A systematic review on Artificial Intelligence applied to predictive cardiovascular risk analysis in liver transplantation

1. Introduction

The domain of healthcare has seen a significant rise in the implementation of artificial intelligence, and liver transplantation is no exception, especially from the perspective of data analytics. Liver transplantation is the last therapeutic option for patients with end-stage liver disease,¹ and is the second most common organ transplantation. Several countries across the world have seen a sharp rise in the number of liver transplants, owing to the population growth, public educational campaigns to enhance awareness of transplantations, and an expansion of organ procurement organizations, among other reasons.² For example, just in the USA, 9236 liver transplants were performed in 2021.³ The requirements and outcomes of liver transplantations involve complex multidimensional relationships of different variables concerning the donor, recipient, surgical procedure, and post-transplantation medical management,⁴ in addition to the clinical indicators. With respect to analysis of candidates for liver transplantation, it is often observed that factors like age, cardiovascular comorbidities such as, coronary heart disease, and other medical conditions such as cirrhosis, acute alcohol hepatitis and nonalcoholic fatty liver disease, significantly impact the risk evaluation of the patients destinated for the transplant. To determine the priority and urgency in the waiting list for liver transplantation, most countries adopted the Model for End-Stage Liver Disease (MELD/MELD-Na) scores for evaluation and subsequent risk profiling of the candidates. The MELD score and its modified versions have provided reasonably robust predictions of short-term mortality among patients with end-stage liver disease, making it a widely used tool for clinical prognosis.⁵ Nevertheless, liver transplantation is a continuously evolving field, and while further studies suggest the presence of several complex variables and new biomarkers and risk predictors, advanced techniques of artificial intelligence are now being developed for improvements in predictive risk analysis in liver transplantation.

Artificial Intelligence is being used in the organ transplant settings in a formalized manner increasingly. Most of the machine learning algorithms are data-adaptive and help producing predictive models that self-learn from clinical data with reasonable accuracy. Liver transplantation involves complex interactions and often nonlinear relationships between demographic, clinical, laboratory, imaging, lifestyle data and several other parameters. Machine learning algorithms are particularly attractive in this context to take advantage of this wide gamut of features and to develop predictive models on that. On one hand, supervised models work on labeled datasets and can help with predictive outcomes. Especially in risk or disease predictions, supervised learning encompasses classification, where the predicted outcome could be a labeled class (disease prediction) and regression, where the predicted outcome could be a continuous range or number (for example, treatment outcome prediction). On the other hand, unsupervised learning uses unlabeled data, examines hidden patterns and relationships between data, for example, performing data clustering.

Machine learning can be a supportive tool in liver transplantation to resolve pre-transplantation predicaments like graft allocation, hepatic steatosis assessment, and the effectiveness of liver segmentation. The Donor Risk Index (DRI) and Model for End-Stage Liver Disease (MELD) are clinical tools widely used for graft allocation. However, since conventional models often fail to consider the complex nonlinear relationships between donor, recipient, and surgical procedure, machine learning tools are used to overcome these limitations. Furthermore, MELD score often fails to reflect the liver disease severity of some patients, which can create a discrepancy in the 90-day waitlist mortality risk. However, well-trained decision-tree-based models can help to create a different waitlist by incorporating all necessary variables, resulting in higher accuracy. Machine Learning also assists in solving post-liver transplantation challenges such as graft rejection, graft failure, patient survival, and post-operative morbidity. Patient survival prediction is one of the most explored research areas for the intersection of liver transplantation and artificial intelligence since predictive models based on machine learning can handle large volumes of uncorrelated data and can assess short-term and long-term post-liver transplantation mortality.

