Received: 12 February 2022

| Revised: 1 June 2022

| Accepted: 3 June 2022

DOI: 10.1111/aor.14334

S Y S T E M AT I C R E V I E W

Machine learning and artificial intelligence in cardiac

transplantation: A systematic review

Vinci Naruka1,2 | Arian Arjomandi Rad2 | Hariharan Subbiah Ponniah2 |

Jeevan Francis3 | Robert Vardanyan2 | Panagiotis Tasoudis4 |
Dimitrios E. Magouliotis | George L. Lazopoulos
4 4,5
| Mohammad Yousuf Salmasi2 |
Thanos Athanasiou1,2,4

1
Department of Cardiothoracic
Surgery, Imperial College NHS Trust, Abstract
Hammersmith Hospital, London, UK Background: This review aims to systematically evaluate the currently available
2
Department of Surgery and Cancer, evidence investigating the use of artificial intelligence (AI) and machine learn-
Faculty of Medicine, Imperial College
London, London, UK
ing (ML) in the field of cardiac transplantation. Furthermore, based on the chal-
3
Faculty of Medicine, University of lenges identified we aim to provide a series of recommendations and a knowledge
Edinburgh, Edinburgh, UK base for future research in the field of ML and heart transplantation.
4
Department of Cardiothoracic Surgery, Methods: A systematic database search was conducted of original articles that ex-
University Hospital Thessaly, Larissa,
plored the use of ML and/or AI in heart transplantation in EMBASE, MEDLINE,
Greece
5
Department of Cardiac Surgery,
Cochrane database, and Google Scholar, from inception to November 2021.
University Hospital of Heraklion, Crete, Results: Our search yielded 237 articles, of which 13 studies were included in this
Greece review, featuring 463 850 patients. Three main areas of application were identi-
Correspondence fied: (1) ML for predictive modeling of heart transplantation mortality outcomes;
Arian Arjomandi Rad, Faculty of (2) ML in graft failure outcomes; (3) ML to aid imaging in heart transplantation.
Medicine, Department of Medicine,
The results of the included studies suggest that AI and ML are more accurate in
Imperial College London, South
Kensington Campus, Sir Alexander predicting graft failure and mortality than traditional scoring systems and con-
Fleming Building, London, SW7 2AZ, ventional regression analysis. Major predictors of graft failure and mortality iden-
UK.
tified in ML models were: length of hospital stay, immunosuppressive regimen,
Email: arian.arjomandi-rad16@
imperial.ac.uk recipient's age, congenital heart disease, and organ ischemia time. Other poten-
tial benefits include analyzing initial lab investigations and imaging, assisting a
patient with medication adherence, and creating positive behavioral changes to
minimize further cardiovascular risk.
Conclusion: ML demonstrated promising applications for improving heart trans-
plantation outcomes and patient-­centered care, nevertheless, there remain im-
portant limitations relating to implementing AI into everyday surgical practices.

KEYWORDS
artificial intelligence, cardiac transplantation, heart transplantation, machine learning

Vinci Naruka, Arian Arjomandi Rad, and Hariharan Subbiah Ponniah contributed equally to this study.

This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided
the original work is properly cited.
© 2022 The Authors. Artificial Organs published by International Center for Artificial Organ and Transplantation (ICAOT) and Wiley Periodicals LLC.

