Accepted Article

DR. ASHLEY LADANA SPANN (Orcid ID : 0000-0003-1278-7712)

MR. MARC ANGELI (Orcid ID : 0000-0002-6809-8820)

Article type : Review

Applying Machine Learning in Liver Disease & Transplantation: A Comprehensive Review

Ashley Spann1, Angeline Yasodhara2, Justin Kang3, Kymberly Watt4, Bo Wang2, Anna Goldenberg2,
Mamatha Bhat3,5

1Division of Gastroenterology, Department of Medicine, Vanderbilt University Medical Center,

Nashville, Tennessee
2Vector Institute for Artificial Intelligence, Toronto, Canada
3Multi Organ Transplant Program, Toronto General Hospital Research Institute, University Health
Network, Toronto, Canada
4Division of Gastroenterology, Mayo Clinic, Rochester, Minnesota
5Division of Gastroenterology, Department of Medicine, University of Toronto, Toronto, Canada

Abstract
Machine learning utilizes artificial intelligence to generate predictive models efficiently and
more effectively than conventional methods through detection of hidden patterns within large data
sets. With this in mind, there are several areas within hepatology where these methods can be applied.
In this review, we examine the literature pertaining to machine learning in hepatology and liver
transplant medicine. We provide an overview of the strengths and limitations of machine learning
tools, and their potential applications to both clinical and molecular data in hepatology. Machine

This article has been accepted for publication and undergone full peer review but has not been
through the copyediting, typesetting, pagination and proofreading process, which may lead to
differences between this version and the Version of Record. Please cite this article as doi:
10.1002/hep.31103
This article is protected by copyright. All rights reserved
 learning has been applied to various types of data in Liver disease research, including clinical,
Accepted Article
demographic, molecular, radiologic and pathologic data. We anticipate that the use of ML tools to
generate predictive algorithms will change the face of clinical practice in Hepatology and
transplantation. This review will provide readers with the opportunity to learn about the ML tools
available, and potential applications to questions of interest in Hepatology.

Introduction
In the changing atmosphere of healthcare and information technology, there is an increasing
opportunity for the use of data science and technology to personalize healthcare and improve delivery
of patient care. At its core, machine learning utilizes artificial intelligence to generate predictive
models efficiently and more effectively than conventional methods through detection of hidden
patterns within large data sets. With this in mind, there are several areas within hepatology where
these methods can be applied. In this review, we examine the literature of the already tested
applications of machine learning in hepatology and liver transplant medicine. We provide an
overview of the strengths and limitations of machine learning tools, and their potential applications to
both clinical and molecular data in hepatology.

Artificial intelligence (AI) and machine learning (ML) algorithms have been increasingly
applied to questions in hepatology in recent years. Electronic health records are a rich source of data,
as are registries and clinically-annotated biobanks. Efforts such as The Cancer Genome Atlas continue
to produce layers of molecular data. The large proportion of the research literature in hepatology
stems from the use of traditional biostatistical methods. These hypothesis-driven studies consist of
examination of preselected variables and their impact on liver-related outcomes such as cirrhosis,
liver cancer, transplantation and mortality. These studies have included prediction models that have
revolutionized clinical practice in hepatology1. Machine learning is an unbiased approach that stands
in complete contrast to this, using any number of variables to permit data-driven discovery. This
hypothesis-free approach has led to identification of similarities and differences in clinical

This article is protected by copyright. All rights reserved

 phenotypes, systematization of patient diagnosis, elucidation of new therapeutic targets, insights into
Accepted Article
the mechanistic basis of disease, and delivery of a data-driven, precision medicine approach2.

Liver diseases are complex and heterogeneous in nature, developing under the influence of
various factors that affect susceptibility to disease. These include sex, ethnicity, genetics,
environmental exposures (viruses, alcohol, diet, chemical), body mass index, and comorbid conditions
such as diabetes. Various types of complex data are generated in hepatology practice and research that
could benefit from AI-based approaches: electronic health record data, transient elastography, other
imaging technologies, histology, biobank data, data from clinical trials, clinical sensors, wearables,
and a variety of molecular data (genomics, transcriptomics, proteomics, metabolomics, immunomics,
microbiomics).

Advantages of Machine Learning

Prediction of a given outcome is often difficult for a clinician to ascertain, given the number of
variables that can affect clinical outcomes. Standard clinical risk assessment models implicitly assume
that each risk factor is related in a linear fashion to clinical outcomes3. Using Machine Learning
Algorithms (MLAs), there is an ability to better incorporate multiple risk factors and identify more
nuanced relationships between risk factors and outcomes. The key advantage of the MLA method
over the traditional predictive model is that MLAs learn from existing data to find novel patterns
between variables and generate predictions4. This is especially powerful when there are many
interactions between the predictors and when data is high-dimensional, which is certainly reflected in
the complexity of liver diseases. MLAs have been shown to improve precision in identifying
individuals at risk of disease5-10.

Machine Learning Tools and Applications to Liver disease (Figure 1)

In supervised learning, tools learn to output the correct labeled target, which can vary from
detection of underlying liver disease in patients, early detection of non-alcoholic fatty liver disease
(NAFLD) with images, or better identification of patients with primary sclerosing cholangitis (PSC)
at risk for hepatic decompensation or otherwise11-14,52. Machine learning in the supervised setting

This article is protected by copyright. All rights reserved

 encompasses tools that can uncover non-linear patterns in the data to predict these various output
Accepted Article
targets. A simple extension of Bayes’ theorem, Naive Bayes classifiers predict class labels by
computing the likelihood of the observed features under each class and returning the class with the
maximum likelihood. k-nearest neighbors (KNN), on the other hand, determines the output based on
the value of classes of the K-nearest training samples. Another example of ML classifier is support
vector machine (SVM) which finds the optimal divisor among classes in the kernel-transformed
hyperplane of the data15. KNN and SVM have been used by Kim et al. to identify a molecular
signature for hepatocellular carcinoma16. Simple models like a decision tree can also be used17. A
decision tree is similar to a flowchart arranged in a tree-like structure, where each step of the
flowchart denotes a test on one or more features and by following the flowchart, one can classify each
sample. Predictions from multiple unique decision trees can be used together in an ensemble. These
ensembles are called random forests (RF) and gradient boosting machines (GBM) 18-19. This has been
used to identify PSC patients with higher risk of hepatic decompensation52. Random forests use an
ensemble of deep decision trees that are trained on different random subsets of the training data in
parallel. The final output of the method corresponds to the mode of all the decision trees’ results.
GBM, on the other hand, uses shallow trees with only one or two levels. These shallow trees are
considered to make predictions that are high in bias and low in variance, as opposed to a full-grown
tree used in random forests that are low in bias and high in variance.

