Artificial intelligence and liver transplantation: Looking for the best donor-recipient pairing

Jvier Brice?o Rfel Cllej Césr Hervás c

a Unit of Liver Transplantation, Department of General Surgery, Hospital Universitario Reina Sofía, Córdoba, Spain

b Maimónides Institute of Biomedical Research of Córdoba (IMIBIC), Córdoba, Spain

c Department of Computer Sciences and Numerical Analysis, Universidad de Córdoba, Córdoba, Spain

Keywords:Donor-recipient matching Artificial intelligence Deep learning Artificial neural networks Random forest Liver transplantation outcome

ABSTRACT Decision-making based on artificial intelligence (AI) methodology is increasingly present in all areas of modern medicine.In recent years, models based on deep-learning have begun to be used in organ transplantation.Taking into account the huge number of factors and variables involved in donor-recipient (DR) matching, AI models may be well suited to improve organ allocation.AI-based models should provide two solutions: complement decision-making with current metrics based on logistic regression and improve their predictability.Hundreds of classifiers could be used to address this problem.However, not all of them are really useful for D-R pairing.Basically, in the decision to assign a given donor to a candidate in waiting list, a multitude of variables are handled, including donor, recipient, logistic and perioperative variables.Of these last two, some of them can be inferred indirectly from the team’s previous experience.Two groups of AI models have been used in the D-R matching: artificial neural networks (ANN)and random forest (RF).The former mimics the functional architecture of neurons, with input layers and output layers.The algorithms can be uni- or multi-objective.In general, ANNs can be used with large databases, where their generalizability is improved.However, they are models that are very sensitive to the quality of the databases and, in essence, they are black-box models in which all variables are important.Unfortunately, these models do not allow to know safely the weight of each variable.On the other hand, RF builds decision trees and works well with small cohorts.In addition, they can select top variables as with logistic regression.However, they are not useful with large databases, due to the extreme number of decision trees that they would generate, making them impractical.Both ANN and RF allowa successful donor allocation in over 80% of D-R pairing, a number much higher than that obtained with the best statistical metrics such as model for end-stage liver disease, balance of risk score, and survival outcomes following liver transplantation scores.Many barriers need to be overcome before these deeplearning-based models can be included for D-R matching.The main one of them is the resistance of the clinicians to leave their own decision to autonomous computational models.

Introduction

The results of liver transplantation have improved considerably due to the advances in surgical technique, the management of immunosuppressive drugs, the better knowledge of post-transplant long-term morbidities, and the greater knowledge of the mechanisms involved in the preservation of the organ.Recipient and graft survivals are around 95% in 1 year, and long-term survivors are common today.Unfortunately, these results are overshadowed by the existence of long waiting lists and the mortality of candidates whilst waiting for a donor.This is a common fact throughout the world, resulting from a disproportion between the number of donors and the growing demand for transplantation.

Current approaches such as expanding the inclusion criteria to elderly and morbidly obese candidates and the implementation of transplant oncology further aggravate the problem.This situation coexists with a different use of donors and organs depending on the regional donation rates.Countries with deceased donor rates of 20-30 per million population (pmp) present a utilization of close to 100%; however, those with the higher donation rates (>50 pmp) underuse donors and grafts (15% of refusals) [1].An aggres-sive strategy through the implementation of preservation machines could achieve a close to 100% successful utilization of allografts [2].

Regardless of the problem of imbalance between number of available donors and candidates, the allocation policy depends on which is the main objective wanted to be prioritized.A policy based on benefiting the highest risk candidate is based on the principle of urgency (“sickest-first”principle).Conversely, apolicy that tries to maximize the benefit of transplantation may favor candidates in better clinical conditions in order to achieve better results (“individual transplant benefit”and “population based transplant benefit”) [3].Along with the policies of prioritizing candidates, which are not exempt from ethical conflicts and controversies, the problem of assigning a specific donor allograft to a patient on the waiting list based on different assignment policies exists.In other words, the characteristics of the donor can impact the candidate’s outcome, especially when his/her clinical situation is critical.From this perspective, a donor with a low donor risk index (DRI)will be suitable for any recipient, regardless of the clinical severity.However, the combination of an expanded criteria donor (ECD) and a high-risk candidate may make the transplant futile.Therefore, an experience-based rule argues not to combine high-risk donors and candidates (risk divergence allocation policy) [ 4 , 5 ].Every recipient has a pre-transplant vulnerability that combines favorable responses to transplantation aspects (e.g., liver dysfunction) and, unfavorable ones that are independent of transplantation (e.g., age).A disproportion of transplant nonresponsive factors combined with an ECD will converge into a futile post-transplant result [5].It can therefore be deducted that the donor-recipient (D-R) matching is one of the biggest dilemmas inherent in organ transplantation.In this review, we focus on D-R matching based on artificial intelligence (AI) and, more specifically, on deep learning.