Artificial Organs. 2022;46:1741–1753.  wileyonlinelibrary.com/journal/aor | 1741

1742 |    MACHINE LEARNING IN CARDIAC TRANSPLANTATION

1 | BAC KG ROU N D This review aims to systematically evaluate the currently

available evidence investigating the use of artificial intelli-
Heart transplantation remains the definitive treatment for gence and machine learning in the field of cardiac trans-
patients with end-­stage heart failure. While the number plantation. Furthermore, based on the challenges identified
of heart transplants across the world has increased, the we aim to provide a series of recommendations and a knowl-
supply of heart donors is yet to increase enough to meet edge base for future research in the field of ML and heart
the demand; therefore, bringing the issues of resource al- transplantation, ultimately aiding patient-­centered care.
location into question.1 The process of graft allocation is
complicated, having to consider both patient and donor
characteristics in pre-­, peri-­and post-­operative settings, 2 | METHODS
thus illustrating the multidimensional nature of the
matching process. Previous studies in heart transplanta- 2.1 | Literature search strategy
tion have demonstrated the use of points-­based scoring
systems, using a selection of identified variables, in order A systematic review was conducted in accordance with
to predict the main endpoints of mortality and graft fail- the Cochrane Collaboration published guidelines and
ure, but such studies observed poor predictability.2 With the Preferred Reporting Items for Systematic Reviews
the increase in demand for donor hearts, prediction of and Meta-­ Analyses (PRISMA) statement. MEDLINE,
a successful transplantation becomes absolutely para- EMBASE, PubMed, Cochrane, and Google Scholar were
mount, and predictability could be improved by inputting searched for original articles from inception to November
a more extensive and updated donor and recipient infor- 2021 that discussed:
mation and the utilization of a more powerful analysis,
machine learning.3 • (P) Patients undergoing cardiac transplantation.
The use of artificial intelligence (AI) has the potential • (I) The use of Artificial Intelligence (AI) or machine
to revolutionize clinical practice. Machine learning (ML) learning (ML).
enables the identification of non-­ linear relationships • (C) Current algorithms used to predict outcomes, if
and contributing variables that have conventionally been available.
thought to be of limited use.4 Utilizing such variables • (O) Outcomes including mortality, graft failure, and
using a ML model allows clinicians to accurately predict their predictors.
prognosis post-­transplantation, quantify the risk of rejec-
tion, and ascertain waitlist mortality for those who may The search terms used included (heart transplantation
not survive long enough to receive a heart, as already illus- OR cardiac transplantation OR heart transplant OR cardiac
trated in kidney and liver transplant recipients.5,6 Previous transplant OR heart allograft OR cardiac allograft OR heart
studies by the International Society of Heart and Lung heterograft OR cardiac heterograft OR heart homograft OR
Transplantation (ISHLT) have attempted to investigate cardiac homograft) AND (machine learning OR artificial in-
mortality rates and ascertain the variables most predictive telligence OR deep learning OR Decision Trees OR Neural
for patient's post-­transplant by utilizing traditional regres- Networks). Further articles were identified through the use
sion models and multivariable analysis.7,8 These models of the “related articles” function on MEDLINE and a man-
remain underutilized in clinical practice due to their rel- ual search of the references lists of articles found through
atively weak and variable predictive powers of outcomes the original search. The only limits used were the mentioned
that are multidimensional in nature. time frame and the English language.
ML models can analyze more variables than traditional
models to thereby build new co-­variate relationships and
identify variables most influential in a particular process. 2.2 | Study inclusion and
Traditional statistical models aim to ascertain the proba- exclusion criteria
bility of an event occurring due to a particular variable.
Furthermore, ML models allow for a greater number of All original articles were included reporting the use of
associated variables to be studied and then build a model machine learning or artificial intelligence in cardiac trans-
based on parameters that influence the outcome the most. plantation. Studies were excluded from the review if: (1)
In cardiac transplantation, this could guide clinicians in inconsistencies in the data impeded extraction of data and
decision making on the allocation of hearts for transplan- (2) the study was performed in an animal model. Reviews,
tation, increase accuracy in predicting graft failure and case reports, preclinical studies, and abstracts from meet-
mortality, and predict those at highest risk for rejection ings were excluded. By following the aforementioned crite-
post-­transplantation. ria, two reviewers (H.S.P. and J.F.) independently selected
MACHINE LEARNING IN CARDIAC TRANSPLANTATION    | 1743

F I G U R E 1 Risk of bias diagram

[Color figure can be viewed at
wileyonlinelibrary.com]

articles for further assessment following title and abstract 2.4 | Risk of bias
review. A third independent reviewer (A.A.R.) resolved any
disagreements between the two reviewers. Potentially eligi- The risk of bias in the selected articles was evaluated by
ble studies were then retrieved for full-­text assessment. two independent reviewers (A.A.R. and H.S.P.) using an
adapted cochrane collaboration risk of bias tool (Figure 1).
The methodological quality of the studies was assessed
2.3 | Data extraction and critical based of domains: (1) Study Participation, (2) Study
appraisal of evidence Response, (3) Outcome Measurement, (4) Statistical
Analysis and Reporting, (5) Study Confounding. An over-
All full texts of retrieved articles were read and reviewed all grading of low, medium, or high risk of bias was then
by two authors (H.S.P. and J.F.) and a unanimous decision allocated.
was made regarding the inclusion or exclusion of stud-
ies. When there was disagreement, the final decision was
made by a third reviewer (A.A.R.) Using a pre-­established 3 | RESULTS
protocol, the following data were extracted: first author,
study design, machine learning technique(s) used, popula- 3.1 | Study selection
tion number, and main outcomes. A data extraction sheet
for this review was developed and pilot-­tested using 3 ran- A total of 237 articles were identified in the literature
domly selected included studies and subsequently was re- search, of which 180 were screened following deduplica-
fined accordingly. Data extraction was performed by two tion and were read in full and assessed in accordance with
review authors (H.S.P. and J.F.). The correctness of the the inclusion and exclusion criteria. A total of 13 studies
tabulated data was validated by a third author (A.A.R). were included in this review following critical appraisal,
1744 |    MACHINE LEARNING IN CARDIAC TRANSPLANTATION

featuring 463 850 patients. The entire study selection pro- what data is available. As such, more data are required to
cess is illustrated in Figure 2. A summary of the studies assess the factors which cause and can predict long-­term
collected and their respective designs, type of outcomes outcomes in the post-­heart transplant patient. It was in-
measured, and its implementation as well as the main re- teresting to note that the time horizon played a part in
ported outcomes are found in Table 1. which variables were most predictive, meaning that pre-
dictive variables were found to differ for 1-­year mortality
compared to 5-­year mortality.16 This calls for a wider array
3.2 | Prediction of graft of data sets to be collected to accurately model factors that
failure and mortality are most influential for specific outcomes, for instance,
waitlist mortality versus 3-­year mortality. Despite this,
There were 12 studies that discussed the use of machine even the current ability of the MI models to predict graft
learning in predicting mortality in heart transplantation failure and morality is a welcome improvement to the
patients,9–­20 comprising 463 807 patients and included a donor graft and recipient matching process and thus pro-
conglomerate of different modeling methods. There was vides a more efficient use of the current limited resources
1 study that discussed the use of machine learning in pre- and thus reduces waiting times and improving prognosis
dicting graft failure in heart transplantation,17 the study for patients.
comprised 15 236 patients.
The outcome of the included studies suggests that AI
and ML are generally more accurate in predicting graft 3.3 | Imaging
failure and mortality than conventional regression anal-
ysis. The study by Kampaktsis and colleagues found that There was 1 study that discussed the use of machine learn-
ML models generally had good predictive power when as- ing within an image-­based context in heart transplanta-
sessing 1-­year outcome, but its predictive power declined tion,21 this comprised 43 patients. Tong et al. developed a
for later outcomes.18 A patient’s journey post-­transplant deep neural network that can identify histological slides
is complex and most likely to be affected by a variety of that fit into rejection and non-­rejection cohorts.21 The re-
multi-­system pathologies observed in the aging popu- sults yielded far more accuracy than manually determin-
lation. ML models can only make predictions based on ing which slide was to be potentially rejected.