Deep neural networks (DNN) have been a tremendous breakthrough in ML, enabling
machines to learn patterns of data by modeling them through a combination of simple non-linear
elementary operations. Neural networks have been applied to predict 3-month graft survival and to
assist with donor-recipient matching for patients with end-stage liver disease as well as predicting the
presence of liver disease from imaging20,21. This can be further extended into convolutional neural
networks (CNN) and recurrent neural networks (RNN) which handle local structures and sequential
data consecutively22,23. Local structure can be important in data, e.g. in images, and it is important to
incorporate this existing structure. CNN uses multiple convolution filters, learned by the network, at
different layers to aggregate information from neighboring pixels. RNN allows temporal
dependability across different timepoints by modifying the architecture to receive input from its past

This article is protected by copyright. All rights reserved

 state. The power of neural networks (NN) can be further applied into survival analysis and time-to-
Accepted Article
event predictions, where NN can be used to predict the risk function or even the parameters of the
distribution, modeling likelihood of the event24-26.

The application of machine learning extends beyond the setting of supervised learning.
Unsupervised learning algorithms have been widely used to automatically discover the patterns
without any labeled data. Classic unsupervised learning methods range from clustering algorithms
such as k-means and graph-based spectral clustering, to dimensionality reduction methods such as
principal component analysis or kernel-based methods27-29,. Deep neural networks generalize some of
these approaches by learning the dataset distribution, whether explicitly or implicitly, and generating
samples from those learned distribution. For example, variational Auto-Encoder (VAE) parameterizes
the distribution of the dataset and trains the neural network to learn the distribution that fits the
training dataset best by maximizing its likelihood30. Generative Adversarial Model (GAN) uses two
separate networks, one to generate fake samples (generator) and another to discriminate whether the
given input is fake or real (discriminator)31. These networks learn adversarially: the goal of one is to
generate samples that are closer to the true distribution, while the other wants to better differentiate
the generated and the true training samples. This method of training results in a model able to
generate samples that are very similar to the training distribution. This method can also be further
extended to impute missing data32,33.

Methods
A comprehensive literature review was conducted by two independent reviewers (ALS and
JK). Two biomedical databases – MEDLINE (PubMed) and Embase (Elsevier) – were searched for
relevant studies through 01/15/2019. The primary search strategy was created in PubMed and
included a combination of text word and Medical Subject Heading (MeSH) terms. Primary search
concepts included machine learning, predictive modeling, deep learning, and liver transplantation as
well as specific etiologies for liver disease, such as Hepatitis C and non-alcoholic fatty liver disease
(NAFLD). The itemized search strategy was then translated to the additional database Embase. The

This article is protected by copyright. All rights reserved

 itemized search strategy can be found in the supporting information. Citations were managed using
Accepted Article
EndNote.
The queries retrieved 487 citations with 305 duplicate citations identified using EndNote,
leaving 182 citations for review. These citations were reviewed manually by two authors (ALS and
JK) for relevance to machine learning, chronic liver diseases and liver transplantation. This process
resulted in selection of 40 papers that included primary data relevant to the topics. The flowchart of
this process is illustrated in Figure 2. The relevant papers are listed in Table 1. Further details
regarding how papers were selected for inclusion or excluded are provided in the supporting
information.

Machine Learning in Liver Diseases

Machine learning has been applied to the identification, prediction, and assessment of multiple
liver diseases. In a study of 126 patients (56 healthy patients with F0, 70 patients with F1 and above),
a support vector machine classification algorithm was employed to analyze the still images obtained
from shear wave elastography (SWE) and categorize patients by the presence or absence of
underlying liver disease. Through an inverse mapping sequence correlating quantifiable color
information with obtained stiffness values, this algorithm was able to accurately discriminate healthy
subjects from those with chronic liver disease at 87.3%, with a sensitivity of 93.5% and specificity of
81.2% when compared to gold standard liver biopsy11. Unfortunately, this study excluded patients
with obesity, limiting generalizability of the algorithm, and etiology of underlying liver disease was
not taken into account. The disadvantages of these studies include the need for pathologist-directed
annotations for training of the machine learning algorithms as well as image preparation and
processing.

Traditional serum-based biomarker indices such as APRI, Fibrosis-4 (FIB-4), AST to ALT
ratio, and NAFLD fibrosis score have been used in clinical practice to identify significant fibrosis and
cirrhosis in patients already diagnosed with different etiologies of chronic liver disease, with modest

This article is protected by copyright. All rights reserved

 accuracy and a large percentage of indeterminate scores. These tests have not been amenable to
Accepted Article
routine use in screening patients for cirrhosis: although their specificity may be adequate, their
sensitivity is insufficient. ML models have been employed to detect fibrosis or cirrhosis secondary to
different liver disease etiologies based on blood tests34-41. These studies have predicted the risk of
cirrhosis in order to reduce the overall number of liver biopsies done on patients with known risk
factors. Diagnostic performance of these models has varied depending on the etiology of liver disease
and study size. The principal limitation of these studies has been the classification of cirrhotic patients
with varying degrees of decompensation into the Stage 4 fibrosis category. Nonetheless, these ML
models have shown promise by consistently outperforming traditional serum-based biomarker
indices.

NAFLD
Perhaps the best-studied condition in which machine learning has been explored for pattern
recognition is non-alcoholic fatty liver disease (NAFLD). Early detection and identification of
patients with NAFLD at risk of disease progression is paramount. Towards this end, efforts have been
made to utilize machine learning methods to not only more accurately and efficiently identify these
patients through image analyses and pathology review, but also to help differentiate severity of the
underlying steatosis.