D-R matching based on biostatistical models

There are several models considering patient characteristics-,donor risks-, and combined D-R-based systems built on conventional biostatistics, mainly logistic regression (LR).All of them have been extensively analyzed and critically reviewed.Unfortunately,none of them has been able to give an adequate response to DR pairing [3].

The Mayo model for end-stage liver disease (MELD) is the most commonly used composite metric for listing candidates based on their severity.The MELD algorithm includes 3 easy-to-determine biochemical parameters that allow to quantify the severity of the candidate’s disease and assign a score based on the sickest-first principle.It is used throughout the world as a prioritization system and has been shown to reduce waiting list mortality but not to impact on post-transplant survival.Furthermore, it allows establishing a minimum score (<15) below which transplantation may not offer the best option for the patient with a liver failure [6].Unfortunately, its drawbacks may outweigh its benefits.MELD lacks of usefulness in predicting post-transplant mortality and does not consider the characteristics of the donor.Basically, the MELD score is not universal (it is only "adequate" in case of patients presenting a liver dysfunction) and it is not useful in its extreme scores, which have been capped (<3 and>40); This score also required the intervention of clinicians not to compromise several indications for liver transplantation by introducing the “MELD exception”in which extra-points are assigned (e.g., hepatocellular cancer) [7].This assignment has been arbitrary in many cases leading to undesirable results [8].Finally, the score has been applied in non-transplant patients for risk assessment in, e.g., ulcerative colitis surgery [9]or liver failure after cardiac surgery [10].For all the here enumerated reasons MELD is not a formula ofE=mc2.

The balance of risk score (BAR) provides a simple and reliable tool to detect unfavorable combinations of donor and recipient factors, and is readily available before the assignment of a liver graft for a specific candidate [11].When compared with similar metrics, BAR score is the best measure to predict 90-day morbidity and mortality with a reasonable accuracy [area under the curve(AUC) = 0.709 and 0.847, respectively][12].Unfortunately, BAR score is not a true “D-R matching”system.Belowa score of 18,liver transplantation can be adequate, irrespective of the D-R combination [3].

The remaining composite metrics range between the simplicity of the D-MELD score [13](combining donor age and MELD)and the greater complexity of the survival outcomes following liver transplantation (SOFT) score [14].Despite multiple attempts to validate them with external databases, none of them was reliable enough to be universally adopted.

From a mathematical point of view, the main problem with all these scores is that they are based on biostatistical models that include LR.These models assume the following four precepts:

?1) The variables can have a linear relationship and specific risks can be established with each of them.While this simplification is operative to our minds and allows to establish some clarity within the scientific method, this is far from being a precise and exact strategy.Medical problems (and almost all biological phenomena) are based on non-linear relationships far from LR.

?2) LR starts from the assumption of independence of the variables.

?3) Biostatistical models try to obtain simplified models,through a selective purge of "non-significant" variables.This is a mathematical game that just does not happen in biology.In nature, all variables matter.

?4) The last of the drawbacks of modern biostatistics refers to its ability to predict unbalanced phenomena.In situations in which the phenomenon is frequent, LR has an adequate predictive capacity and allows to extract the risk factors.However, many of the medical problems are not balanced, which means they comprise majority group(s) and minority group(s).In the former, the prediction is acceptable; with a small sample size, the statistical models have a very low predictive capacity.In the field of liver transplantation, this situation is relatively frequent: the deaths of transplanted patients are proportionally scarce compared to survivors; and very long-term survivors are fewand therefore,the prediction of events is poor in this population.The only solution to solve this drawback is to use large cohorts of patients where infrequent phenomena are sufficiently present(and therefore predictable).