With respect to risk assessment in liver transplantation, cardiovascular diseases hold an important impact. Specific cardiovascular responses and symptoms in cirrhosis can be deleterious for patients’ health and suggest an increased risk in liver transplantation. Patients with cirrhosis requiring liver transplant often exhibit conditions like cirrhotic cardiomyopathy, which is present in up to 30% of patients with cirrhosis⁶ and is characterized by increased cardiac output and inept ventricular response to stress. Patients with underlying cardiovascular conditions are prone to complications in the post-liver transplantation tenure, highlighting the need for the analysis of underlying cardiovascular status during the assessment of the waiting list for transplantation to determine whether the patients are suitable for the procedure. While there exists empirical data and research looking into the various aspects of cardiovascular implications on liver transplants, there is a paucity of a systematic framework that can be incorporated into clinical practice for improved risk analysis and perioperative management of transplant candidates. With the introduction of artificial intelligence, especially with the techniques of machine learning, the domain of identification of risk predictors and biomarkers has been amplified largely.

Considering the importance of assessing cardiovascular risks in liver transplantation setup, in this work, we present a systematic review on the aspect of using artificial intelligence applied to cardiovascular risk assessment in liver transplantation, aiming at identifying the principal lines of development, key issues, and challenges, as well as the future direction.

2. Methods

This systematic review has been performed focusing on the implementation of artificial intelligence (AI) for the assessment and predictive analysis of cardiovascular risks in liver transplantation. Digital repositories- IEEE Xplore, PubMed, ScienceDirect, Springer Link, and Wiley Online have been explored for a qualitative analysis of publications. For the review, we used a systematic structure,⁷ consisting of definition of the key research questions, precise definition of the research scope, search and screening of relevant publications, analysis in terms of the research questions, and finally, a general perception of the current and future lines of the field. To perform a specific selection of the studies to be included in this review, we have established the definition of inclusion criteria. At first, only journal articles, conference papers, and book chapters written in English were considered from the digital repositories. Secondly, the year of publication was set from 2000 to 2023. Regarding the scope of the studies, the incorporation of artificial intelligence was set, followed by the consideration of the specific domain of predictive analysis of cardiovascular risks in liver transplantation.

The following sets of search keywords have been used in different combinations: artificial intelligence, machine learning, liver transplant, liver transplantation, and cardiovascular. Following the search, a careful analysis of the publications has been made with respect to the inclusion-exclusion criteria (Figure 1). During this research, 119 papers were identified which were deemed pertinent to the study. Subsequently, 10 duplicate papers and an additional 33 papers were excluded due to irrelevance to the systematic study, subject to the inclusion-exclusion criteria. Furthermore, 64 papers were eliminated from consideration due to inadequate coverage of the designated research topics.

Figure 1. Search pipeline.

With respect to the analysis of the relevant literature, research questions have been established (Table 1).

3. Results

After an exhaustive search, from a total of 119 papers subsequently passing through preliminary filtering and application of the inclusion-exclusion criteria and subject relevance i.e. the use of artificial intelligence in liver transplantation with respect to cardiovascular risks, 12 papers were obtained and considered as the core of this systematic review. Each publication has been critically analyzed in terms of the research questions modeled before, and finally, a general perspective of the outcome of the review is presented.

Figure 2 depicts the outcomes of the search methodology. A notable observation is the insufficient coverage of all three research topics in several papers, which focused solely on singular or dual aspects. This underscores the underexplored intersection of artificial intelligence with liver transplantation and cardiovascular risks, suggesting a potential domain with limited prior investigation or scholarly studies.

Figure 2. Results obtained following search pipeline.

As a background of this study, several reviews related to the application of artificial intelligence in liver transplantation were analyzed (Table 2). These reviews provide a general approach to the state of the art of artificial intelligence in this field.

