---
title: "Artificial intelligence applications in liver transplantation: scoping review of 196 studies"
id: "plos-one-15-the-applications-of-artificial-intelligence-in-liver-transplantation-a"
canonical_url: "https://medichelpline.com/clinical-feed/plos-one-15-the-applications-of-artificial-intelligence-in-liver-transplantation-a"
content_type: "clinical_feed_article"
specialty: "Gastroenterology"
source_name: "PLOS ONE (Medicine)"
source_url: "https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361120"
published_at: "2026-10-23T14:00:00.000Z"
evidence_level: "Journal Feed"
license: "CC-BY-NC-4.0 / Informational Use"
---
# Artificial intelligence applications in liver transplantation: scoping review of 196 studies
## Provenance & Clinical Metadata
- **Canonical URL:** https://medichelpline.com/clinical-feed/plos-one-15-the-applications-of-artificial-intelligence-in-liver-transplantation-a
- **Specialty:** [Gastroenterology](https://medichelpline.com/clinical-feed/gastroenterology.md)
- **Primary Source:** PLOS ONE (Medicine)
- **Source URL:** [Original Journal Publication](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361120)
- **Published At:** 2026-10-23T14:00:00.000Z
- **Evidence Rating:** Journal Feed
## Executive GIST (TL;DR)
- The scoping review identified 196 primary articles on **artificial intelligence** applied to **liver transplantation**, sourced from MEDLINE and Embase between 2000 and September 2026. - Publications concentrated on long-term post-operative complications (45.9%) and pre-transplant decision-making (21.9%), with smaller proportions addressing organ retrieval (9.7%), intraoperative issues (3.6%), and short-term post-operative outcomes (16.8%). - The literature is heavily retrospective: 95.9% of studies used retrospective designs; only one study included a prospective randomized evaluation. - Multi-centre datasets were uncommon: 12.8% of studies used multi-centre data, 30.6% used national registries, the remainder were single-centre. - External validation was performed in 13.8% of studies; the majority reported internal-only validation or did not report external testing. - Geographically, the United States produced the largest number of studies (36.7%), followed by China (18.4%); the UK accounted for 4.6% of studies. - Methodologically, the review framed AI along a spectrum from classical statistics and rule-based systems through machine learning classifiers to deep learning and transformer-based models, and discussed Explainable AI tools such as SHAP for interpretability. - Notable application areas included improving allocation beyond **MELD**, predicting tacrolimus dosing, graft assessment at retrieval, intraoperative computer vision tasks, and extraction of EHR concepts via **NLP**. - Major limitations in the evidence base were heterogenous reporting, low external validation rates, dataset fragmentation, and few prospective or randomized evaluations; investigators recommended harmonised, higher-granularity datasets and prospective evaluation within clinical workflows.
## Clinical Analysis & Structured Key Points
The applications of artificial intelligence in liver transplantation: A comprehensive scoping review | PLOS One Browse Subject Areas ? Click through the PLOS taxonomy to find articles in your field. For more information about PLOS Subject Areas, click here . Article Authors Metrics Comments Media Coverage Peer Review Reader Comments Figures Figures Abstract Liver transplantation (LT) is a complex surgical intervention, with intricate decision-making and risks at all stages before, during and after the procedure itself. Artificial intelligence (AI) may assist the process but has seen limited use. This scoping review assessed primary articles reporting on clinical AI in LT in MEDLINE and Embase from 2000 to September 2026. 196 articles were identified: focusing on the pre-operative period before transplant (21.9%), organ retrieval (9.7%), intra-operative issues (3.6%), short-term post-operative complications (16.8%), long-term post-operative complications (45.9%) and natural language processing (2.0%). 95.9% had retrospective designs and only 12.8% had direct multi-centre data, with the rest from national registries or single centres. 