Human decisions and decisions assisted by AI

The best available metrics are capable of giving an adequate prediction of 70% in each D-R match (e.g., BAR score).Some models are absolutely incapable of predicting post-transplant evolution(e.g., for the MELD score this value reaches 50%), which are comparable to heads or tails decisions.Therefore, its usefulness in supporting the clinical decision of the best D-R match is limited.

Clinical decisions are a compendium of objective and subjective thinking.Objective thinking is based on scientific evidence, memory, and previous experiences.But our subjective thinking also influences decisions based on intuitive and, emotional intelligence,as well as on the subconscious.To a lesser or greater extent, a D-R match is influenced by empathy and emotional resilience.Together with the available evidence (“head-on”) and experience and technical expertise (“hands-on”), the decision of this match may be affected by other subjective aspects (“heart-on”).

Fig.1.Classifiers based on AI most useful for D-R matching.Artificial neural networks and random forest have been used successfully and have demonstrated their superiority over biostatistical-based models.AI: artificial intelligence; D-R: donor-recipient.

In a singular D-R match, more than 100 parameters can be handled, between donor factors, recipient variables and (certain) logistical aspects.In addition, a given donor must be assigned to an indeterminate number of candidates on the waiting list, presenting a myriad of clinical situations.A correct assignment therefore exceeds by far our decision-making capacity, although we usually do it correctly in 3 out of 4 cases.To improve this aspect, objective classifiers (“heart-off”), based not only on evidence, memory and experience, but also on big data and multi-objective algorithms must be used.A more precise and reliable solution is the use of classifiers based on AI and, more specifically, deep-learning classifiers.

Best AI classifiers in organ transplantation

Deep-learning offers multiple classifiers that could be used in practically all areas of modern medicine [15].A serious problem,for those unfamiliar with these models, is the selection of the most appropriate classifiers for the specific problem to be studied( Fig.1 ).

In recent years, classifiers based on AI have been used in some areas of hepatology: to predict the risk of cirrhosis due to viral hepatitis; the survival in patients with primary sclerosing cholangitis and to classify elastography patterns [16].Specifically, in liver transplantation, models based on AI have been used to predict long-term post-transplant outcomes, to improve D-R matching and decision-making in transplant oncology, to predict sepsis of transplant recipients in intensive care units (ICU), to recognize pattern through precision pathology, and finally to individually adjust immunosuppressive therapies ( Fig.2 ).

Among hundreds of possible classifiers, ANN and RF have been used for D-R matching.In our experience, Bayesian models have not offered advantages over metrics based on biostatistics (unpublished results).ANN mimics the functional architecture of a biological neuron.In summary, it consists of three differentiated parts: a layer of input variables (input layers) and a layer of output variables (output layers); between them exists a third layer of hidden variables and/or relationships generated from the input layers (hidden layers).For the process of building, the database is divided into two independent parts: a larger one (comprising 75% or 90% of the cases) used for training (training group) and a smaller one (comprising the remaining cases not included for training) in which the results obtained are checked (testing group) [17].In situations where one of the output targets is very infrequent (very unbalanced problems), the database can be randomly split several times sequentially into the training and testing groups to ensure that all infrequent cases are included (and therefore predictable).Typically in these situations the database is randomly divided into ten different times (ten-fold randomization).The ANN-based model has several inherent characteristics: (1) the input variables establish relationships with different strengths; (2) the weight of apparently useless variables under study can be weaker (but not null); in ANN, all variables count; (3) the output variables can be a single one (uni-objective algorithm) or several outputs can be explored(multi-objective algorithms), and based on this, with each algorithm an individual probability can be obtained; (4) by definition,ANN-based models are black-box models because the precision of the result can be known, but not the value of the differentiated risk of each of the model variables; (5) due to its way of building,ANN can have the problem of overtraining and overfitting, which can decrease its predictability and (6) models based on ANN are sensitive to the existence of missing data, considerably reducing their prediction capacity when there is a high proportion of missing data in one or more of their input layers [ 18 , 19 ].