Reviews,^8,9,13 and¹⁷ focus on the various applications of artificial intelligence and machine learning on liver transplantation and hepatology, highlighting the strengths, downfalls, and future perspectives. In,⁸ 21 papers were broadly reviewed between a period of 1997 to 2019; 4 papers focusing on the pre-liver transplantation period i.e., prediction of 3-months Waitlist mortality, outcomes after hospitalization; 7 papers on prediction of short-term mortality post-liver transplantation; 3 papers on prediction of long-term survival post liver transplantation and 7 papers focusing on post-liver transplantation complications. The potential fields of interest for future development of machine learning suggested by⁸ were resolution of bacterial infections, allocation of extended criteria donors and donors after circulatory death, and immunosuppression management. Paper⁹ focused on the use of machine learning in liver transplantation for both pre-transplant and post-transplant applications, by analyzing 29 articles out of a total 216 articles, identified from various databases, between a period of 2015-2021. Various machine learning algorithms were assessed for different applications, however XGBoost deemed to be most suitable for liver transplantation-clinical decisions given its rapid processing and high accuracy.¹³ analyzed 40 papers from MEDLINE (PubMed) and Embase (Elsevier), suggesting the efficient replacement of machine learning in place of biostatistics.¹⁷ discusses aspects of AI in liver transplantation, with a focus on applications which require accurate decisions such as organ allocation and survival prediction. The paper highlights classifiers such as ANNs, decision tree classifiers, random forest, and Naïve Bayes models.¹⁰ is a systematic review assessing 9 articles and 18,771 Liver transplants with a focus to predict survival of grafts using Artificial Intelligence. A maximum AUROC of 0.84 was achieved using the ANN models, compared to linear regression and standard predictive modeling.^11,12 are systematic reviews about the applications of AI for predicting NAFLD and NASH. NAFLD is one of the leading causes of liver failure and it has been found that cardiovascular and renal risks are highly prevalent in NAFLD patients as well.¹¹ analyzes summarized diagnostic/assessment models which perform diagnostic tasks on medicinal data and optimize their function for better accurate diagnostic effects.¹² has summarized various domains concerning NAFLD by analyzing 4 papers from Electronic Medical Records (EMR), 1 paper from Imaging, 1 from Microbiome and 5 papers from Histology.¹⁴ is a systematic review of studies that predict patient survival after liver transplantation using AI techniques and compares the proposed models with existing models such as SOFT score, MELD score, DRI score and BAR score. Among all the models of all the studies analyzed, a highest AUROC of 0.9 was achieved by ANN model on different datasets.¹⁵ is a review that highlights the application of ML methods in assessing and predicting various diseases. The studies primarily investigate the prognosis of NAFLD and acute liver failure, using decision tree models. The studies demonstrate the growing application of data mining techniques to analyze medical data and provide insights for clinical diagnosis and risk assessment in common diseases.¹⁶ discusses the increasing use of AI in organ transplantation, with an emphasis on donor-recipient matching. Deep Learning models (ANN, Random Forest) handle a multitude of variables and hold potential to improve organ allocation by enhancing predictability. Both ANN and RF outperform the metrics such as MELD and post-LT survival outcome scores.

The selected reviews collectively illustrate the significant strides made in applying artificial intelligence and machine learning to liver transplantation. Particularly supervised learning models like decision trees, random forests, and artificial neural networks have shown superior accuracy in predicting short-term and long-term survival, complications, and graft rejection compared to traditional clinical models. The reviews also highlight the emerging role of artificial intelligence in addressing complex challenges such as optimizing organ allocation and managing immunosuppression. Furthermore, the ability of machine learning to process and learn from vast and multidimensional datasets makes it a valuable tool in this domain, offering potential improvements in both the precision and efficiency of clinical decision-making in liver transplantation.

Narrowing down the analysis to the specific area of cardiovascular risks in liver transplantation (Figure 2), Table 3 highlights the studies that used artificial intelligence for the same.

