP) or hypothermic oxygenated (HOPE)—which is specified in Table 1.

On ROBINS-I, no study was at low risk of bias; 7 were at moderate and 13 at serious risk, driven chiefly by confounding (12/20 serious—inherent to non-randomized designs) and selective reporting (9/20). The two reviewers reached concordant overall judgements (Figure 2). A domain-level summary is provided in Supplementary Figure S1.

Between four and seven studies contributed to each pooled comparison—all fewer than ten, which precluded formal small-study (publication-bias) testing [26]. Findings are presented by endpoint (a risk ratio below 1 favours NRP); the pooled estimates are summarized in Figure 3, per-study forest plots appear in Supplementary Figure S4, GRADE summaries in Table 2 and pooled incidences in Table 3.

Four to eight studies contributed to each endpoint—again fewer than ten, precluding small-study testing—and heterogeneity was low (I² ≤ 4%) for every endpoint except early allograft dysfunction (I² = 34%); no outcome differed significantly and every estimate was of very low certainty by GRADE (Table 2; per-study forest plots in Supplementary Figure S5). Incidences are given as NRP versus DBD.

The NRP-versus-non-NRP comparison was repeated after excluding all grafts managed with ex situ perfusion: Mohkam et al. (comparator entirely NMP) and Schurink et al. (dual HOPE in both arms) were removed, and the three-arm cohort of Gaurav et al. contributed through its static-cold-storage arm only. Every effect persisted in the same direction and remained statistically significant, by endpoint: ischemic cholangiopathy RR 0.14 (0.08–0.26; 5 studies; I² = 0%); graft loss 0.41 (0.32–0.53; 4 studies; I² = 0%); recipient death 0.47 (0.35–0.64; 3 studies; I² = 0%); other biliary complications 0.54 (0.37–0.79; 4 studies; I² = 30%); hepatic artery thrombosis 0.48 (0.28–0.83; 4 studies; I² = 0%); primary non-function 0.49 (0.27–0.92; 5 studies; I² = 0%); and early allograft dysfunction 0.58 (0.39–0.86; 5 studies; I² = 78%) (Supplementary Table S3; Supplementary Figure S3). Heterogeneity was therefore unchanged by the exclusion, and the pooled incidences and GRADE certainty reported above apply. The NRP-versus-DBD comparison was inherently free of ex situ perfusion. Estimates were also robust to alternative estimators (fixed versus random effects, Mantel-Haenszel, Peto) and to exclusion of the salvage study and of higher-risk studies (Supplementary Table S1). All estimates were reproduced in R/metafor and in Python with exact agreement.