Machine Learning and Imaging Analyses in NAFLD

Machine learning computational methods have been used for pattern recognition in order to
accurately and efficiently identify patients with liver disease. In an attempt to automate the staging of
disease Vanderbeck et al generated a classification algorithm utilizing 47 unique liver biopsy images
with manual annotations by two pathologists12. Through utilization of a color analysis protocol, the
algorithm was able to identify key liver biopsy features (macrosteatosis, bile ducts, portal veins and
sinusoids) with high precision and recall >82%12. Since not all patients undergo liver biopsy, similar
measures have been applied to non-invasive methods of analysis such as shear wave elastography
(SWE) imaging. Through the utilization of neural networks applied to ultrasound images, Kuppuli et
al were able to differentiate abnormal ultrasounds in fatty liver disease from normal images with an

This article is protected by copyright. All rights reserved

 accuracy of 96.75%13. Similarly, in a study of 55 severely obese patients that combined the use of
Accepted Article
neural networks for feature extraction and support vector machine algorithms for classification, Byra
et al were able to automatically detect the amount of steatosis present in the liver (AUC 0.98) from
ultrasound image sequences when compared to gold standard liver biopsy without the need for
preceding image processing or eliminating inter-operator variability14.

Predictive Modeling in NAFLD

Through analyses of EMR data, machine learning allows for assessment of large volumes of
clinical data for the detection of underlying NAFLD42-44. To identify those patients with the greatest
likelihood of having underlying hepatic steatosis, gradient boosting techniques were applied to 254
clinical variables obtained from a population of 4,222 individuals to develop a predictive model. After
further isolation of variables, the resultant scoring system was able to discriminate well between
patients with and without hepatic steatosis in the training data (AUC 0.860) and the validation data
(AUC 0.876)43. In a comparison study of different MLAs, Ma et al found that all MLAs performed
better than the Fatty Liver Index (FLI; F-measure 0.318) and the Hepatis Steatosis Index (HSI; F-
measure 0.524), with the Bayesian network model performing the best of 11 MLAs in the
classification of patients with NAFLD from a set of 2,522 cases of Chinese patients (F-measure
0.655)44. In an effort to objectively classify patients with steatosis due to NAFLD as opposed to
alcohol liver disease, serological biomarkers were obtained from patients with either known NAFLD
or ALD along with other clinical data including elastography measurements45. Different machine
learning approaches such as logistic regression, decision trees, SVM and random forests were
evaluated. The best MLAs were able to accurately distinguish NAFLD from both non-cirrhotic ALD
(SVM, AUC 0.91 ± 0.0056) and cirrhotic ALD (RF; AUC 0.98 ± 0.0018)45. Each of these examples
support that given a clinical context, MLAs may perform differently than in other scenarios and
therefore it is necessary to evaluate multiple MLAs for a specific problem to identify the most
appropriate algorithm to apply.

Viral Hepatitis

This article is protected by copyright. All rights reserved

 Not all patients with hepatitis C infection will progress to cirrhosis, therefore it remains
Accepted Article
potentially cost-effective to identify the subset of patients at risk for developing cirrhosis. Thus, after
performing whole blood genotyping of 420 Caucasian patients without cirrhosis as determined by
liver biopsy, Huang et al utilized machine learning methods to derive a predictive risk score (Cirrhosis
Risk Score – CRS) for HCV cirrhosis over time46. The gene signature was developed from a
combination of 7 SNPs after utilizing a heuristic, repetitive iteration ranking system to identify which
combination of genes was best able to consistently and robustly improve detection. This scoring
system was applied to a multi-center cohort of 154 patients compared to conventional risk
stratification methods based on clinical factors alone and was superior in differentiating high versus
low risk for cirrhosis development. A similar machine learning method was later applied to genetic
analyses of different HCV strains and was subsequently able to identify relevant genetic markers
associated with fast and slow rates of fibrosis progression in HCV47. With these same principles,
Shousha et al were able to combine data mining strategies and IL28B genotyping to predict advanced
fibrosis in HCV patients using a neural network algorithm that had higher performance than both
APRI and FIB-448. These studies highlight the opportunity for genetic assay design through the
utilization of computational modeling.

In the absence of genetic data, machine learning has also played a role in clinical data-based
predictive assessments. Wei et al compared machine learning methods with the FIB-4 score for
detection of viral hepatitis-related cirrhosis34. Gradient boosting outperformed other MLAs, with an
AUROC of 0.87 as compared to the AUROC for FIB-4 at 0.8334. In a cohort of chronic hepatitis C
patients from the HALT-C trial, Konerman et al utilized longitudinal clinical data to improve upon
standard statistical methods (AUC 0.79) with random forest (AUC 0.86) and boosting MLAs (AUC
0.84) for prediction of fibrosis progression, liver-related death, hepatic decompensation, an increase in
Child-Turcotte Pugh score ≥ 7, hepatocellular carcinoma or liver transplantation within 1 year49. The
success of these algorithms was later validated in a more heterogenous cohort of 1,007 chronic
hepatitis C patients for the prediction of both 1-year (AUROC 0.78) and 3-year (AUROC 0.76)
outcomes50. Unfortunately, the algorithm was not able to be validated for prediction of fibrosis
progression due to small sample size. In a follow-up study utilizing a cohort of 72,683 veterans with

This article is protected by copyright. All rights reserved

 chronic hepatitis C, boosted survival tree-based models utilizing longitudinal data consistently
Accepted Article
outperformed prediction of development of cirrhosis in chronic hepatitis C patients at 1, 3 and 5 years
as opposed to standard cross-sectional statistical methods51.

Primary Sclerosing Cholangitis

ML methods have been investigated in patients with primary sclerosing cholangitis (PSC) both
preceding and following transplantation. In an effort to be able to better identify patients with PSC at
risk for hepatic decompensation, Eaton et al. examined a large cohort of 509 PSC patients to derive an
algorithm predictive of hepatic decompensation (ascites, variceal hemorrhage, and encephalopathy)
utilizing the gradient boosting machines (GBM) algorithm. This algorithm was validated in a smaller
cohort of patients with accurate prediction of hepatic decompensation (C-statistic, 0.90; 95% CI 0.84-
0.95)52. In sensitivity analyses, the algorithm remained accurate regardless of baseline phenotype of
PSC, degree of bilirubin elevation or hepatic involvement. In an attempt to identify those patients
with PSC most likely to survive post liver transplantation, Andres et al. produced a novel multi-time
point calibrated model for the prediction of individual survival after liver transplantation derived from
SRTR database information for 2,769 PSC LT recipients53. This model was able to outperform
traditional Cox regression analysis in the evaluation of long-term survival for PSC LT recipients.