13.8% of studies externally validated their AI tools. Only one study performed a prospective evaluation of their tool with a randomized controlled trial. Progress towards clinically deployable tools will require harmonised, higher-granularity datasets, consistent reporting, and prospective evaluation within real-world transplant workflows. Citation: Borakati A, Agilinko J, Wu J, Hunter S, Alhaq A, Froghi F, et al. (2026) The applications of artificial intelligence in liver transplantation: A comprehensive scoping review. PLoS One 21(10): e0361120. https://doi.org/10.1371/journal.pone.0361120 Editor: Amelia J. Hessheimer, Hospital Universitario La Paz, SPAIN Received: March 26, 2026; Accepted: October 6, 2026; Published: October 23, 2026 Copyright: © 2026 Borakati et al. This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability: All data are in the manuscript and/or supporting information files. Funding: Aditya Borakati is supported by the UK Engineering and Physical Sciences Research Council (EPSRC) [Grant reference number EP/Y035216/1] Centre for Doctoral Training in Data-Driven Health (DRIVE-Health) at King’s College London, with additional funding from the Roger Williams Institute of Liver Studies at King’s College London and a Fellowship in Clinical Artificial Intelligence from the National Health Service Digital Academy. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Competing interests: The authors have declared that no competing interests exist. Introduction Liver transplantation is a highly effective treatment for end-stage liver disease, but it is complex and carries significant risks before, during, and after the procedure. The decision to offer a transplant involves careful consideration of patient wishes, surgical fitness, prognosis, graft availability, and, in some countries, resource and financial constraints. While some livers are donated by living individuals, most grafts in the Western World are recovered from deceased donors following brainstem (DBD) or circulatory death (DCD) [ 1 ]. Unfortunately, the demand for transplants far exceeds supply, and in the USA, pretransplant mortality remains 13.4 deaths per 100 patient-years, with 12.4% of candidates removed from the liver waiting list in 2024 having died or become too sick for transplant [ 2 ]. Artificial intelligence (AI) may help address some of these challenges by supporting risk prediction and decision making. For example, to allocate these scarce organs, most countries use risk scores—such as MELD or UKELD—to prioritise patients based on predicted mortality without a transplant [ 3 ], though alternative approaches considering post-transplant survival are also debated, with the UK being the first to implement an allocation system based on overall Transplant Benefit [ 4 ]. At the time of organ retrieval, surgeons assess graft suitability based on donor status, graft appearance, and recipient condition, but there are no standardised protocols for these decisions. During surgery, choices regarding implantation technique and anatomical variations can further complicate the process. After transplantation, patients face acute complications such as haemorrhage, vascular issues, and rejection, as well as long-term risks including graft failure, infection, malignancy, renal failure, and cardiovascular disease, all of which are exacerbated by immunosuppressive therapy. In each stage of the transplantation process, AI has the potential to standardize many of the aforementioned issues including through more accurate waitlisting, graft assessment and prediction of post-operative outcomes. The stages in the liver transplantation process, along some of the long-term complications, are outlined in Fig 1 . Download: PNG larger image TIFF original image Fig 1. Stages of the Liver Transplantation process with examples of post-operative complications (some graphics in this figure were generated by generative AI software: Microsoft Copilot, Microsoft, Redmond, WA, USA, the authors grant permission for publication of this figure under a Creative Commons Attribution License (CCAL) CC BY 4.0). https://doi.org/10.1371/journal.pone.0361120.g001 Background to artificial intelligence and its evolution Over the last decade, and notably since 2022, the popularity of transformer-based large language models such as ChatGPT has resulted in a surge in artificial intelligence (AI) applications. AI has the potential to transform healthcare, by assisting clinical decisions [ 5 ], automation of tasks, and improving efficiency. Complex multifactorial clinical activities such as liver transplantation are therefore well placed to benefit from AI. AI and machine learning (ML) are often used interchangeably and have varying definitions in the literature. In this review, we define AI as an umbrella term referring to any tool designed to perform tasks that would otherwise require human intelligence. ML is considered a subset of AI involving algorithms that discover patterns and improve performance on a task through exposure to data [ 6 ]. To rigorously evaluate these technologies, however, a clear methodological stratification is required, tracking a chronological evolution from transparent statistical models to complex, “black-box” deep learning frameworks. Historically, classical statistical models and rule-based expert systems formed the earliest foundations of clinical AI. Rule-based systems rely on hardcoded, human-defined logical paths (e.g., “if-then” clinical algorithms) to automate decision-making. As