¹has analyzed the cardiometabolic risks during pre-liver transplantation and post-liver transplantation of patients registered in the National Liver Transplantation Program (NLTP) program in Uruguay. A size of 165 patients of the 2014-2019 cohort was considered, considering vascular age as the significant factor. A maximum AUC score of 0.74 and AIC score of 0.8125 was obtained using Logistic Regression. Similarly,¹⁸ aimed to perform a risk-group analysis by dividing the NLTP patients into clusters based on their health parameters, with a specific focus on Cardiovascular risk.^29,28 are studies which focus on the analysis of ECGs to predict and evaluate cardiovascular complications after liver transplantation, with the help of AI algorithms.¹⁶ included 2 cohorts of adult patients with end-stage liver disease either undergoing LT evaluation or receiving transplant between 2017 and 2019 for their retrospective study. The study analyzed the performance of an AI based algorithm (AI-ECG) which reflected an AUROC of 0.69 for prediction of subsequent low left ventricular ejection fraction. Similarly,²⁸ developed a deep learning model to provide an AI-Cirrhosis-ECG (ACE) score to correlate cirrhosis with cardiovascular dysfunction. The model resulted in an AUC of 0.908 with consistency. Similarly,²⁶ utilized CT scans to assess cardiovascular risk. The study used automated ML algorithms focused to evaluate biomarkers on non-contrast chest CT (Emphy-Alg, LD-Alg, CCS-Alg) which was followed by multivariate logistic regression and statistical analysis. The algorithms demonstrated their potential to predict CVD incidence and mortality. The studies suggest that AI can be a potential tool for identifying cardiovascular complications in transplant evaluations.²⁰ aimed to predict the post liver transplantation cardiovascular events and mortality, wherein 1459 patients undergoing liver transplantation between January 2008 and December 2019 were studied. Various machine learning algorithms were analyzed, however the GBM model XGBoost achieved the best results, with an AUC=0.71 for overall survival and AUC=0.72 for cardiovascular mortality. Similarly,²³ aims to predict risks and early signs of chronic diseases by analyzing the health data. A cohort of 104 patients from the NLTP, Uruguay program between 2014-2017 was considered and k-means was applied on the data for clustering. The study related the risk groups before liver transplantation with cardiometabolic risks using vascular age.

Non-alcoholic Fatty Liver Disease (NAFLD) can lead to enhanced cardiovascular risks which may lead to the need for a liver transplant.^28,22 aim to study the correlation of Cardiovascular risks with NAFLD²⁸; studies 2239 hypertensive patients to analyze the most relevant features of NAFLD using adverse machine learning algorithms, resulting in a maximum AUC=0.79.²² aimed to develop a machine learning model that integrates various risk factors to identify cardiovascular disease in liver transplant patients. Seven various Supervised Learning methods were analyzed, with Random Forest giving the highest AUC of 0.849, in the UK Biobank cohort. Similarly.³⁰ demonstrated the use of Random Forests in predicting the patients with a high risk for developing NASH. Data of patients who attended the Cardiovascular Risk Unit of Mostoles University Hospital (Madrid, Spain) from 2005 to 2021 was used. E-health records were analyzed using 6 models entailing different pre-processing strategies. Random Forest models provided impressive results (best model accuracy- 0.87, worst model accuracy- 0.79), highlighting the power of ML to provide insightful results in disease prediction.²⁴ aimed to assess the use of CT-derived fractional flow reserve (CT-FFR) in differentiating coronary stenosis in patients being evaluated for LT. The patients were preliminary tested to rule out coronary artery disease (CAD), and CT-FFR was computed using a ML algorithm. The diagnosis reflected an accuracy of 0.85, suggesting a promising noninvasive tool to exclude hemodynamically significant coronary stenosis, thereby reducing the need for invasive measures to achieve the same.²⁵ focused on the correlation between metabolic-associated fatty liver disease (MAFLD) and cardiovascular disease (CVD) risk. The study divided the 191 patients into two groups: 144 without CVD and 47 with CVD. The best model operating on 5 most discriminative features correctly identified 114/144 low-risk and 40/47 high-risk patients, highlighting the potential of ML in predicting identifying MAFLD in patients with a CVD risk.²⁷ aimed to investigate the correlation between hepatosteatosis (HS) and cardiovascular disease (CVD) risk in HIV-positive individuals. CT images of 1306 subjects (HIV-positive and HIV-negative individuals) were analyzed to determine the presence of HS and CVD using Logistic regression and Random Forest models. The results suggested that metabolic dysfunction may contribute to the increased CVD risk in this population.