FIGURE 2 PRISMA flow chart [Color figure can be viewed at wileyonlinelibrary.com]

TABLE 1 Summary of the studies included in the systematic review

Type of study;
Country; Database Algorithm/model/method of Population
Study used implementation number Aim Main reported outcomes
21
Tong et al. M, P; USA; Whole Deep Neural Network 43 Automate prediction of heart Shape and distribution of nuclei in tissue images
slides from transplant rejection using dominate algorithm prediction. NN can
the Children's histopathological whole-­slide significantly reduce overfitting and achieve
Healthcare of imaging more stable accuracy compared to NN without
Atlanta regularization and drop out
Medved et al.11 M, NP; USA; UNOS ANN 27 444 Predict outcome 180, 365, 730 days Extracted top 10 variables (weighted by importance)
after entering HTx list (Outcome that affect outcome
include waiting, transplanted, or
dead)
Medved et al.10 M, NP; USA; UNOS NN (IHTSA and LuDeLTA) 49 566 ML predicts the status of the patient The predicted mean survival for allocating according
MACHINE LEARNING IN CARDIAC TRANSPLANTATION

in the que and then the post-­ to wait time was about 4300 days, clinical rules
transplant survival 4300 days, and using neural networks 4700 days
Medved et al.13 M, NP; USA; UNOS IHTSA + IMPACT 27 705 Compare two risk models (IHTSA IHTSA model had improved performance and
and IMPACT) to predict short-­ accuracy compared to the IMPACT model. IHTSA
and long-­term mortality after shows better discrimination on one-­year mortality.
heart transplantation IHTSA predicts short-­term mortality with greater
accuracy than traditional risk-­based models based
on logistic regression. AUROC: IHTSA (0.643) and
IMPACT (0.608). Calibrated IHTSA (0.688) and
IMPACT (0.606)
Yoon et al.16 M, NP; USA; UNOS Trees of Predictors 95 275 Construct a tree of predictors (ToPs) AUC for 3-­month was 0.660, while the best clinical
and utilize its predictive power for risk scoring method only achieved 0.587. ToPs
predicting 3 months, 1-­, 3-­, 10-­year achieved better prediction of both survival and
mortality after HTx mortality. ToPs identifies the most relevant
features and is adaptable to changes in clinical
practice
Miller et al.14 M, NP; USA; UNOS ANN, CART, RF 2802 Predict 1-­, 3-­, 5-­year mortality after Good predictive value for mortality but poor
pediatric heart transplantation sensitivity. Due to lack of registry data, MLs
ability to predict mortality post-­transplant is
fundamentally limited. AUROC: Calibrated—­RF
Testing 1 Year (1.25), ANN Testing 1 Year (0.73),
CART Testing 1 Year (0.46). RF Testing 3 Years
(0.60), ANN Testing 3 Years (0.26), CART Testing
3 Years (0.38), RF Testing 5 Years (0.86), ANN
Testing 5 Years (0.20), CART Testing 5 Years (0.33)
  
|
1745

(Continues)
TABLE 1 (Continued)
1746
|

Type of study;
Country; Database Algorithm/model/method of Population
  

Study used implementation number Aim Main reported outcomes

9
Miller et al. M, NP; USA; UNOS LR, SVM, RF, Decision Tree, 56 477 Develop a risk-­prediction model for Major univariate predictors of 1-­year mortality were
NN assessing 1-­year mortality post-­ consistent with previous findings and included
heart transplantation using ML age, renal function, body mass index, liver function
tests, and hemodynamics. Machine Learning
models showed similarly modest discrimination
capabilities compared with traditional models
(C-­statistic 0.66, all). The neural network model
showed the highest C-­statistic (0.66) but was only
slightly superior to the simple logistic regression,
ridge regression, and regression with LASSO
models (C-­statistic = 0.65, all)
Hsich et al.19 M, NP; USA; RSF 33 069 Identify variables of importance for Strong and weak predictive variables were identified.
Scientific Registry waitlist mortality using Random Complex interactions were identified such as an
of Transplant Survival Forests additive risk in mortality. Most predictive variables
Recipients (SRTR) for waitlist mortality are in the current tiered
allocation system except for eGFR and serum
albumin which have an additive risk and complex
interactions
Agasthi et al.17 M, NP; USA; ISHLT GBM 15 236 Predict mortality and graft failure Model utilized 87 variables in a non-­linear fashion to
5 years after orthotopic heart accurately predict mortality/graft failure. Provided
transplantation top 10 most influential variables for predicting
5-­year mortality/graft failure. AUROC: Mortality
(0.717) and Graft failure (0.716)
Dolatsara et al.20 M, NP; USA; UNOS LR, XGB, LDA, RF, ANN, CART 103 570 First stage—­use independent First stage produces AUROC between 0.60 and 0.71
machine learning models to for years 1–­10. Second stage of calculating survival
predict transplantation outcomes probabilities guarantees monotonicity
for each time period. Second
stage—­Calibrate survival
probabilities over time using
isotonic regression
Ayers et al.12 M, NP; USA; UNOS Deep Neural Network, LR, 33 657 Predict 1-­year mortality post Ensemble ML model outperformed traditional risk
AdaBoost, RF orthotopic heart transplantation models in predicting mortality. Model was made
from preoperative variables. AUROC: LR (0.649),
RF (0.691), DNN (0.691), Adaboost (0.653). Final
Ensemble ML Model (0.764)
MACHINE LEARNING IN CARDIAC TRANSPLANTATION
 MACHINE LEARNING IN CARDIAC TRANSPLANTATION    | 1747