data availability expanded, these gave way to data-driven, statistical approaches like logistic regression and Cox proportional hazards regression. Within this tier, advanced automated regularization techniques—such as LASSO (Least Absolute Shrinkage and Selection Operator) —are frequently classified as machine learning because they autonomously handle variable selection and model shrinkage to optimize predictive power without manual human curation. While structurally distinct from true AI, these standard approaches are frequently evaluated alongside traditional, task-specific machine learning (ML) classifiers . This tier includes models like decision trees, random forests, support vector machines (SVM), and gradient boosting architectures (e.g., XGBoost, CatBoost). These models are typically reported in the medical literature to handle structured datasets, predicting outcomes either as discrete categories (classification) or a range of continuous numbers (regression) [ 7 ]. A major evolutionary milestone occurred with the advent of deep learning (DL) . Powered by artificial neural networks (ANNs) and convolutional neural networks (CNNs), these models are inspired by biological networks with complex, multi-layered, interconnected inputs and outputs [ 8 ]. The current state-of-the-art is driven by transformer models and natural language processing (NLP) . Built on deep neural networks, transformers possess the unique ability to “focus attention” on important segments of text while identifying context across entire datasets [ 9 ]. When scaled, these transformers undergo generative pre-training (GPT) , an autonomous learning process where the model learns the underlying structure of language by predicting the next word in vast textual repositories (such as the entire public internet). These foundational steps have created large language models (LLMs) which power popular ‘chatbot’ style AI such as ChatGPT [ 10 ]. These highly advanced models can bridge multiple domains, including generating clinical narratives, reading unstructured health records, image analysis, and advanced prediction modelling. As clinical modelling has evolved along this spectrum, a critical trade-off has emerged between predictive performance and model interpretability. While classical regression models offer direct, intuitive coefficients (such as odds or hazard ratios), advanced ML and deep learning algorithms operate as “black boxes” with hidden non-linear layers that prevent clinicians from deducing why a specific prediction was made. In high-stakes environments like organ allocation or post-operative risk tracking, this lack of transparency severely hinders clinical trust. Consequently, the paradigm of Explainable AI (XAI) has emerged, using post-hoc frameworks like Shapley Additive Explanations (SHAP) to assign quantitative importance to predictor variables [ 11 ]. Understanding this methodological spectrum and the necessity of interpretability frameworks is essential to accurately map the literature. The purpose of this paper is to evaluate the state of the art of AI applications across this stratified taxonomy at each stage of the liver transplantation process and identify current evidence gaps in the literature. Methods We conducted a scoping review of the literature, searching for all published articles on the use of artificial intelligence in applications related to liver transplantation. The review was conducted in accordance with the Joanna Briggs Institute guidelines on scoping reviews [ 12 – 14 ] and reported using the PRISMA-ScR checklist [ 15 ] ( S1 File ). The protocol for the scoping review was pre-registered on the Open Science Framework: https://osf.io/tuvrz/ [ 16 ]. Eligibility criteria All primary research articles evaluating artificial intelligence tools on patients eligible for, undergoing or those who have undergone a liver transplant were included. There was no age restriction on the patient subjects. Pre-prints, in addition to published, peer-reviewed articles, were eligible for inclusion. Review articles, correspondence, and editorials were not eligible for inclusion. Non-English language articles were also excluded. Search strategy MEDLINE and Embase were searched from 2000 to September 2026, using a combination of Medical Subject Headings and free-text terms related to liver transplantation and artificial intelligence ( S2 File ). The latter database also incorporates pre-print articles from databases such as arXiv and medrXiV and these were also included. Title and abstract screening was performed on the Rayyan platform (rayyan.ai [ 17 ]), by two independent reviewers (AB and JA). Any disagreements were resolved through consensus or discussion with a third reviewer (DW). Data items and charting process One reviewer (AB) reviewed full-text publications and extracted relevant data to a pre-piloted spreadsheet in Microsoft