Hepatocellular Carcinoma
In a study of 59 tissue samples obtained from explanted livers of liver transplant recipients,
Kim et al. generated cDNA microarrays with over 9,000 genes per sample16. Through utilization of k-
nearest neighbors and support vector machine methods, they were able to identify a molecular
signature of 30 genes significantly altered in cirrhotic patients at high risk for HCC16. Unfortunately,
molecular data is not readily available for the evaluation of patient risk, therefore other methods need
to be employed to be able to identify those at greatest risk for development of HCC. In a prospective
evaluation of 442 patients with Child A or B cirrhosis, Singal et al. sought to develop and compare
predictive models utilizing regularly collected clinical data and conventional predictive methods in
comparison to ML algorithms54. A random forest approach for decision tree-based analysis
performed significantly better (C-statistic 0.71 [CI 0.63-0.79]) than conventional regression modeling

This article is protected by copyright. All rights reserved

 (C-statistic 0.64, 95% CI 0.54-0.73). These algorithms were then applied to a validation cohort from
Accepted Article
the HALT-C cohort (1050 HCV patients), and once again the ML algorithm was better able to
identify those patients who would go on to develop HCC with sensitivity of 80.7% with specificity of
46.8%, compared to the original model generated from the HALT-C cohort (sensitivity 70.7%,
specificity 41.6%). Despite the ability to outperform conventional methods, there are some limitations
that remain in this study. Particularly, this ML algorithm was derived from a cohort of patients at a
single tertiary care center and therefore may not be fully representative and generalizable. In addition,
the algorithm was limited to input of clinical data at enrollment and did not utilize longitudinal data
for algorithm refinement. Despite these limitations, this study was able to highlight the greater
capabilities and stability of ML in high-risk patient identification and disease development across
multiple cohorts.

MACHINE LEARNING IN LIVER TRANSPLANTATION

Screening and Selection of LT Recipients
It is necessary to not only accurately identify those patients who are most likely to benefit
from liver transplantation, but also optimize graft and patient survival. In efforts to address this, ML
has been employed to improve equity of donor-recipient matching by employing rule-based systems
using artificial neural networks (ANN)55. Utilizing a compilation of clinical data from 1003 liver
transplant recipients and their donors, as well as data from the retrieval and pre-transplant process,
ANNs were utilized to generate separate algorithms for prediction of 3-month graft survival20,56.
When compared to several current scoring systems that utilize either isolated donor/recipient scores or
combined donor/recipient factors, the ANN more accurately predicted 3-month graft survival (AUC
0.81) and graft loss (AUC 0.82). A separate study combined donor, recipient and transplant-related
variables in predicting 30-day risk of graft failure57. Random forest applied on 15 donor and recipient
characteristics at transplant resulted in an excellent AUC of 0.818 (95% CI 0.812-0.824) in predicting
risk of graft failure, further confirming the utility of ML tools in donor-recipient matching57.

Further to this, a more recent study used a ML tool known as the Optimal Classification Tool
(OCT) to predict a given individual’s 3-month mortality on the waiting list or risk of delisting58. This

This article is protected by copyright. All rights reserved

 method is based on the principle that these hierarchically organized structures of nodes (classification
Accepted Article
trees) can make predictions by sequentially splitting the data. The authors accounted for both
traditional indications for transplant with biological MELD score as well as exception point
indications such as HCC. Using the OCT would result in a significant decrease in 418 deaths on the
waiting list, by allocating organs based on the expected 3-month mortality on the waiting list. This
study demonstrates the power of ML tools in effecting change that could render organ allocation more
equitable.

Prediction of Post-Transplant Survival & Complications

ML has also been used to predict outcomes such as acute kidney injury (AKI) and diabetes
after transplant. ML tools were employed on preoperative and intraoperative anesthesia and surgery-
related variables to predict post-operative AKI, which has been associated with increased mortality59.
Gradient boosting machine performed best among all ML methods to predict AKI of all stages (AUC
0.90, 95% CI 0.86–0.93), as compared to AUC for standard logistic regression analysis of 0.61 (95%
CI 0.56–0.66). Post-transplant diabetes mellitus (PTDM) is a major complication associated with a 2-
fold higher risk of cardiovascular events, graft loss and infections in the long term60. Different MLAs
were used to identify key predictors of PTDM in the U.S. Scientific Registry for Transplant
Recipients (SRTR)61. Increasing age, male sex, and obesity were recipient factors correlating with
increased risk of PTDM. Sirolimus as primary immunosuppressant carried a 33% higher risk of
PTDM than tacrolimus61.

Pitfalls of Machine Learning

Though Machine learning approaches have been shown to improve upon standard biostatistics
in selected medical conditions, one should keep in mind that the tools work off of data. Therefore, the
generated algorithms are only as good as the quality of data provided. Various factors can affect the
performance an accuracy of a ML model. Careful study design, data integration strategy, choice of
ML model, and relevance of this model to the clinical setting are all critical factors. Certain types of
ML models may be favored for specific types of data: CNN is usually used for images, RNN for

This article is protected by copyright. All rights reserved

 sequential data, and survival analysis for time-to-death. For other types of general classification, all
Accepted Article
different methods will generally be evaluated to see what works best. Additionally, one should be
careful not to overfit a model to the training data, which is a common error in ML.

Even though ML is often perceived as a black box model, a lot of work has been done to make
ML models interpretable and there are a few tools available to build a more interpretable yet powerful
model62-63. It is important to look carefully at what the model is learning. For example, some
questions one might ask can be: (1) What factors does the model find to be important? Are they
confounders? (2) Which subpopulation of patients is the model consistently misclassifying? (3) Is
there a data shift distribution between the training data and where the model would be applied?

Conclusion
Machine Learning approaches may potentially improve upon biostatistical methods to address
questions across medicine. Though standard biostatistics can be sufficient for many questions in
hepatology, ML can improve upon this, particularly when studying questions of clinical prediction.
ML approaches have been increasingly used in recent years in hepatology to examine the wealth of
clinical, molecular, radiologic and pathologic data available in liver disease. One can anticipate that
using these tools to decipher the complexity of liver disease could enhance the identification of more
optimal biomarkers and therapeutic strategies, and ultimately a more precision medicine approach to
the practice of hepatology.

