Introduction

Can computers think? – This was the question on which Richard Bellman elaborated in his book in 1978 [1]. At that time, the era of modern artificial intelligence began. Early computational models and algorithms were developed to assist in medical diagnoses and treatment planning. This integration of artificial intelligence (AI) into medicine, including surgery, has revolutionised clinical medicine. Among various surgical disciplines, abdominal surgery became one of the first domains to harness the power of AI. However, the significant rise of AI specifically in liver transplantation (LT) began around the mid-2010s, with the most prominent advancements in the past 5 years. Currently, there are various AI applications in this field, but many more are expected. Several of the latest developments related to AI in LT should be highlighted. First, there is an increasing number of machine-learning models used to assess donor-recipient matching (random forest, gradient boosting) by integrating donor demographics, comorbidities, and histology. Many of these models today are outperforming MELD-based allocation [2–4]. Furthermore, in 2024, a transformer-based multi-task deep learning model with a fairness-enhancing algorithm was developed to accurately and equitably predict multiple post-liver transplant risk factors using a large-scale U.S. transplant dataset [5]. Similar models are being developed to predict postoperative pneumonia and cardiovascular incidents after LT [6, 7].

Second, many advancements are seen in deep‐learning–based image analysis (convolutional neural networks), which aims to quantify liver damage (steatosis or fibrosis) on donor computed tomography (CT)/magnetic resonance imaging (MRI) scans [8, 9].

Third, liver machine perfusion combined with AI opened a new horizon in coping with the deficit of donor organs. In this sense, AI may help to optimise perfusate composition, flow, and pressure [10]. Computer vision as an intraoperative tool is still in the experimental phase, but as in other surgical branches, it has huge potential in LT as well.

The utilisation of AI in liver diseases has already been explored in several publications; nevertheless, the majority of these studies did not focus specifically on LT, and some encompassed a wider range of gastroenterological disorders [11–13]. In the literature we found just one review study that demonstrates the use of AI specifically in LT, but it primarily focused on preoperative patient management and organ allocation [14]. Recently, a review about the role of AI in chronic liver diseases and LT was published by Spann et al. [15]. The paper offers important insights about the subject matter. The authors focus primarily on AI theory, chronic liver diseases, and hepatology in general, and they emphasise that AI is playing an increasingly significant role in all stages of LT, from preoperative planning to long-term patient follow-up. It also discusses barriers for widespread adoption of AI in clinical practice.

Our review focuses specifically on LT, and it demonstrates AI in a more simplified and understandable way. Thus, a current review is needed to catalogue all relevant AI tools and critically compare these to the present gold standards. It offers a comprehensive, focused review of the latest AI applications in LT while also analysing advantages and potential downfalls of this rapidly advancing technology. We discuss the application of AI across the entire LT timeline, from diagnosis and donor evaluation through intraoperative planning, postoperative care, and long-term follow-up, with a focus on everyday clinical implementation (Table I). Continued rapid advancement of AI in this field is expected, which is a result of innovation as well as improvements in existing computer technologies. Therefore, it is necessary for clinicians to continuously follow the development of AI, which necessitates the publication of review articles that address current AI achievements and their practical application.

Organ viability assessment

AI can assist in evaluating the quality and viability of donor livers using imaging and other data, optimising organ utilisation. Although histopathological examination is considered the gold standard for hepatic steatosis (HS) grading, the use of AI in evaluating HS during procurement is showing promising results. One prospective study compared the use of a machine learning (ML) model and standard histopathological examination in the assessment of HS. The ML model required an intraoperative smartphone liver picture and showed an accuracy of 89% in graft classification (≥ 30 vs. < 30% HS) [16]. LiverColor is a similar software platform that also uses AI image analysis to classify liver grafts regarding HS levels, with 90% accuracy in predicting > 30% HS [17].