Abbreviations: AdaBoost, adaptative boosting; ANN, artificial neural network; CART, classification and regression tree; GBM, gradient boost machines; HTx, heart transplantation; IHTSA, international heart transplant
4 DISC USSION

survival algorithm; KNN, k-­nearest neighbor; LDA, linear discriminant analysis; LR, logistic regression; LuDeLTA, Lund deep learning transplant algorithm; M, multicenter, NM, non-­multicenter, NP, non-­prospective;
XGBoost (0.769), GBM (0.786), ANN (0.755), Naïve

1 Year Adaboost (0.689), 3 year Adaboost (0.60528),

(0.801), AdaBoost (0.641), LR (0.688), SVM (0.714),

nearest neighbor models (0.526). IMPACT (0.569).

RF model performed optimal predictive power. Top 5
most important variables for short-­term prognosis

Overall, ML showed good predictive accuracy of

mortality after HTx. AUROC: 1 Year—­Adaboost
Adaboost achieved highest predictive performance.

(0.689), LR (0.642), DT (0.649), SVM (0.637), K-­

was ALB, age, LA, RBC, HB level. AUROC: RF
This systematic review explored the data on utilizing AI
in a heart transplant setting. Thirteen papers were in-
cluded in this study to investigate its use in heart trans-
plantation. The majority of papers discussed the use of
ML models in accurately predicting mortality and survival
post-­transplantation. Others discussed models which pre-

5 Year Adaboost (0.6283)

Main reported outcomes

dict the risk of rejection pre-­transplant and ML use for

predicting waitlist mortality.

4.1 | ML predictors of graft

(0.500)

failure and mortality

4.1.1 | Length of hospital stay

Develop a risk-­prediction model for

Develop a risk-­prediction model for

orthotopic heart transplantation

orthotopic heart transplantation

assessing 1-­year mortality post

assessing 1-­year mortality post

Remarkably, ML models have been found to depend more

on factors that are not of high importance in traditional
statistical models.22 Indeed, when predicting graft failure
and mortality, donor variables such as age were generally
found to be of less importance in ML models, while the
length of hospital stay was of high importance.14,15,17 In
using ML

using ML

this setting, variables that affect length of stay should be

optimized to ensure graft patency and survival, and more
Aim

data points are required to ensure accurate prognostic pre-

dictions. The literature provides no clear explanation for
the strong predictive power of length of hospital stay on
Population

graft patency and mortality. However, increased compli-

number

cations such as bleeding or incidence of infection, and the

18 625
381

P, prospective; RF, random forest; RSF, random survival forest; SVM, support vector machine.

severity of such complications are known to lead to an in-

crease in the length of hospital stay.23,24 Additionally, the
Algorithm/model/method of

Adaboost, SVM, Decision Tree,

length of hospital stay is difficult to ascertain accurately

LR, SVM, RF, XGB, AdaBoost,

pre-­operatively. Therefore, further studies are needed to

establish the potential predictive factors of these clini-
cal outcomes, and its subsequent predictive potential in
implementation

mortality and graft failure. However, while the exact re-

GBM, ANN

KNN, LR

lationship between the length of hospital stay and graft

mortality could not be explained in the current ML mod-
els, there is an unequivocal cost-­benefit of reducing the
length of stay for both patients and healthcare systems.25
Country; Database

M, NP; USA; UNOS

Database from

4.1.2 | Immunosuppression regimen

NM, NP; China;
Type of study;

Hospital

Anti-­rejection immunosuppression medications are typi-

(Continued)

used

cally given post-­transplant, but most databases do not

collect data on patients' immunosuppression regime.13,17
Studies highlighted that this factor was influential in pre-
dicting graft failure, more so than predicting mortality.
15

Kampaktsis
TABLE 1

Zhou et al.

et al.18

This may be due to such regimes decreasing the chances

Study

of host rejection, but causing toxic side effects to the kid-

ney, for example.26 The overall toxicity increases mortality
1748 |    MACHINE LEARNING IN CARDIAC TRANSPLANTATION

but does not affect graft failure as much. It is important to failure mortality.11,14,15,17,18 This is consistent with previ-
note that many databases did not collect data on patients' ously published studies.29
immunosuppression regime, perhaps due to the perceived Two biomarkers found to have a major influence on
lack of importance, and as such, ML models may pave graft failure were pre-­ transplant creatinine and biliru-
for more broader data collection to increase the models' bin.12,14,15,17,18 Previous studies have highlighted this rela-
prognostic accuracy. Incorporation of patient's immuno- tionship as creatinine and bilirubin are useful indicators to
suppression regimes will not only aid prognostic accuracy assess overall kidney and liver health, both of which are cru-
but could also potentially aid with the optimization of im- cial in cardiovascular health in the post-­transplant patient.30
munosuppression regimes for each patient. Episodes of Additionally, transplantation itself could affect kidney func-
graft rejection are associated with subtherapeutic immu- tion due to reduced renal blood flow and the side effects of
nosuppressive drug levels and given the various pharma- potent immunosuppressive drugs post-­ transplantation.31
codynamic and pharmacokinetic factors that are usually Serum creatinine also stands as a biomarker for end-­organ
involved, the use of ML algorithms on large datasets failure, and as such, more identification of biomarkers
would enable a multi-­dimensional analysis of these fac- could pave for predicting the likelihood of organ failure and
tors and thus could potentially identify the ideal regime graft failure by identifying the factors which are conducive
for each patient.27 to an increase in serum creatinine.