Acknowledgements: The authors acknowledge the assistance of Marc Angeli and Elisa Pasini in
formatting and the submission process of this manuscript.

This article is protected by copyright. All rights reserved

Accepted Article
References
1. Kamath PS, Wiesner RH, Malinchoc M, et al. A model to predict survival in patients with end-stage
liver disease. Hepatology. 2001;33(2):464-470.
2. Chen JH, Asch SM. Machine Learning and Prediction in Medicine - Beyond the Peak of Inflated
Expectations. N Engl J Med. 2017;376(26):2507-2509.
3. Obermeyer Z, Emanuel EJ. Predicting the Future - Big Data, Machine Learning, and Clinical Medicine.
N Engl J Med. 2016;375(13):1216-1219.
4. Berglund E, Lytsy P, Westerling R. Adherence to and beliefs in lipid-lowering medical treatments: a
structural equation modeling approach including the necessity-concern framework. Patient Educ
Couns. 2013;91(1):105-112.
5. Alghamdi M, Al-Mallah M, Keteyian S, Brawner C, Ehrman J, Sakr S. Predicting diabetes mellitus using
SMOTE and ensemble machine learning approach: The Henry Ford ExercIse Testing (FIT) project. PLoS
One. 2017;12(7):e0179805.
6. Kavakiotis I, Tsave O, Salifoglou A, Maglaveras N, Vlahavas I, Chouvarda I. Machine Learning and Data
Mining Methods in Diabetes Research. Comput Struct Biotechnol J. 2017;15:104-116.
7. Maeta K, Nishiyama Y, Fujibayashi K, et al. Prediction of Glucose Metabolism Disorder Risk Using a
Machine Learning Algorithm: Pilot Study. JMIR Diabetes. 2018;3(4):e10212.
8. Brisimi TS, Xu T, Wang T, Dai W, Paschalidis IC. Predicting diabetes-related hospitalizations based on
electronic health records. Stat Methods Med Res. 2018:962280218810911.
9. Zou Q, Qu K, Luo Y, Yin D, Ju Y, Tang H. Predicting Diabetes Mellitus With Machine Learning
Techniques. Front Genet. 2018;9:515.
10. Tu JF, Pan HY, Ying XH, Lou J, Ji JS, Zou H. Mast Cells Comprise the Major of Interleukin 17-Producing
Cells and Predict a Poor Prognosis in Hepatocellular Carcinoma. Medicine (Baltimore).
2016;95(13):e3220.
11. Gatos I, Tsantis S, Spiliopoulos S, et al. A Machine-Learning Algorithm Toward Color Analysis for
Chronic Liver Disease Classification, Employing Ultrasound Shear Wave Elastography. Ultrasound Med
Biol. 2017;43(9):1797-1810.

This article is protected by copyright. All rights reserved

 12. Vanderbeck S, Bockhorst J, Komorowski R, Kleiner DE, Gawrieh S. Automatic classification of white
Accepted Article regions in liver biopsies by supervised machine learning. Hum Pathol. 2014;45(4):785-792.
13. Kuppili V, Biswas M, Sreekumar A, et al. Author Correction to: Extreme Learning Machine Framework
for Risk Stratification of Fatty Liver Disease Using Ultrasound Tissue Characterization. J Med Syst.
2017;42(1):18.
14. Byra M, Styczynski G, Szmigielski C, et al. Transfer learning with deep convolutional neural network
for liver steatosis assessment in ultrasound images. Int J Comput Assist Radiol Surg.
2018;13(12):1895-1903.
15. Cortes CV, V. Support-vector networks. Machine Learning. 1995;20:273.
16. Kim JW, Ye Q, Forgues M, et al. Cancer-associated molecular signature in the tissue samples of
patients with cirrhosis. Hepatology. 2004;39(2):518-527.
17. Bertsimas D, Dunn J, Pawlowski C, et al. Applied Informatics Decision Support Tool for Mortality
Predictions in Patients With Cancer. JCO Clin Cancer Inform. 2018;2:1-11.
18. Breiman L. Random Forests. Machine Learning. 2001;45(5).
19. Friedman JH. Greedy Function Approximation: A Gradient Boosting Machine. The Annals of Statistics.
2001;29(5):1189-1232.
20. Briceno J, Cruz-Ramirez M, Prieto M, et al. Use of artificial intelligence as an innovative donor-
recipient matching model for liver transplantation: results from a multicenter Spanish study. J
Hepatol. 2014;61(5):1020-1028.
21. Liu X, Song JL, Wang SH, Zhao JW, Chen YQ. Learning to Diagnose Cirrhosis with Liver Capsule Guided
Ultrasound Image Classification. Sensors (Basel). 2017;17(1).
22. LeCun YB, L. ; Bengio, Y.; Haffner, P. Gradiant-based Learning Applied to Documents Recognition. proc
Of The IEEE. 1989(November).
23. Rumelheart DES, P.; McClelland, J.L; and Hinton, G. E. Schemata and Sequential Thought Processes in
PDP Models. Psycoloctsl Processes. 1982.
24. Katzman JL, Shaham U, Cloninger A, Bates J, Jiang T, Kluger Y. DeepSurv: personalized treatment
recommender system using a Cox proportional hazards deep neural network. BMC Med Res
Methodol. 2018;18(1):24.
25. Christ PF, Ettlinger F, Kaissis G, et al. SurvivalNet: Predicting patient survival from diffusion weighted
magnetic resonance images using cascaded fully convolutional and 3D Convolutional Neural