Random forest classifiers are deep-learning models based on the construction of decision trees.For each of the possible outputs, a different decision tree is built.The database is "split"into different nodes at the discretion of the researcher.They require little pre-treatment of the database and, therefore, avoid over training and overfitting of the data.Unlike ANNs, which improve the greater the number of cases, the RF models can be used in smaller databases.These classifiers work well with missing data.As with biostatistics-based models, a limited number of variables with their corresponding real risk can be obtained [20].Unfortunately, RF models are not useful with very large databases because the number of decision trees can become impossible to manage.Therefore, faced with very unbalanced problems that require large databases, RF models are not suitable.

ANN for D-R matching

In essence, a liver transplantation can be divided into three clearly differentiated phases: a pre-transplant period, the transplant itself and a post-transplant period.In each of them a multitude of variables intervene and appear.In the pre-transplant phase,there are only candidate variables.At the time of transplantation,variables of the donor and recipient, related to logistics and intraoperative events are combined.In the postoperative phase, donor and recipient variables converge, along with other external and internal variables (e.g., immunosuppression,de-novocomorbidities)( Fig.3 ).

For the D-R match, the factors of the candidates on the list and the factors of the donor at the time of the offer must be considered.One can also include important variables unknown at the time of donor offer, but roughly estimated (e.g., cold ischemia time, degree of steatosis).The latter are pervasive variables, andhow they affect or not the final model can be analyzed.Therefore,the database will include data from the donor, the recipient and those indirectly estimated.Those variables with a high percentage of missing values (>10%) should be eliminated, since the prediction of the model would not be reliable.Once the database is preprocessed, it is divided into two groups: one for training (training group; 75% of the D-R pairs) and one for verifying the prediction of the model (testing group; 25% of the D-R pairs).These percentages can be modified to 90%/10% training-testing, according to the prevalence of the objective to be analyzed.In the case of D-R matching, the objectives are the survival of the recipient and of the graft in the short and medium term, phases in which donor factors and those derived from the ischemia-reperfusion injury take part preferentially, as well as the candidate’s factors that respond to transplantation (e.g., those related to the urgency of the candidate on the waiting list).

Fig.2.Modern applications of artificial intelligence classifiers in the field of hepatology.D-R: donor-recipient.

Fig.3.For each liver transplantation, three phases can be identified: pre-transplant, transplant and post-transplant phases.Donor and recipient factors are combined with logistical and external factors.All these variables can be recruited to build a deep learning-based D-R matching model.D-R; donor-recipient; BMI: body mass index; DCD:donation after cardiac death; I/R: ischemia/reperfusion injury; NASH: non-alcoholic steatohepatitis; ECD: extended criteria donors.

Based on the results worldwide published, graft and recipient survival is an unbalanced problem in practically all series.There is a very likely outcome (surviving recipient/functioning graft) and a rare outcome (recipient death/graft loss).This situation may force the use of database randomization techniques, as previously described (ten-fold randomization).In summary, for D-R matching a model that includes several different algorithms (multi-objective algorithms), an algorithm can be built to predict graft/recipient survival; and another to predict recipient/graft loss.In mathematical terms, for the majority class (recipient/surviving graft) the concept of correct classification rate (CCR) or accuracy is used.It is defined as the percentage of correctly classified training patterns.This model tries to maximize the probability that a D-R pair belongs to the “surviving”class.For the second algorithm, whose objective is to predict recipient/graft losses (minority class), the concept of minimum sensitivity (MS) is used.It is defined as the minimum value of the sensitivities of each of the classes.The model tries to maximize the probability that a D-R pair belongs to the“non-surviving class”.For each D-R pair, a probability of successful outcome (ANN-CCR) and of loss (ANN-MS) is obtained, allowing so to obtain, two probabilities for each D-R pair.However, if only both algorithms are obtained, two problems arise.(1) Naturally, the model would always choose the pair with the highest numerical probability of surviving and the lowest numerical probability of being lost.As an example between a pair with a graft survival of 98.4% and another with a survival of 98.2%, the model will always choose the former.From a clinical point of view, this difference is irrelevant.(2) Consequently, the model would choose the most favorable combinations that, naturally, would benefit the candidates in a better clinical situation, while the sickest ones will be hardly matched.To face these two problems, that is, on the one hand, to establish the quantitative limits so that the numerical results have clinical relevance; and, on the other hand, that the sickest-first principle is respected, two actions must be taken.The first of these is to apply the ANN-CCR and ANN-MS algorithms only to those candidates with the greatest urgency (for example,the highest MELD patients on the list).The second action is to establish a rule-based system capable of establishing when a survival probability difference in the ANN-CCR model and a loss probability difference with the ANN-MS model are relevant.Likewise, the rule-based system must resolve possible ties between D-R pairs.In order to see the superiority of the ANN models, it is useful to compare them with the results that would be obtained with other conventional statistical metrics (e.g., LR) and with the published scores based on them (e.g., MELD, BAR, SOFT scores) ( Fig.4 ).