Excel for Windows. Data extracted included study title, first author, year, country of origin of the study population, timing on the transplant process (pre-transplant, organ retrieval, intraoperative, post-operative complications) and details about the AI model if available, including the type of model. Additional information recorded included study design (retrospective vs. prospective), data source (single-centre/ multi-centre/ registry) and validation approach for the model (internal only or on external datasets). A formal risk of bias assessment was not conducted due to the heterogeneity of studies included and these are not required in a scoping review as per PRISMA-ScR guidelines [ 15 ]. Synthesis of results Eligible papers are presented through a narrative synthesis. Included papers were partitioned into the different stages of the perioperative transplant pathway, i.e., pre-transplant, organ retrieval, intraoperative, post-operative complications and the purpose of the AI tool. Each paper was assigned to a single stage; papers whose main focus was a natural language processing method were charted as a separate category. Evidence gaps were identified through discussion between the review team and summarised in text, as well as a table. Bibliometric analysis We quantified the number of included publications by year in a bar chart and generated a world map, showing the number of publications by country. In addition we conducted a simple bibliometric analysis using the Bibliometrix software [ 18 ], with default settings. Results Search results We identified 2,193 records through our database search, of which 300 were duplicates. After title and abstract screening of the remaining 1,893 records, we selected 466 records for full-text screening. This identified 196 articles that met our search criteria and were included in the final review ( Fig 2 ). Bibliographic details of all included articles are available in the S3 file . Download: PNG larger image TIFF original image Fig 2. PRISMA flowchart of inclusion and exclusion of records. https://doi.org/10.1371/journal.pone.0361120.g002 163 (83.2%) articles were published from 2020 onwards, from 2020 there is a year-on-year rise in publications to 2025; 2026 includes publications to September only ( Fig 3 ). The largest number of studies used datasets from the United States (n = 72, 36.7%), followed by China (n = 36, 18.4%). The United Kingdom produced the sixth most studies (n = 9, 4.6%). Studies using data from more than one country were counted once for each country. Fig 4 displays a map with the frequencies of included articles per country. Download: PNG larger image TIFF original image Fig 3. Number of publications on AI in liver transplantation by year of publication. https://doi.org/10.1371/journal.pone.0361120.g003 Download: PNG larger image TIFF original image Fig 4. World map showing distribution and frequency of included articles by country (figure generated in Python using Plotly; two studies using multinational registry data are not shown. The authors grant permission for publication of this figure under a Creative Commons Attribution License (CCAL) CC BY 4.0). https://doi.org/10.1371/journal.pone.0361120.g004 Bibliometric analysis In addition to the above, a bibliometric analysis of the 177 included articles indexed in Scopus (19 were not indexed and could not be analysed) is given in S4 File . This analysis notably highlights the contribution of individual institutions such as the University of Toronto, the University Health Network and the Mayo Clinic, which appeared 24, 22 and 18 times respectively in author affiliations. The most locally cited article, cited by 20 of the other included articles, was by Lau et al [ 19 ]. Study design 188 (95.9%) studies were conducted retrospectively. 25 (12.8%) studies had data from multiple centres, while 60 (30.6%) used national registry data and the remainder were single-centre. Only 27 (13.8%) studies externally validated their models on populations unrelated to their training data. Overall themes The overall themes charted in the scoping review are shown in Fig 5 . Download: PNG larger image TIFF original image Fig 5. Network diagram of themes charted in review; HCC: Hepatocellular Carcinoma. The Transplant Evaluation node under Natural Language Processing shows the one NLP study that also addressed transplant evaluation. https://doi.org/10.1371/journal.pone.0361120.g005 43 studies (21.9%), focused on the pre-transplant period. 19 articles (9.7%) focused on the organ retrieval process. These were all studies that aimed to assess the quality of the donor graft using AI techniques. 7 articles (3.6%) studied issues intra-operatively, these included predicting intra-operative transfusion requirements and identifying anatomical structures intra-operatively using computer vision. 