This article is protected by copyright. All rights reserved

 Networks. Paper presented at: 2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI
Accepted Article 2017); 18-21 April 2017, 2017.
26. martinsson E. WTTE-RNN: Weilbull Time To Event Recurrent Neural Network. Gothenburg, Sweeden:
Engineering Mathematics and Computational Science, University of Gothenburg; 2017.
27. von Luxburg U. A tutorial on spectral clustering. Statistics and Computing. 2007;17(4):395-416.
28. Wang B, Mezlini AM, Demir F, et al. Similarity network fusion for aggregating data types on a genomic
scale. Nat Methods. 2014;11(3):333-337.
29. Wang B, Zhu J, Pierson E, Ramazzotti D, Batzoglou S. Visualization and analysis of single-cell RNA-seq
data by kernel-based similarity learning. Nat Methods. 2017;14(4):414-416.
30. Auto-Encoding Variational Bayes [computer program]. Version 10. Cornell University: Cornell; 2013.
31. Generative Adversarial Networks [computer program]. Version 1. Cornell University: Cornell; 2014.
32. Yoon JJ, J.; van der schaar, M.;. GAIN: Missing Data Imputation using Generative Adversarial Nets.
Arxiv. 2018.
33. Lou YC, X.; Zhang, Y.; Xu, J.; Xiaojie, Y. Multivariate Time Series Imputation with Generative
Adversarial Networks. Paper presented at: Nerual Information Processing Systmes2018.
34. Wei R, Wang J, Wang X, et al. Clinical prediction of HBV and HCV related hepatic fibrosis using
machine learning. EBioMedicine. 2018;35:124-132.
35. Cao Y, He K, Cheng M, et al. Two classifiers based on serum peptide pattern for prediction of HBV-
induced liver cirrhosis using MALDI-TOF MS. Biomed Res Int. 2013;2013:814876.
36. Cao Y, Hu ZD, Liu XF, Deng AM, Hu CJ. An MLP classifier for prediction of HBV-induced liver cirrhosis
using routinely available clinical parameters. Dis Markers. 2013;35(6):653-660.
37. Wang D, Wang Q, Shan F, Liu B, Lu C. Identification of the risk for liver fibrosis on CHB patients using
an artificial neural network based on routine and serum markers. BMC Infect Dis. 2010;10:251.
38. Yip TC, Ma AJ, Wong VW, et al. Laboratory parameter-based machine learning model for excluding
non-alcoholic fatty liver disease (NAFLD) in the general population. Aliment Pharmacol Ther.
2017;46(4):447-456.
39. Stoean R, Stoean C, Lupsor M, Stefanescu H, Badea R. Evolutionary-driven support vector machines
for determining the degree of liver fibrosis in chronic hepatitis C. Artif Intell Med. 2011;51(1):53-65.
40. Islam MM, Wu CC, Poly TN, Yang HC, Li YJ. Applications of Machine Learning in Fatty Live Disease
Prediction. Stud Health Technol Inform. 2018;247:166-170.

This article is protected by copyright. All rights reserved

 41. Pournik O, Dorri S, Zabolinezhad H, Alavian SM, Eslami S. A diagnostic model for cirrhosis in patients
Accepted Article with non-alcoholic fatty liver disease: an artificial neural network approach. Med J Islam Repub Iran.
2014;28:116.
42. Perveen S, Shahbaz M, Keshavjee K, Guergachi A. A Systematic Machine Learning Based Approach for
the Diagnosis of Non-Alcoholic Fatty Liver Disease Risk and Progression. Sci Rep. 2018;8(1):2112.
43. Meffert PJ, Baumeister SE, Lerch MM, Mayerle J, Kratzer W, Volzke H. Development, external
validation, and comparative assessment of a new diagnostic score for hepatic steatosis. Am J
Gastroenterol. 2014;109(9):1404-1414.
44. Ma H, Xu CF, Shen Z, Yu CH, Li YM. Application of Machine Learning Techniques for Clinical Predictive
Modeling: A Cross-Sectional Study on Nonalcoholic Fatty Liver Disease in China. Biomed Res Int.
2018;2018:4304376.
45. Sowa JP, Atmaca O, Kahraman A, et al. Non-invasive separation of alcoholic and non-alcoholic liver
disease with predictive modeling. PLoS One. 2014;9(7):e101444.
46. Huang H, Shiffman ML, Friedman S, et al. A 7 gene signature identifies the risk of developing cirrhosis
in patients with chronic hepatitis C. Hepatology. 2007;46(2):297-306.
47. Lara J, Lopez-Labrador F, Gonzalez-Candelas F, Berenguer M, Khudyakov YE. Computational models of
liver fibrosis progression for hepatitis C virus chronic infection. BMC Bioinformatics. 2014;15 Suppl
8:S5.
48. Shousha HI, Awad AH, Omran DA, Elnegouly MM, Mabrouk M. Data Mining and Machine Learning
Algorithms Using IL28B Genotype and Biochemical Markers Best Predicted Advanced Liver Fibrosis in
Chronic Hepatitis C. Jpn J Infect Dis. 2018;71(1):51-57.
49. Konerman MA, Zhang Y, Zhu J, Higgins PD, Lok AS, Waljee AK. Improvement of predictive models of
risk of disease progression in chronic hepatitis C by incorporating longitudinal data. Hepatology.
2015;61(6):1832-1841.
50. Konerman MA, Lu D, Zhang Y, et al. Assessing risk of fibrosis progression and liver-related clinical
outcomes among patients with both early stage and advanced chronic hepatitis C. PLoS One.
2017;12(11):e0187344.
51. Konerman MA, Beste LA, Van T, et al. Machine learning models to predict disease progression among
veterans with hepatitis C virus. PLoS One. 2019;14(1):e0208141.