These studies (Table 3) provide an in-depth idea of how artificial intelligence is being applied to assess cardiovascular risks in liver transplantation, highlighting the diversity in the algorithms used, cohort sizes, and specific cardiovascular outcomes examined. The use of logistic regression, deep learning, and various supervised and unsupervised learning methods across different studies demonstrates the versatility of applying artificial intelligence in this domain. Several studies focused on pre-liver transplantation cardiovascular risk assessment, employing models like logistic regression and k-means clustering to predict cardiometabolic risks based on health parameters and vascular age. Others utilized AI-enabled electrocardiograms (AI-ECG) and deep learning models to predict cardiovascular complications and dysfunction in liver transplant candidates, achieving promising AUC scores that underscore the predictive power of AI. Additionally, machine learning algorithms, including Random Forest, Support Vector Machines (SVMs), and Gradient Boosting Models (GBM), were applied to predict major adverse cardiovascular events post-transplantation, often potentially outperforming traditional clinical models in terms of accuracy and predictive capability. In the context of non-alcoholic fatty liver disease (NAFLD) and metabolic-associated fatty liver disease (MAFLD), which are linked to heightened cardiovascular risks, AI models have been used to identify relevant features and predict cardiovascular diseases with high accuracy. These studies also employed various imaging techniques, such as computed tomography (CT) scans, coupled with AI algorithms to noninvasively assess coronary artery disease risk and evaluate cardiovascular biomarkers, further illustrating AI's potential in improving pre-transplant evaluation and post-transplant outcomes.

With respect to RQ1, the studies reveal a variety of AI algorithms employed to predict cardiovascular risks in liver transplantation, reflecting diverse methodological approaches. Logistic regression is commonly used for its simplicity and interpretability, particularly effective in smaller cohorts and for binary classification tasks. Unsupervised learning techniques like k-means are applied to cluster patients based on health parameters, useful for identifying underlying patterns without labeled data. Convolutional Neural Networks (CNNs) are utilized in AI-enabled electrocardiograms (AI-ECG) to extract features from ECG data for risk prediction. Other methods such as Random Forests, SVMs, LASSO, and Gradient Boosting Models (GBM) are preferred for their ability to handle complex, high-dimensional data and their robustness in various predictive tasks. Deep learning models, employed for ECG analysis and CT-derived fractional flow reserve (CT-FFR), are effective in capturing intricate relationships within large datasets. The choice of algorithms often correlates with the nature of the dataset, the type of prediction needed, and the existing knowledge of intra-data relationships.

In terms of RQ2, the results across different studies highlight AI's potential in risk prediction. AI models have often shown improved accuracy and reliability over traditional models, as they can better manage complex datasets and uncover hidden patterns. The use of AI-enabled ECGs and CT-FFR models showed significant promise in noninvasive risk assessment, highlighting the effectiveness of AI methods in providing accurate and reliable predictions, which are crucial for improving clinical outcomes in liver transplantation.

Regarding RQ3, validation of AI models stands as a critical point due to several factors. The absence of one-size-fits-all structure in AI models often limit the functionality of a predictive model to a specific cohort or dataset and is not directly transferable to other cohorts, unlike several traditional clinical models. Limited information has been obtained on the validation and clinical reproducibility of the AI-based predictive models. In addition to validating through clinical trials or retrospective analysis of large patient cohorts, it is crucial to ensure that the models are robust and applicable in real-world settings.

For RQ4, several design and implementation challenges were identified from a general perspective. Ensuring the availability of high-quality, comprehensive datasets is crucial for training robust AI models. Balancing model complexity with interpretability remains a challenge, especially in clinical settings where transparency is essential. Integrating AI models seamlessly into existing clinical workflows requires careful consideration of usability and acceptance by healthcare professionals. Moreover, addressing ethical and legal considerations related to patient privacy, data security, and the ethical use of AI in healthcare is paramount.