Histopathological assessment during procurement requires a pathologist to always be available and is subjective. HEPASS (HEPatic Adaptive Steatosis Segmentation) is a fully automated AI algorithm used for the detection of lipid droplets in haematoxylin and eosin (H&E)-stained liver donor slides. The method demonstrated 97.27% accuracy in HS quantification [18]. A different study tested the six most used classification ML algorithms in the context of establishing the value of macrovesicular HS in liver donor biopsies stained with Sudan. All classifiers showed favourable results overall, with Naïve Bayes and KNN being the best in terms of speed and accuracy [19].

On the other hand, in living-donor liver transplantation (LDLT), accurate donor liver volumetry (LV) is essential for ensuring donor safety, as well as adequate graft-to-recipient weight ratio. AI and ML algorithms, especially convolutional neural networks (CNN), can be used to automatically segment the liver and assist in achieving a more accurate LV [20]. In LDLT, just as in deceased donor LT, HS remains one of the main factors in the selection of appropriate donors. A DONATION model uses AI technology to analyse parameters obtained by noninvasive tests to isolate healthy individuals with a low probability of significant HS, potentially exempting them from pre-donation liver biopsy [21].

Donor-recipient matching

AI algorithms can analyse large datasets to improve matching based on compatibility and urgency, potentially increasing success rates. With the persistence of disproportionate numbers of donors and waitlisted liver transplant patients leading to long waiting times and high mortality rates in certain patient groups, new solutions to donor-recipient (D-R) matching problems have emerged. AI tools can improve D-R matching, thus leading to better outcomes and cost-effectiveness. The most used classifiers are artificial neural networks (ANNs) and random forests, while the most frequent objectives have been the prediction of post-transplant graft survival and waitlist mortality [21, 22]. One multicentre study compared ANNs for D-R matching to validated scoring systems (MELD, D-MELD, DRI, P-SOFT, SOFT, and BAR). ANNs proved to be significantly more accurate in predicting the 3-month graft survival and loss than classical validated scoring systems [22, 23]. There are multiple ethical considerations regarding solid organ transplantation, and AI use in D-R matching is no exception. A study assessing public attitudes regarding AI use in liver allocation demonstrated that most participants (69.2%) found it acceptable, believing it to be more consistent and less biased than humans [23, 24]. Recently, Gambella et al. developed and validated the first AI algorithm for assessing liver graft steatosis based on Banff recommendations on 292 consecutive allograft liver biopsies. Results indicated that such an algorithm provides reliable automated assessment of liver steatosis, which can predict short- and long-term organ viability [25]. This study resulted in innovative evidence and considerations in the field of AI and LT. However, as discussed by Li et al. in their letter to the editor, the study has several limitations, and we mention the main ones below [26]. First, the sample size is relatively low, and the majority (75%) contain fat content below 30%, while some important data on comorbidities and laboratory findings are not provided. Second, it is highlighted that in this article the authors did not use standard transplantation clinical outcomes, which would have contributed to a better assessment of the method’s clinical significance and applicability. Third, there is no comprehensive explanation of the algorithm’s underlying principles, which is important when analysing the comparative advantages of one method over another. We agree that all these shortcomings may restrict the generalisability and applicability of the described method. In addition, the limitations related to retrospective and monocentric design should be outlined. Despite all this, the article has great value, and if such methods are applied in clinical practice, it would contribute to more efficient and safe allocation of donor organs. Similar prospective studies on a larger number of samples could in the future result in wider clinical application of predictive models for liver steatosis assessment.

Predictive analytics

Machine learning models can predict post-transplant outcomes, helping clinicians make informed decisions about patient management. Predictive analytics emerges as a transformative component of AI applications in LT. AI enables the integration and analysis of complex data, such as donor characteristics, recipient condition, and perioperative factors, to provide more accurate individual outcome prediction. Unlike conventional scoring systems such as Model for End Stage Liver Disease (MELD), Donor Risk Index (DRI), and Survival Outcome Following Liver Transplantation (SOFT), which rely on static and often linear models, AI-driven approaches can dynamically learn patterns from past data to improve how we assess risk and support clinical decision-making. This makes it possible to make better decisions, such as whether to accept an organ, how to prioritise patients on the waiting list, and how to plan post-transplant care. AI methods, including neural networks and random forest models, can be updated regularly with new information, making them flexible and applicable in different clinical settings. Recent studies have demonstrated that machine learning (ML) approaches, particularly artificial neural networks (ANN), outperform traditional models. A systematic review study found that ANN-based models have achieved area under curve (AUC) scores of 0.82 to 0.84, compared to SOFT (0.57–0.64) and DRI (0.68), indicating significantly improved predictive accuracy [24]. AI-driven predictive analytics therefore represent a significant advancement in LT, enabling more precise, data-driven decisions that can enhance patient outcomes and optimise resource allocation. This is also a limitation because some models cannot be globally applied and should be used in the same population as the data sets from which they were derived. Other disadvantages include the need for periodical updating and loss of explainability (the so-called black box issue).