This article is protected by copyright. All rights reserved

 52. Eaton JE, Vesterhus M, McCauley BM, et al. Primary Sclerosing Cholangitis Risk Estimate Tool (PREsTo)
Accepted Article Predicts Outcomes of the Disease: A Derivation and Validation Study Using Machine Learning.
Hepatology. 2018.
53. Andres A, Montano-Loza A, Greiner R, et al. A novel learning algorithm to predict individual survival
after liver transplantation for primary sclerosing cholangitis. PLoS One. 2018;13(3):e0193523.
54. Singal AG, Mukherjee A, Elmunzer BJ, et al. Machine learning algorithms outperform conventional
regression models in predicting development of hepatocellular carcinoma. Am J Gastroenterol.
2013;108(11):1723-1730.
55. Cruz-Ramirez M, Hervas-Martinez C, Fernandez JC, Briceno J, de la Mata M. Predicting patient survival
after liver transplantation using evolutionary multi-objective artificial neural networks. Artif Intell
Med. 2013;58(1):37-49.
56. Ayllon MD, Ciria R, Cruz-Ramirez M, et al. Validation of artificial neural networks as a methodology for
donor-recipient matching for liver transplantation. Liver Transpl. 2018;24(2):192-203.
57. Lau L, Kankanige Y, Rubinstein B, et al. Machine-Learning Algorithms Predict Graft Failure After Liver
Transplantation. Transplantation. 2017;101(4):e125-e132.
58. Bertsimas D, Kung J, Trichakis N, Wang Y, Hirose R, Vagefi PA. Development and validation of an
optimized prediction of mortality for candidates awaiting liver transplantation. Am J Transplant.
2019;19(4):1109-1118.
59. Lee HC, Yoon SB, Yang SM, et al. Prediction of Acute Kidney Injury after Liver Transplantation:
Machine Learning Approaches vs. Logistic Regression Model. J Clin Med. 2018;7(11).
60. Davidson J, Wilkinson A, Dantal J, et al. New-onset diabetes after transplantation: 2003 International
consensus guidelines. Proceedings of an international expert panel meeting. Barcelona, Spain, 19
February 2003. Transplantation. 2003;75(10 Suppl):SS3-24.
61. Bhat V, Tazari M, Watt KD, Bhat M. New-Onset Diabetes and Preexisting Diabetes Are Associated
With Comparable Reduction in Long-Term Survival After Liver Transplant: A Machine Learning
Approach. Mayo Clin Proc. 2018;93(12):1794-1802.
62. caruana RL, Y.; Gehrke, J.; Koch, P.; Sturm, M.; Elhadad, N. Intelligible Models for HealthCare:
Predicting Pneumonia Risk and Hospital 30-day Readmission. Paper presented at: 21th ACM SGKDD
International Conference on Knowledge Discovery Data Mining2015; Sydney, NSW, Australia.
63. InterpretML. 1019. https://github.com/Microsoft/interpret.

This article is protected by copyright. All rights reserved

Figure Legends:

Table 1: Summary of included research papers for review

Figure 1: Schematic diagram of machine learning tools and applications to liver disease

Figure 2: Flowchart of search strategy and selection of studies for inclusion

Supplemental Figure 1: Growth of publications in PubMed with “hepatology” and “machine learning”.
Data compiled using Medline (PubMed) trend (http://dan.corlan.net/medline-trend.html).

Supplemental Figure 2: Artificial Intelligence in Precision Hepatology Medicine

 CATEGORY AUTHOR, YEAR PATIENT GROUP KEY OUTCOMES
Accepted Article
CHRONIC LIVER
DISEASE, CIRRHOSIS
Gatos et al,
2017
Clinical data from 126 patients (56
healthy, 70 with CLD)
Highest accuracy in classification of healthy to CLD was 87.3%
Introduced a machine-learning algorithm that quantifies color information in
terms of stiffness values
New objective parameters and criteria for CLD diagnosis
Liu et al, 91 ultrasound images, 44 from normal Fine-tuning of a deep convolutional neural network (CNN) model can extract
2017 people and 47 from people with liver capsule from ultrasound imaging and accurately classify cirrhotic versus
cirrhosis non-cirrhotic images
FATTY LIVER DISEASE Meffert et al, 4222 Pomeranian patients used to Fatty Liver Index (FLI, Hepatic Steatosis Index (HSI), and the developed risk
(FLD), HEPATIC 2014 develop scoring system; 2177 South scoring system were all able to discriminate between patients with and without
STEATOSIS, NON- German patients used as an external hepatic steatosis, but predicted risk and observed risks did not match well
ALCOHOLIC FATTY validation data set when applied to external data
LIVER DISEASE (NAFLD)
Sowa et al, Cohort of 31 NAFLD patients with BMI ML techniques that rely on ALT/AST ratio, adipokines, and cytokines were able
2014 <30, 51 ALD patients with cirrhosis, and to distinguish NAFLD and ALD; as well as distinguish between cirrhotic and non-
51 ALD patients without cirrhosis cirrhotic patients with ALD
Vanderbeck et 47 liver biopsies from patients with Classification algorithm performed with 89% overall accuracy. Worked best for
al, normal liver histology (n=20) and NAFLD identifying macrosteatosis; identification of central veins & portal arteries less
2014 (n=27) robust

Kuppili et al, 63 patients (27 normal, 36 abnormal) in Superior performance of ELM compared to SVM for all protocols, and all types
2017 US liver database of US data sets in terms of sensitivity, specificity, accuracy

Yip et al, 922 subjects from a population Laboratory parameter-based ML model (NAFLD ridge score) is a robust
2017 screening study were randomly divided reference for detecting NAFLD in the general population
into training and validation groups
Byra et al, 55 severely obese patients admitted for AUROC with proposed approach (Inception-ResNet-v2 deep convolutional
2018 bariatric surgery (mean age 40.1 ± 9.1, neural network) was higher than both the hepatorenal index method and the
mean BMI 45.9 ± 5.6, 20% of males) gray-level co-occurrence matrix algorithm
Islam et al, Dataset developed with 10 attributes Logistic Regression (LR) proved to be the best technique among other
2018 that included 994 liver patients (553 techniques (RF, SVM, and ANN) when taking into account accuracy, sensitivity,
female, 461 male) specificity, positive predictive value, and negative predictive value in prediction
of FLD
Ma et al, 10,508 patients of which 2522 patients Bayesian network model improves F-measure scores and is the best performing
2018 met the diagnostic criteria of NAFLD in accuracy, specificity, and sensitivity among 11 total MLAs
Novel ML techniques can have screening and predictive value for NAFLD
Perveen et al, Multiclass labeled dataset of 7 risk Proposed Decision Tree based method could help with management of NAFLD
2018 factors for 40,637 individuals over a patients evaluating risk and progression
period of 10 years
HEPATOCELLULAR Kim et al, Tissue samples from 59 patients with Identification of gene signatures that can be used as markers for diagnosing
CARCINOMA 2004 end stage CLD who received liver early onset HCC in high-risk populations using BRB ArrayTools – an Excel-based
transplantation platform
Tumor and matched nontumor tissue
samples from 74 patients