4.1.3 | Recipient age and congenital 4.2 | Multi-­level functioning of AI in

heart disease heart transplant: From the laboratory to
post-­transplantation patient care
Recipient age was also found to be an important predic-
tor as opposed to traditional models, which placed impor- The use of ML algorithms is not restricted to predicting
tance on the donor age.14,15,17 It is interesting to note that mortality and graft failure. In combination with other ap-
ML models generally had a better predictive power for pa- plications, AI and ML could aid a patient's journey within
tients above the age of 60. Most of the population in the heart transplantation by predicting the potential benefits
databases were over 60, and, therefore, the models built of transplantation by analyzing initial lab investigations
on these data were more suited to patients over that age and imaging, ascertaining graft failure and mortality after
threshold. Younger patients may have unmeasured vari- transplantation, and assisting a patient with medication
ables, including variables influenced by congenital heart adherence and creating positive behavioral changes to
disease (CHD), for example, which were unaccounted for minimize further cardiovascular risk.
in most models. One study highlights that the diagnosis Endomyocardial biopsy is a gold standard investiga-
of CHD was the most crucial factor in 1-­, 3-­, and 5-­year tion to screen for the risk of heart rejection. Due to the
mortality.14 With the rise in adult CHD prevalence and time-­consuming nature of screening all histological slides
surgery, and its potential implications in transplantation, manually, utilizing AI and ML could offer an alternative
more data and models are required to ascertain the utility approach to identifying those at risk of rejection.32
of AI within specific age groups.28 This includes incorpo- Medved et al. discuss the use of AI in the allocation
rating CHD-­specific variables that would otherwise not be of hearts and predicting waitlist mortality.10 Two models
needed in the general adult cardiac patient. were created, one which simulated the removal of a pa-
tient from the waitlist and the other to predict survival
post-­transplant. The survival of the patients allocated by
4.1.4 | Other major factors ML models was also evaluated. The results showed that
patients allocated by deep neural networks had reduced
As reported in previous studies, prolonged ischaemic time waitlist mortality and longer survival post-­transplant.
was also found to be significant in predicting graft failure AI has the potential of assisting patients and clinicians in
and mortality.12,16–­18,28 Its influence on 5-­ year mortal- assessing patient-­specific responses to post-­transplant med-
ity was not so significant. This is unsurprising given the ication. Previous studies in other fields have investigated
fact that hearts undergoing prolonged ischemic time were the use of ML models simulating patient-­specific responses
more likely to fail during the initial stages after transplan- to treatment, to subsequently indicate what treatments pa-
tation; hence, having a higher predictive power for 1-­year tients should and should not receive. A study by Labovitz
mortality than 5-­year mortality. Additionally, donor BMI et al. investigated the use of an AI application on smart-
and recipient BMI were found to increase the risk of graft phones to improve patient adherence to anticoagulation.33
The utilization of AI to positively impact patient behavior
MACHINE LEARNING IN CARDIAC TRANSPLANTATION    | 1749

on their adherence to medication management will be of prior to transplantation.17 Hence, while ML identifies this
particular use in the post-­transplant patient, due to the myr- factor as highly prognostic of the endpoints in question,
iad of treatments patients receive for immunosuppression, information regarding the risk factors associated with the
anticoagulation, antihypertension, and others. length of hospital stay needs to be identified, optimized,
and then incorporated into the algorithm. Additionally,
the notion of length of the hospital being the strongest pre-
4.3 | Challenges with the dictor of mortality as opposed to donor age also questions
implementation of ML the degree of heterogeneity between the models and more
importantly, the databases on which they were developed,
Despite ML proving to be better performing at predicting as well as the methodology that was used, with some stud-
endpoints such as mortality and graft failure, both multi- ies excluding post-­operative variables.
dimensional in nature, as compared to more traditional It must also be noted that the predictive ability of the
methods of scoring systems and regression models, there ML models is as only strong as the initial data set was de-
are a few hurdles currently preventing wider implemen- veloped and validated with.34 Studies to date have used a
tation. One such challenge is ascertaining the risk factors range of databases and variation was observed even within
for post-­op variables that are identified as predictors of the the derivation and validation subgroups derived from the
endpoints (Figure 3). Agasthi et al. discuss the most impor- same database. Furthermore, what was even more hinder-
tant factor in mortality and graft failure to be the length ing was the number of variables that were removed purely
of hospital stay—­a post-­op variable that is hard to predict due to a lack of completeness and granularity. A potential