Fig.4.Steps for the building of a D-R matching model of survival and non-survival in liver transplantation based on ANN.After the database processing, a training group and a testing group are created.Two algorithms are obtained (ANN-CCR and ANN-MS) that are applied following a rule-based model on the set of the sickest candidate list.D-R: donor-recipient; ANN: artificial neural networks; CCR: correct classification rate; MS: minimum sensitivity.

A multicenter study from 11 Spanish liver transplant units including 1003 consecutive liver transplantation performed between 20 07 and 20 08 was conducted to look at 3-month graft survival.Fifty-seven variables (26 from the recipient, 19 from the donor, 6 from the retrieval procedure and 6 from the transplant procedure)were reported for each D-R pair.ANN-CCR and ANN-MS models were built with a ten-fold randomization.ANN-CCR reached a 90.79% prediction of the probability of graft survival and an AUC of 0.80.ANN-MS showed the best capacity for graft loss prediction with 71.42% for the best model and an AUC of 0.82.As expected,models based on LR were able to predict graft survival, although slightly inferior to ANN models.The BAR score reached the best survival prediction probability (AUC = 0.68).However, the ability to predict graft loss from all these scores was very low [21].

In order to explore and validate this methodology, a second study was carried out with 858 consecutive D-R transplantation pairs performed at King’s College Hospital (KCH) in London during the same period of time ( Fig.5 ).Models designed for KCH had excellent prediction capabilities for both 3 months (ANN-CCR AUC = 0.94; ANN-MS AUC = 0.94) and 12 months (ANN-CCR AUC = 0.78; ANN-MS AUC = 0.82, almost 15% higher than the best obtained by other scores such as BAR).Moreover, these results were superior to those reported in the Spanish study.The differences between the two studies were explained by the inherent problems of the ANNs described in the previous section including the great difference between input variables related to both donor and recipient populations (e.g., different donors, indications, race proportions).Furthermore, the KCH database was more homogeneous and the percentage of missing values was significantly lower.This means that each ANN model needs to be trained for a specific purpose in a single distinctive population [22].

Recently, our group has tried to analyze how ANNs work in a big database using D-R pairs from the United Network for Organ Sharing (UNOS); a total of 39 189 D-R pairs with 4 different endpoints (3-month and 1-, 2- and 5-year recipient survival rates) were included.Two groups of modeling techniques were used: 1) classical statistical methods, including LR and Na?ve Bayes,and 2) standard machine learning techniques, including ANN, RF,gradient boosting, and support vector machines, among others.For the 5-year endpoint, LR (AUC = 0.654) outperformed various machine learning techniques, such as ANN (AUC = 0.599) or RF (AUC = 0.644) [23].The reason why these results are so different from those obtained in our previous studies with singlecenter and/or more limited databases was the presence of a considerable percentage of missing values for many relevant variables in the UNOS database.In fact, after removing all variables with missing values>10%, only 28 variables remained in the model.This disadvantage of AI models with the UNOS database has also been found for the same reasons in heart transplantation [24].AImodels demonstrated similarly modest discrimination capabilities compared with traditional models (C-statistic ≤0.66).The ANN model demonstrated the highest C-statistic (0.66) but this was only slightly superior to the simple LR (C-statistic = 0.65).