Computer vision

Computer vision in LT is an active area that focuses on extracting actionable information from images and videos to support the decision-making process. While many ideas of computer vision applications are still going through the experimental phase, many promising results have already been achieved. From a surgical aspect, the most important are intraoperative visualisation and surgical navigation, especially when a laparoscopic/robotic technique is used. Computer vision may assist in producing virtual reality (a useful tool for real-time detection of anatomy and anatomic variants), or it may help with postoperative imaging review and detection of complications.

Immunosuppressive therapy

Immunosuppressive therapy is the mainstay of posttransplant management, and it allows long-term survival. It is crucial for preventing organ rejection while minimising the risk of infections and other complications. Large volumes of patient data, including genetic profiles, can be analysed by AI. AI can analyse a patient’s medical background and their reactions to prior treatments to create individualised immunosuppressive therapies. By customising drugs and doses for each patient traits, AI can minimize side effects while maximising effectiveness. In addition, AI can help with ongoing patient monitor-ing via wearable technology.

Post-operative care

Real-time biomarker and vital sign analysis allows for early detection of signs of rejection or infection, enabling timely interventions. In addition, AI-driven systems can improve medication adherence through reminders and educational resources tailored to indi-vidual patients [27]. Possible complications following LT include graft failure, infections, cardiovascular complications, and more. Many studies have used AI to identify these complications and influence survival. For example, a 2021 study by Chen et al. developed six predictive models using machine learning methods to predict pneumonia in patients who underwent LT. The study concluded that the best-performing model, XGBoost, might accurately predict postoperative pneumonia based on specific features such as laboratory findings, operation time, and total anaesthesia time [6, 28]. Another rare postoperative complication associated with high mortality is acute graft-versus-host disease (GVHD). Using machine-learning, a study from 2022 developed a practical model that could identify patients at high risk for developing GVHD, allowing for additional monitoring with blood chimerism testing [29].

Resource management

AI can streamline logistics and resource allocation, ensure efficient use of medical resources, and reduce wait times. In recent years, studies have shown many different ways in which AI can improve the LT process, from preoperative planning to long-term postoperative care.

The MELD score, as an assessment of waitlist mortality risk, is currently the most widely used score for determining urgency in receiving LT. A study from 2021 showed that machine learning and deep learning models outperformed the MELD score for all prediction cases, using variables such as AST, ALT, and haemoglobin as well as standard MELD variables [30].

AI can also be a useful tool in the assessment of graft HS, an important factor for predicting liver dysfunction risk after transplantation. One study has shown that machine learning and automatic texture analysis of RGB images from smartphone cameras achieved promising classification sensitivity, specificity, and accuracy and could in the future be used to assist surgeons in HS assessment inside the operating room [31].

Finally, AI can be of use in predicting postoperative adverse outcomes. In a study [30] from 2024, looking at data from 160,360 patients, a deep learning model was developed to predict five post-transplant risks: cardiovascular complications, infection, malignancy, diabetes, and graft rejection. The study concluded that the model significantly reduced the task discrepancy by 39% and that multi-task learning outperforms single-task predictions.

Future prospects and challenges

It is not simple to evaluate what the current practical role of AI in LT is, because there is a lack of long-term prospective studies, which is understandable due to the novelty of this field. There is a global agreement that AI should augment, not replace, clinicians, so it is hard to expect that a certain AI tool will become the gold standard for the treatment and diagnosis of specific diseases, especially if we take into account legal and ethical aspects. The most prominent advantage of AI compared to standard techniques lies in donor-recipient matching and graft survival predictions, where AI tools like ANNs and random forests have demonstrated more accurate predictions than current validated scoring systems like MELD or SOFT [32, 33].