F I G U R E 3 Challenges and
recommendations of ML in heart
transplantation research. [Color figure
can be viewed at wileyonlinelibrary.com]
1750 |    MACHINE LEARNING IN CARDIAC TRANSPLANTATION

reason for this is the fact that many of the databases incor- factors such as immunosuppression regime or the causa-
porated data from the 80s and 90s—­since then, computa- tive factors for length of hospital stay were not studied.
tional methods, the use of electronic health records, and Our analysis has shown that such factors may pave way
improvements to data collection guidelines have greatly for a more powerful predictive capacity for ML models.
improved the quality of the datasets with regards to both Such coherent models may allow surgeons of the future to
granularity and validity of the data. Despite this, it is make better decisions on the allocation of hearts, manage-
worth noting that the use of registries can still blunt the ment of patients post-­transplantation, and guide patient
phenotyping of complex patients which could ultimately in the decision for heart transplantation.
affect the predictive ability of ML methods.9 Further in- Yoon et al. discuss the use of Tree of Predictors (ToP).16
spection of the datasets used raises concerns about the ho- ToPs are predictive models which set binary rules to iden-
mogeneity of the databases—­while subgroup analysis was tify the strongest dependent variables for predictions.35
attempted in many of the studies, the nature of data sets, Each tree consists of branches, nodes, and leaves. With
predominantly based on a Caucasian cohort, limits the nodes having further sub-­nodes. In this case, patients
generalisability of the models. Zhou et al. did demonstrate were split into clusters and sub-­clusters based on their
the effectiveness of ML models for assessing the progno- specific patient-­donor compatibility features, and a model
sis of heart transplantation patients in a predominantly was created for each specific cluster. This allows for iden-
Chinese population, however, any consequential contri- tifying the most relevant covariant for predictive models
bution to generalisability was limited by the small sample and utilizes them to ensure greater accuracy in predicting
size (381 patients) as well as the focus on the short-­term survival pre-­and post-­transplantation. Greater optimiza-
prognosis, thus highlighting the need for further studies tion and personalization for patient decision making in
before wider implementation of ML models.15 cardiac transplantation will allow for better allocation
Moreover, the heterogeneity in the ML methods used of resources in a clinical setting. However, these models
as well as what the endpoints measured limits compari- utilized patients from retrospective studies, and as such,
son between the current studies, and hence the question there remains a risk of overfitting the model to prospec-
of what the best model(s) are remains to be solved. The tive patient cohorts, which may not necessarily translate
need for constant updatability to the various novel inter- to greater accuracy in current clinical practice.
ventions is another area that must be considered when Further validation and development of ensemble mod-
implementing ML algorithms—­previous studies have in- els may allow for a unique website or software, whereby
corporated data from a 30-­or 40-­year period and in this clinicians can input patient variables to calculate the likely
time, novel interventions such as the left ventricular as- prognosis. It is important to note that the results yielded
sist devices have significantly improved patient survival from our studies are likely to be under-­represented the po-
as well as changes to organ allocation sequences.9,16 tential of ML models due to the restricted datasets that
Explainability of ML models also raises reservations with were inputted. This demonstrates the potential for further
regards to wider implementation—­albeit ML models have multiple non-­linear ML models to be combined to hold a
shown to be highly predictive, this often comes at the ex- more predictive power for accurately estimating prognosis
pense of explainability to both patients and clinicians and post-­transplantation. Additionally, the methodology de-
at times, there is a great theoretical and practical divide veloped in our studies can be applied in other specialties
in the factors identified in machine learning algorithms to form a wider application of ML models.
and the current clinical practice. An example of such mis-
match was seen by Zhou et al. who observed smoking to
be a protective factor—­contrary to previous literature and 4.5 | Nationwide data accessibility
scientific understanding.15
Machine learning algorithms are notably data-­driven and
perform optimally in scenarios where training models are
4.4 | Future steps of AI-­assisted heart developed using larger databases. A common issue with
transplantation smaller databases is that they contain a disproportionately
larger quantity of poor data points, as well as outliers and
AI and ML techniques have been proven to generally random errors. As a result, they encourage the principle of
improve the accuracy of predicting prognosis post-­heart overfitting whereby a machine learning algorithm models
transplantation. Although highly predictive, such models the data to include these erroneous points and inciden-
still require a validated dataset. This includes the need tally describes random errors rather than the interplay
for prospective multi-­center studies collecting data on the between variables in a dataset. This forces the outcomes
various elements of heart transplantation. Conventionally, of these algorithms to be far less generalizable. One of the
MACHINE LEARNING IN CARDIAC TRANSPLANTATION    | 1751