With respect to RQ5, future research directions in the field of artificial intelligence for cardiovascular risk assessment in liver transplantation are expected to prioritize the enhancement of model interpretability and transparency. This focus aims to bolster clinical trust and facilitate the broader integration of AI technologies into healthcare settings. The development of hybrid AI models that amalgamate the predictive capabilities of various algorithms could potentially elevate both the accuracy and robustness of risk assessments. Moreover, future studies are anticipated to extend the validation of these AI models across more diverse and larger cohorts, which is essential for confirming their applicability and generalizability across different patient demographics. The integration of AI with innovative technologies such as wearable health devices and telehealth services of day-to-day clinical settings promises to revolutionize patient monitoring by enabling continuous and proactive management of individuals at elevated risk. Additionally, future research will need to rigorously address the ethical, legal, and social challenges posed by the deployment of AI in medical contexts, ensuring that these technologies are used responsibly and equitably.

Nevertheless, the studies reviewed in this work suggest that AI provides valuable tools for enhancing cardiovascular risk assessment in liver transplantation, offering more precise and efficient predictive models compared to traditional methods. This can significantly impact clinical decision-making, leading to better management and outcomes for liver transplant patients.

4. Discussion

The disruptive development of artificial intelligence has seen the domain of healthcare being reengineered to incorporate predictive modeling in different areas. In both domains of liver transplantation and cardiovascular diseases, machine learning is being used profusely, but quite independently. From an implementational perspective, the application of artificial intelligence, especially machine learning algorithms to predict cardiovascular risk in patients suffering from liver diseases, undergoing liver transplantation, or recovering post-surgery. Liver transplantation patients in several the intersection between the fields of liver transplantation and cardiovascular diseases, given the high impact of cardiovascular risks in affecting the patients undergoing liver transplantation. For example, cardiovascular disease is the leading cause of mortality among patients with nonalcoholic fatty liver disease, which in turn, is the most prevalent liver disease worldwide.²² Through this work, the research questions presented at the beginning are answered. Primarily, the review identified the key methods used in this domain, and their subsequent results and validation parameters. Discussions on the key challenges and future scope follow.

Machine learning stands as a highly potential tool in the intersection of cardiovascular diseases and liver transplantation, not only from the predictive point of view, but also from the perspective of discovering new patterns and relationships between the hepatic and cardiometabolic features, useful for eventual prediction of risks. The lack of studies in this joint area of artificial intelligence applied to cardiovascular diseases in liver transplantation is the evidence highlighting the need of transdisciplinary interactions. It involves challenges at different levels. At the first place, the integration of the data sets is a crucial challenge, considering the difficulties in developing and handling a unified database shared between the cardiovascular and hepatic profiles despite the focus being on liver transplantations. Secondly, in the domain of mortality in liver transplantation, the interference of multifarious factors makes it difficult to isolate cardiovascular impact. Cardiovascular complications in orthotopic liver transplantation are quite seldom; however, since they compromise the blood flow of the transplant, they are the most feared complications with a high incidence of both graft loss and mortality. In this respect, diagnosis and therapeutic management of vascular complications constitute a major challenge in terms of enhancing the success rate of liver transplantations.³¹ Especially from the implementational perspective of artificial intelligence, data quality as well as quantity are significant challenges. The diversity among hepatic and cardiometabolic data from different countries makes it difficult to develop unified models to avoid one-size-fits-all approach for risk predictions. Also, the lack of prospective studies poses a significant challenge on the confidence of the machine learning algorithms, besides its sensitivity and reliability when applied to different cohorts with differed attributes. However, despite the current challenges, the need of integrating cardiometabolic and hepatic aspects while applying artificial intelligence to predict cardiovascular risks in liver transplantation is the call of the hour. By incorporating AI methodologies into routine risk assessment protocols and throughout the Electronic Health Records (EHR)-creation procedure, healthcare practitioners can harness the power of advanced algorithms to refine patient stratification, optimize resource allocation, and facilitate timely intervention strategies. As we navigate toward an era of personalized medicine, the convergence of artificial intelligence and cardiovascular risk prediction in liver transplantation emerges as a pivotal avenue for transformative advancements, fostering a paradigm shift towards more effective and patient-centric healthcare practices.