33 articles (16.8%) studied short-term post-operative outcomes, within 90-days of transplant including mortality, graft failure, acute kidney injury and sepsis. 90 articles (45.9%) explored complications in the long-term post-operatively. These were predominantly complications such as mortality or graft loss, but also included tumour recurrence for transplants performed for oncologic purposes (pre-dominantly hepatocellular carcinoma). NB- Not all studies map to one of these themes so are not counted. A further key theme is the use of machine learning models for prediction of tacrolimus dosing. There were also four further articles (2.0%) that assessed natural language processing (NLP) methods, most often to extract concepts from electronic health records (EHRs) on datasets including liver transplant recipients [ 20 – 22 ]. Pre-transplant Transplant assessment Four studies assessed the role of AI in assessing potential recipients for their suitability for a transplant. Strauss and colleagues performed a qualitative study evaluating patients’ opinions of a hypothetical AI clinical decision support tool for listing a patient for transplant. While there was some optimism for the potential of AI to improve fairness and reduce discrimination in listing, there was concern that AI may not be trustworthy and may not act in the patient’s best interests [ 23 ]. Hasjim et al. [ 24 ] (preprint) went one step further and used a large language model to simulate four participants in a transplant selection committee: surgeon, hepatologist, cardiologist and social worker. This simulation was able to accurately identify contra-indications to transplant in 98.4% of cases and predicted 6-month mortality with 94.9% accuracy. However, there were significant errors with this model as it failed to identify significant prognostic factors such as a lack of social support or severe cardiorespiratory disease. Zaver [ 25 ] and co-workers identified AI enhanced electrocardiograms (ECGs) as an adjunct to the transplant pre-assessment, where patients are assessed for fitness for an elective liver transplant. Using a neural network on ECG tracings, they were then able to predict post-transplant atrial fibrillation and left sided heart failure, although with weak AUCs of 0.69 for each. A fourth study, by Robitschek et al. [ 22 ], used a large language model to quantify the effect of psychosocial risk factors and social determinants of health on listing decisions, and is described in the natural language processing section. Allocation policy 17 articles applied machine learning methods to prioritise patients for liver transplant. Several aimed to improve upon the MELD score, which is widely used for waiting list prioritisation and especially in the US. For example, Cooper [ 26 ] (preprint), Cucchetti [ 27 ], and Salkic [ 28 ] developed neural network models that showed modest to significant improvements in predicting 90-day mortality compared to MELD, with improvements in area under the receiver operating curve (AUC) ranging from 0.05 to 0.16. Some models [ 26 ] also reported increased fairness in allocation, with women and patients with primary sclerosing cholangitis prioritized compared to MELD. However, these studies were
## Related Clinical Research

- [Renal Angina Index predicts severe AKI in Vietnamese PICU: prospective validation](https://medichelpline.com/clinical-feed/pubmed-42909503.md) (DOI: 10.1007/s00467-026-07563-5)
- [When to stop KRT in critically ill children: evidence and practical pediatric nephrology guidance](https://medichelpline.com/clinical-feed/pubmed-42909501.md) (DOI: 10.1007/s00467-026-07588-w)
- [Macrophage-Derived 25-Hydroxycholesterol Potentiates Statin-Associated Acute Kidney Injury](https://medichelpline.com/clinical-feed/biorxiv-2-macrophage-derived-25-hydroxycholesterol-drives-statin-associated-acute-kidney.md)
- [Soluble urokinase plasminogen activator receptor (suPAR) across the kidney–cardiometabolic continu](https://medichelpline.com/clinical-feed/frontiers-in-immunology-17-soluble-urokinase-plasminogen-activator-receptor-across-the-kidney.md)
- [3m-SMEF Protects Against Ischemia-Reperfusion Acute Kidney Injury by Preserving Na⁺/K⁺-ATPase](https://medichelpline.com/clinical-feed/biorxiv-8-modified-3rd-generation-synchronization-modulation-electric-field-3m-smef.md)

## Navigation
- [← Back to Gastroenterology Feed](https://medichelpline.com/clinical-feed/gastroenterology.md)
- [← All Clinical Specialties](https://medichelpline.com/clinical-feed.md)
## Medical & Regulatory Disclaimer

> [!CAUTION]
> MedicHelpline content is structured for research, educational, and professional discovery purposes. It does not constitute individual medical advice, clinical diagnosis, or treatment recommendations.
> Always verify dosing, contraindications, and regulatory alerts against official product labeling and primary regulatory sources before clinical decision-making.