key issues, when AI is applied to heart transplantation, and ML could aid a patient's journey within heart trans-
remains the lack of detailed data, be it structured or un- plantation by predicting the potential benefits of trans-
structured. Since the number of transplants performed plantation by analyzing initial lab investigations and
worldwide is relatively low, even the busiest single hos- imaging, ascertaining graft failure and mortality after
pitals will only have hundreds of cases in their registry. transplantation, and assisting a patient with medication
In order to take full advantage of this technology, there adherence and creating positive behavioral changes to
remains the need for access to nationwide registries which minimize further cardiovascular risk. Nevertheless, this
collect data and variables in a granular fashion. study also identified the need for higher quality, more
granular, and extensive databases since the models are
only as good as the initial information that is fed into
4.6 | Limitations them. Crucially, the heterogeneity in data restricted the
use of such models to adults over the age of sixty. More
This systematic review comes with certain limitations. multi-­center prospective and nationwide datasets are
Much of the data included in this systematic review were required to address these concerns whereby parameters
from retrospective observational studies, which is con- involved in heart transplantation are collected, regard-
ducive to bias and confounding. Additionally, due to the less of the traditionally perceived importance.
different databases utilized by each individual study, a
meta-­analysis is unachievable due to the heterogeneity in ACKNOWLEDGMENT
the variables included, and due to the type of ML models Imperial College London paid for the open access through
that were utilized. To test the full potential of ML models JISC. Open access funding enabled and organized by
and AI, larger multi-­center prospective studies are needed. ProjektDEAL.
Further studies will need to consider a broader range of
variables, especially those which are commonly not in- AUTHOR CONTRIBUTIONS
cluded due to the perceived lack of importance—­for exam- Arian Arjomandi Rad, Hariharan Subbiah Ponniah, and
ple, the immunosuppression regime post-­transplantation. Vinci Naruka: Concept/design, data collection, data inter-
Due to the timeliness of the following review and in pretation, drafting article, critical revision of article, ap-
view of the recent rapid advances in the field, we started proval of article. Jeevan Francis: data curation, drafting
the following work aiming to be able to rapidly provide article, approval of article. Robert Vardanyan, Panagiotis
the readers with a high-­quality review of this important Tasoudis, Dimitrios E. Magouliotis, George L. Lazopoulos,
topic. Therefore, we did not initially register the protocol Mohammad Yousuf Salmasi, and Thanos Athanasiou:
of our work on Prospero. Although we had taken steps Concept/design, data interpretation, critical revision of
before commencing this review to scan the literature for article, approval of article.
any ongoing or existing reviews on this topic, not finding
any similar work being present, we understand the im- CONFLICT OF INTEREST
portance of registering protocols of systematic reviews on The authors report no relationships that could be con-
PROSPERO to avoid duplication and overlapping works, strued as a conflict of interest.
and the following remains a limitation of this review.
DATA AVAILABILITY STATEMENT
The data that support the findings of this study are avail-
5 | CO N C LUSION able from the corresponding author, upon reasonable
request.
The implementation of machine learning models in
heart transplantation has illustrated the scope for this ORCID
powerful tool which could greatly enhance current clini- Arian Arjomandi Rad https://orcid.
cal practice by improving the predictability of outcomes. org/0000-0002-4931-4049
Several studies demonstrated the use of machine learn- Hariharan Subbiah Ponniah https://orcid.
ing was superior to traditional models of scoring systems org/0000-0002-5471-9120
and regression models in predicting endpoints of heart
transplantation, thus proving vital to improving the TWITTER
chances of successful transplantation and the chances Vinci Naruka @VinciNaruka
of a successful donor-­recipient match. The use of ML Arian Arjomandi Rad @AArjomandiRad
algorithms is not restricted to predicting mortality and Hariharan Subbiah Ponniah @Hari_SubPon7
graft failure. In combination with other applications, AI Panagiotis Tasoudis @TasoudisPanos
1752 |    MACHINE LEARNING IN CARDIAC TRANSPLANTATION

