The authors screened MEDLINE and Embase from 2000 to September 2026 and identified 196 eligible primary articles reporting clinical AI in liver transplantation. Screening began with 2,193 records and, after deduplication and full-text review, yielded the final corpus of 196 studies. Publication activity accelerated markedly from 2020 onwards, with 83.2% of included articles published since 2020 and a steady year-on-year rise through 2025; the 2026 data include records only to September.
Data sources were geographically concentrated. The United States produced the largest share of studies (n = 72, 36.7%), followed by China (n = 36, 18.4%). The United Kingdom accounted for nine studies (4.6%). Bibliometric analysis of the subset indexed in Scopus highlighted recurrent institutional contributors — for example, several institutions appeared repeatedly among author affiliations — but the review notes that 19 of the included articles were not indexed in Scopus and thus excluded from that specific bibliometric subset.
Nearly all studies were retrospective: 188 of 196 (95.9%). Multi-centre datasets were uncommon; 25 studies (12.8%) drew on data from multiple centres, while 60 studies (30.6%) used national registry data and the remainder were single-centre. External validation was limited: only 27 studies (13.8%) tested model performance on populations unrelated to their training data. One study among the entire corpus performed a prospective evaluation using a randomized controlled trial design; the review indicates that no other prospective randomized evaluations were reported.
The authors partitioned included studies according to the perioperative transplant pathway. The distribution was as follows: pre-transplant assessment (21.9%), organ retrieval and graft assessment (9.7%), intraoperative tasks (3.6%), short-term post-operative outcomes within 90 days (16.8%), and long-term post-operative complications (45.9%). A small set of studies (2.0%) focused explicitly on NLP applications, mostly extracting concepts from electronic health records for liver transplant cohorts.
Four studies evaluated AI for transplant candidacy assessment, ranging from qualitative patient-opinion research to neural network analyses. One qualitative study asked patients their views on a hypothetical AI decision-support tool for listing and found mixed trust. A preprint used a large language model to simulate multidisciplinary selection committee roles and reported high accuracy for identifying contra-indications and 6-month mortality predictions, but it missed important prognostic factors such as limited social support and severe cardiorespiratory disease. Another study applied neural-network analysis to ECG tracings to predict post-transplant atrial fibrillation and left-sided heart failure; reported discriminatory performance in that work was modest, with AUCs around 0.69 for each outcome.
Seventeen articles applied machine learning to allocation decisions. Several groups developed neural-network approaches intended to outperform the MELD score for short-term mortality prediction. Individual studies reported improvements in area under the receiver operating curve compared with MELD by variable margins; some authors also reported shifts in prioritisation that affected subgroups such as women and patients with specific liver diseases. The review notes these allocation-focused models varied in design and validation and, in many cases, lacked external testing.
Nineteen articles addressed organ retrieval, typically using AI to assess donor graft quality. Intraoperative work was limited (seven articles), with studies predicting needs such as transfusion or using computer vision to identify anatomical structures. Short-term post-operative studies (33 articles) targeted outcomes within 90 days — including mortality, graft failure, acute kidney injury, and sepsis. The largest single category was long-term post-operative complications (90 articles), which addressed mortality, graft loss, tumour recurrence after oncologic transplants (predominantly hepatocellular carcinoma), renal failure, infection, and cardiovascular disease linked to immunosuppression.
Across dosing studies, machine learning models were also developed to predict tacrolimus dosing, reflecting interest in medication management after transplant.
Four studies (2.0% of the corpus) applied NLP methods. These chiefly extracted structured concepts from electronic health records for liver transplant populations. Examples included a study using an LLM to quantify psychosocial risk factors and social determinants of health in listing decisions, and other NLP projects aimed at transplant evaluation tasks. The review mapped these studies separately because their primary methodological approach differed from structured-data machine learning pipelines.
The review team observed consistent limitations across the literature: heterogeneous and inconsistent reporting, dominance of retrospective designs, limited external validation, and fragmented datasets concentrated in single centres or national registries. Only a small minority of studies evaluated tools prospectively in real-world clinical workflows; randomized prospective evaluation was reported in a single study. Investigators argued that progress toward clinically deployable tools will require harmonised, higher-granularity datasets, standardised reporting practices, and prospective evaluation integrated within transplant workflows. The authors also emphasized the need for interpretability frameworks — for example, Explainable AI approaches such as SHAP are discussed — to build clinical trust when moving from model development to deployment.
Limitations of the scoping review itself were handled according to PRISMA-ScR guidance; the team did not perform a formal risk-of-bias assessment given study heterogeneity. The protocol was pre-registered on the Open Science Framework, and supplementary files provide complete bibliographic and analytic details. The final recommendation from the review team is pragmatic: harmonize datasets, increase external validation, and prioritize prospective clinical evaluation to determine whether AI tools can safely and usefully integrate into liver transplantation practice.