From a statistical point of view, AI outperformed clinicians in many clinical tasks, from interpretation of radiological images to histological evaluation of liver parenchyma. When considering AI applications in LT presented previously, several benefits should be highlighted. First, AI-based predictive analytics is by default better than standard statistics (e.g. logistic regression), and it is better than the surgeon’s “common sense”. Therefore, today, AI-based analytics can predict graft survival, recipient outcomes, and long-term complications with much higher accuracy, especially in complex, high-dimensional risk estimations, since AI can capture nonlinear interactions among various donor/liver/recipient characteristics. Second, immunosuppressive regimens can be personalised by predicting the risk of rejection, infection, or drug toxicity based on pharmacokinetics, genomics, and clinical data. Despite this, due to legal issues, it is hard to believe that medication prescription will not remain the responsibility of the doctor.

Third, as described previously, AI can contribute to better donor-recipient matching by integrating multi-modal data and predicting compatibility beyond ABO/Rh and HLA matching. Again, regulatory and ethical aspects may limit this benefit of AI in this step of LT, because it will not be easy to ensure alignment of AI use with organ allocation policies. Lastly, there is evidence that even large language models such as ChatGPT4 can analyse histologic images and assess steatohepatitis grade with success comparable to pathologists [34]. All this suggests that AI is rapidly evolving, and today it is an extraordinary tool for improving LT outcomes, but its practical implementation is somewhat delayed due to the aforementioned issues. For more reliable comparative studies between humans and AI, we need to define concrete use cases and success metrics, solve ethical issues associated with it, and establish external collaborations to validate models across diverse populations.

The limit of AI in med-icine including LT is unpredictable. Future developments in machine learning and big data analytics are anticipated to further integrate AI into LT. Real-time decision support technologies that give surgeons vital information during operations could be used in future applications to significantly improve accuracy and results. AI has the potential to completely transform organ transportation logistics by guaranteeing that organs are matched and transported as efficiently as possible. Developments in AI-driven personalised patient care will surely improve recovery and lower the risk of organ rejection. AI will also further simplify administrative work, freeing up healthcare providers to concentrate more on patient care. However, the most exciting part of this technology is related to the development of autonomous robotics, and time will show us if and when this will become reality.

On the other hand, with great power comes great responsibility, and clinicians and scientists should be aware of challenges and limitations related to AI and use it cautiously. The most significant challenge is related to data availability and data quality because AI model training requires a huge amount of complete and reliable datasets. Biased or incomplete data would affect AI model accuracy. Second, AI always raises some ethical questions, especially those related to patient privacy, as well as the fact that important decisions are made by an AI algorithm rather than a doctor. Also, both clinicians and patients should understand AI decision processes, which is crucial for trust. Third, implementing AI technologies requires investment in infrastructure, including data servers and training of clinicians.

Data quality and availability remain significant hurdles because AI systems require large, high-quality datasets to function effectively. Complexity of outcomes and accurately predicting these (graft survival, rejection risk, or complications) involves multifaceted data, which makes AI modelling very challenging. Ethical considerations, particularly concerning patient privacy and the transparency of AI decision-making processes, must be addressed. Additionally, integrating AI technologies into existing medical workflows can be complex and resource-intensive, requiring substantial investment in infrastructure and training. Most costs are related to developing, validating, and maintaining AI systems, and the whole process requires multidisciplinary teams combining data scientists, engineers, and clinicians.

Conclusions

Artificial intelligence (AI) has enormous potential to greatly enhance LT results and efficiency. However, achieving this potential will necessitate resolving present integration, data, and ethical issues. The future of LT will be greatly influenced by the ongoing cooperation between medical experts and AI developers as technology develops.

Funding

No external funding.

Ethical approval

Not applicable.

Conflict of interest

The authors declare no conflict of interest.

References