16. Yoon J, Zame WR, Banerjee A, Cadeiras M, Alaa AM, van der
REFERENCES
Schaar M. Personalized survival predictions via trees of pre-
1. Saidi RF, Kenari SKH, Saidi RF. Challenges of organ shortage dictors: an application to cardiac transplantation. PLoS One.
for transplantation: solutions and opportunities. Int J Organ 2018;13(3):e0194985.
Transplant Med. 2014;5(3):87–­96. 17. Agasthi P, Buras MR, Smith SD, Golafshar MA, Mookadam F,
2. Nilsson J, Ohlsson M, Höglund P, Ekmehag B, Koul B, Anand S, et al. Machine learning helps predict long-­term mor-
Andersson B. The international heart transplant survival al- tality and graft failure in patients undergoing heart transplant.
gorithm (IHTSA): a new model to improve organ sharing and Gen Thorac Cardiovasc Surg. 2020;68(12):1369–­76.
survival. PLoS One. 2015;10(3):e0118644. 18. Kampaktsis PN, Tzani A, Doulamis IP, Moustakidis S, Drosou
3. Guijo-­ Rubio D, Gutiérrez PA, Hervás-­ Martínez C. Machine A, Diakos N, et al. State-­of-­the-­art machine learning algorithms
learning methods in organ transplantation. Curr Opin Organ for the prediction of outcomes after contemporary heart trans-
Transplant. 2020;25(4):399–­405. plantation: results from the UNOS database. Clin Transplant.
4. Vinoth A, Datta S. Fundamentals of machine learning. 2022. p. 2021;35(8):e14388.
1–­27. 19. Hsich EM, Thuita L, McNamara DM, Rogers JG, Valapour M,
5. Yoo KD, Noh J, Lee H, Kim DK, Lim CS, Kim YH, et al. A ma- Goldberg LR, et al. Variables of importance in the scientific reg-
chine learning approach using survival statistics to predict graft istry of transplant recipients database predictive of heart trans-
survival in kidney transplant recipients: a multicenter cohort plant waitlist mortality. Am J Transplant. 2019;19(7):2067–­76.
study. Sci Reports. 2017;7(1):1–­12. 20. Ahady Dolatsara H, Chen YJ, Evans C, Gupta A, Megahed FM.
6. Lau L, Kankanige Y, Rubinstein B, Jones R, Christophi C, A two-­stage machine learning framework to predict heart trans-
Muralidharan V, et al. Machine-­ learning algorithms pre- plantation survival probabilities over time with a monotonic
dict graft failure after liver transplantation. Transplantation. probability constraint. Decis Support Syst. 2020;137:113363.
2017;101(4):e125–­32. 21. Tong L, Hoffman R, Deshpande SR, Wang MD. Predicting heart
7. Squiers JJ, Saracino G, Chamogeorgakis T, MacHannaford rejection using histopathological whole-­slide imaging and deep
JC, Rafael AE, Gonzalez-­Stawinski GV, et al. Application of neural network with dropout. 2017 IEEE EMBS International
the International Society for Heart and Lung Transplantation Conference on Biomedical & Health Informatics (BHI); 2017.
(ISHLT) criteria for primary graft dysfunction after cardiac p. 1–­4.
transplantation: outcomes from a high-­volume Centre. Eur J 22. Rajula HSR, Verlato G, Manchia M, Antonucci N, Fanos V.
Cardio-­Thoracic Surg. 2017;51(2):263–­70. Comparison of conventional statistical methods with machine
8. Foroutan F, Alba AC, Stein M, Krakovsky J, Chien KGW, Chih S, learning in medicine: diagnosis, drug development, and treat-
et al. Validation of the International Society for Heart and Lung ment. Medicina. 2020;56(9):455.
Transplantation primary graft dysfunction instrument in heart 23. Weintraub WS, Jones EL, Craver J, Guyton R, Cohen C.
transplantation. J Hear Lung Transplant. 2019;38(3):260–­6. Determinants of prolonged length of hospital stay after coro-
9. Miller PE, Pawar S, Vaccaro B, McCullough M, Rao P, Ghosh R, nary bypass surgery. Circulation. 1989;80(2):276–­84.
et al. Predictive abilities of machine learning techniques may 24. Al-­Attar N, Johnston S, Jamous N, Mistry S, Ghosh E, Gangoli
be limited by dataset characteristics: insights from the UNOS G, et al. Impact of bleeding complications on length of stay and
database. J Card Fail. 2019;25(6):479–­83. critical care utilization in cardiac surgery patients in England. J
10. Medved D, Nugues P, Nilsson J. Simulating the outcome Cardiothorac Surg. 2019;14(1):1–­10.
of heart allocation policies using deep neural networks. 25. Baek H, Cho M, Kim S, Hwang H, Song M, Yoo S. Analysis of
Conference Proceedings: Annual International Conference length of hospital stay using electronic health records: a statisti-
of the IEEE Engineering in Medicine and Biology Society. cal and data mining approach. PLoS One. 2018;13(4):e0195901.
IEEE Engineering in Medicine and Biology Society. Annual 26. De Mattos AM, Olyaei AJ, Bennett WM. Nephrotoxicity of
Conference; 2018. p. 6141–­4. immunosuppressive drugs: long-­term consequences and chal-
11. Medved D, Nugues P, Nilsson J. Predicting the outcome for lenges for the future. Am J Kidney Dis. 2000;35(2):333–­46.
patients in a heart transplantation queue using deep learning. 27. Kim IC, Youn JC, Kobashigawa JA. The past, present and future
Proceedings of the Annual International Conference of the of heart transplantation. Korean Circ J. 2018;48(7):565–­90.
IEEE Engineering in Medicine and Biology Society (EMBC); 28. Mulder BJM. Epidemiology of adult congenital heart dis-
2017. p. 74–­7. ease: demographic variations worldwide. Netherlands Hear J.
12. Ayers B, Sandholm T, Gosev I, Prasad S, Kilic A. Using machine 2012;20(12):505–­8.
learning to improve survival prediction after heart transplanta- 29. Chouairi F, Milner A, Sen S, Guha A, Stewart J, Jastreboff AM,
tion. J Card Surg. 2021;36(11):4113–­20. et al. Impact of obesity on heart transplantation outcomes. J
13. Medved D, Ohlsson M, Höglund P, Andersson B, Nugues Am Heart Assoc. 2021;10(23):21346.
P, Nilsson J. Improving prediction of heart transplanta- 30. Chokshi A, Cheema FH, Schaefle KJ, Jiang J, Collado E,
tion outcome using deep learning techniques. Sci Reports. Shahzad K, et al. Hepatic dysfunction and survival after or-
2018;8(1):1–­9. thotopic heart transplantation: application of the MELD scor-
14. Miller R, Tumin D, Cooper J, Hayes D, Tobias JD. Prediction ing system for outcome prediction. J Hear Lung Transplant.
of mortality following pediatric heart transplant using machine 2012;31(6):591–­600.
learning algorithms. Pediatr Transplant. 2019;23(3):e13360. 31. Taiwo AA, Khush KK, Stedman MR, Zheng Y, Tan JC.
15. Zhou Y, Chen S, Rao Z, Yang D, Liu X, Dong N, et al. Prediction Longitudinal changes in kidney function following heart
of 1-­year mortality after heart transplantation using machine transplantation: Stanford experience. Clin Transplant.
learning approaches: a single-­center study from China. Int J 2018;32(11):e13414.
Cardiol. 2021;339:21–­7.
MACHINE LEARNING IN CARDIAC TRANSPLANTATION    | 1753

32. Cunningham KS, Veinot JP, Butany J. An approach to endomyo-

cardial biopsy interpretation. J Clin Pathol. 2006;59(2):121–­9. How to cite this article: Naruka V, Arjomandi
33. Labovitz DL, Shafner L, Reyes Gil M, Virmani D, Hanina Rad A, Subbiah Ponniah H, Francis J, Vardanyan
A. Using artificial intelligence to reduce the risk of non-
R, Tasoudis P, et al. Machine learning and artificial
adherence in patients on anticoagulation therapy. Stroke.
2017;48(5):1416–­9.
intelligence in cardiac transplantation: A systematic
34. Paullada A, Raji ID, Bender EM, Denton E, Hanna A. Data and review. Artif. Organs. 2022;46:1741–­1753. https://
its (dis)contents: a survey of dataset development and use in doi.org/10.1111/aor.14334
machine learning research. Patterns. 2021;2(11):100336.
35. Shipe ME, Deppen SA, Farjah F, Grogan EL. Developing predic-
tion models for clinical use using logistic regression: an over-
view. J Thorac Dis. 2019;11(Suppl 4):S574–­84.