Fig.5.Validation of ANN methodology in an external database [King’s College Hospital (KCH)].DB: database; ANN: artificial neural networks; MELD: model for end-stage liver disease; D-MELD: donor age and MELD score; DRI: donor risk index; SOFT: survival outcome following liver transplantation.

As noted above, the predictability of an AI model depends on the database where it was created and cannot be extrapolated to a different database.It is important to note that the homogeneity of the database determines the predictability of the classifier.This statement assumes that the ANNs for D-R matching can only be applied in databases that followsimilar rules for prioritization and inclusion of candidates, and that are also sufficiently homogeneous.ANN-based models will work very well in local and regional liver transplantation programs, but they will not work as universal systems with high dispersion criteria.

RF for D-R matching

Models based on RF methodology have recently been used to test their ability to predict post-transplant graft failure and have been compared with conventional metrics such as the DRI [25]and the MELD score.In a small cohort of 180 liver transplantation performed between 2013 and 2015 in the Melbourne Center, Lau et al.demonstrated the superiority of RF with an AUC of 0.787, compared to ANN (AUC = 0.734) and DRI (AUC = 0.595).More interesting was the ability of this methodology to select in the final model 15 out of 276 variables with an AUC of 0.715.The percentage of missing values in these 15 top variables ranged between 0 and 72.22%, with 5 variables having>10% missing values (27.78%-72.22%).This fact demonstrates the ability of the RF models to include variables with a high percentage of incomplete data.However, RF classifiers are of little use when using larger databases due to the large number of decision trees they can generate.In addition, they require the selection of significant variables and the elimination of others, a methodology similar to the one used in LR [17].

Future perspectives

AI-based models are being used with increasing frequency in the field of organ transplantation and, specifically, in the prediction of graft and recipient survival, as well as in the assignment of a specific donor to a (list of) potential recipient(s) without affecting the principles of urgency and utility (survival benefit).The current available studies still harbor many design problems, and the most important challenge is that the generation of AI-based models from imperfect databases and the poor applicability to predict outcome in other databases than those from which they were created.

It is also not very well-known which classifier fits best to each situation, or if a different classifier is needed for each challenge.Regardless of these assertions, it seems obvious that AI-based methodology significantly outperforms the decision-making process based on LR models.Based on data from the US Scientific Registry of Transplant Recipients (SRTR) including 190 input variables from 42 146 patients, Nitski et al [26].assessed the ability of deep-learning algorithms to predict complications resulting in death after liver transplantation during different time periods.The obtained results were tested in a cohort of 3269 patients with 63 input variables registered in the Canadian University Health Network (UHN).As both databases presented a low number of missing values, their prediction capacity was very reliable.The ANNs models generated AUCs ranging from 0.847 to 0.871 for prediction of death by graft failure within the first posttransplantation year; these values were reproduced in the testing group.

The results of liver transplantation impact on patients and available resources.A functioning graft implies a lower economic investment compared to an initially dysfunction or primary nonfunction graft due to a longer ICU hospitalization time, a need for re-transplantation or both.A methodology capable of predicting the best performance of grafts in patients on the waiting list would markedly reduce costs, leading to both an individual and social benefit [27].The future frontiers that AI-based algorithms should overcome in order to generalize D-R matching are displayed in Table 1.Probably, the most important one is to overcome the mental barrier of the clinician to combine statistical metrics and models based on deep learning [18].At the moment, AI models are not self-driven cars!

Table 1Future barriers to overcome for the implementation of AI models in D-R matching [18].

In conclusion, although the use of AI-based models is in its infancy, there is no doubt that in the future they could represent invaluable tools for decision-making in D-R pairing and the management of the liver transplantation waiting lists.

Acknowledgments

None.

CRediT authorship contribution statement

Javier Brice?o:Conceptualization, Funding acquisition, Methodology, Supervision, Writing - original draft, Writing - review & editing.Rafael Calleja:Data curation, Writing - review & editing.CésarHervás:Conceptualization, Supervision, Writing - review & editing.

Funding

This work was supported by a grant from Mutua Madrile?a XVIII Convovatoria de ayudas a la investigación.

Ethical approval

Not needed.

Competing interest

No benefits in any form have been received or will be received from a commercial party related directly or indirectly to the subject of this article.