This article is protected by copyright. All rights reserved

 Singal et al, 442 patients with Child A or B cirrhosis Machine-learning algorithm had significantly better diagnostic accuracy
Accepted Article 2013 that were prospectively followed until
HCC development, LT, death or study
compared to the UM regression model and the HALT-C model at identifying
patients high-risk for developing HCC
termination
LIVER Cruz-Ramirez et 1003 liver transplants, patients A multi-objective evolutionary algorithm (MPENSGA2) showed better
TRANSPLANTATION al, undergoing partial, split, or living donor performance when compared to the mono-objective algorithm
2013 liver transplantation, combined/multi- Rule-based system minimizes prediction probability error and maximizes
visceral transplants were excluded probability of success in liver transplants
Briceno et al, Evaluation of 64 donor and recipient Artificial neural networks predicted with better accuracy 3-month graft survival
2014 variables from a set of 1003 liver and 3-month graft loss for each D-R pair than for other scores (MELD, D-MELD,
transplantations from a multicenter DRI, P-SOFT, SOFT, and BAR scores)
study including 11 Spanish centers
Lau et al, Liver Transplant Database in Australia Donor Risk Index predicts outcome with a relatively high AUC_ROC value (0.68)
2017 from 1988 to 2013, including brain-dead while the combination of factors used in Donor Risk Index yields a higher
and cardiac death organ donors while AUC)RIC score (0.764)
excluding paediatric recipients and Using a combination of donor, transplant, and recipient characteristics known
living-related donors at time of transplant allows for high accuracy in donor-recipient matching
Ayllon et al, 822 patients from King’s College Development of a rule-based system to facilitate decision making of the most
2018 Hospital (KCH), excluding pediatric liver appropriate D-R matching with better performance than MELD with
transplants (LT), living donor LTs, and improvements from previously reported results from the multicentric MADR-E
hepatocellular carcinoma database
Bertsimas et al, Waitlist, deceased donor, transplant, Optimized prediction of mortality (OPOM) objectively and accurately prioritizes
2018 and follow-up data from the Organ candidates for liver transplantation based on disease severity
Procurement and Transplantation OPOM allocation shows reduced mortality in the ML model compared to MELD
Network (OPTN) Standard Transplant
Analysis & Research (STAR) dataset
Bhat et al, All adult LT recipients between 1987 and Machine learning identified risk factors for new onset diabetes after
2018 2016 divided into training (70%) and transplantation (older age, male, and obesity)
validation (30%) data sets Donor features do not affect risk
Lee et al, 1211 patients who underwent LDLT or Gradient boosting machine had the best performance with highest AUROC
2018 DDLT, pediatric cases and cases with when predicting AKI after liver transplantation
missing baseline serum creatinine were Development of an internet-based risk estimator based on the gradient
excluded boosting model
PRIMARY SCLEROSING Andres et al, 2769 patients that received liver PSSP, a software tool, was able to predict individual survival distributions for
CHOLANGITIS (PSC) 2018 transplantation for Primary Sclerosing novel patients
Cholangitis (PSC) in the Scientific PSSP showed great D-calibration, and passed the single-time calibration test in
Registry of Transplant Recipients contrast to the Cox regression model, which displayed worse calibration and
failed at 10 years
Eaton et al, 509 subjects validated in an PREsTo accurately predicts hepatic decompensation in PSC and performs better
2018 international multicenter cohort (n=278) than other systems
PREsTo is also accurate in individuals with a bilirubin < 2.0 mg/dL
VIRAL HEPATITIS Huang et al, Independently enrolled validation Area-under-the-ROC curves of the developed score (CRS) was 0.75 in the
2007 cohort of 154 Caucasian patients (ages training cohort. In the validation cohort, AUC was 0.53 for clinical factors, 0.73
18-75 years, anti-HCV and HCV RNA for CRS, and 0.76 when the two were combined. CRS is a better predictor than

This article is protected by copyright. All rights reserved

 positive; excluded for co-existing liver clinical factors in differentiating high-risk vs low-risk for cirrhosis
Accepted Article Wang et al,
diseases, HIV, HBV, and HCC)
339 chronic hepatitis B patients divided Artificial neural network model (ANN) based on routine and serum markers
2010 randomly into 2 subsets (training and predict risk for fibrosis with high accuracy
validation set) Most important factors in predictive model were age, AST, platelet, and GGT
116 additional patients enrolled
Stoean et al, 722 patients with CHC infection and 24 Evolutionary method is superior to standard vector machines while offering
2010 indicators direct expression of influence of indicators on corresponding stages of fibrosis
Cao et al, 239 patients (124 LC patients with CHB, MLP, a new classifier had high accuracy in discriminating cirrhotic patients from
2013 115 patients with CHB) non-cirrhotic patients by using ML method based on the given clinical
parameters
MLP could be used in HBV-induced cirrhosis assessment
Cao et al, 162 serum samples from 44 LC patients LC-NB and LC-MLP are two novel methods for prediction of HBV-induced LC
2013 with CHB, 46 patients with CHB, and 72 with high sensitivity, specificity, and accuracy compared to existing models
healthy individuals Serum biomarkers can be determined for discriminating LC and non-LC
Lara et al, Sequences of HCV 1b isolates from 44 Strong genetic association to rate of fibrosis progression indicates quantifiable
2014 liver transplanted or non-transplanted effect of HCV heterogeneity on liver disease severity
immunocompetent patients with slow RFP-relevant genetic markers may be useful for prediction of RFP regardless of
and fast rates of fibrosis progression (17 transplant status
fast, 25 slow); 2 patients excluded due
to insufficient data
Konerman et al, Cohort of 184 patients with fibrosis Prediction models that use longitudinal data outperform baseline models as
2015 progression and 349 patients with they are able to capture nonlinear disease progression in chronic hepatitis C
clinical outcome
Konerman et al, 1007 adults with CHC without hepatic Using routinely collected data from a heterogeneous cohort of CHC patients
2017 decompensation, HCC, LT, HBV or HIV allows accurate assessments for risk of clinical outcomes
co-infection
Shousha et al, 427 patients with chronic hepatitis C IL28B genotype selected by REPTree to be the best predictor for advanced
2018 who were naïve candidates for an fibrosis. Using Auto-WEKA, the multilayer perception (MLP) was selected to be
antiviral therapy the best predictive algorithm with regards to sensitivity, specificity, and ROC
area when compared to APRI, FIB-4, and REPTree
Wei et al, 490 HBV infected subjects, models Gradient boosting based prediction system displayed improvements relative to
2018 validated in an independent HBV dataset FIB-4 in HBV and HCV cohorts. Higher cut-offs for gradient boosting and FIB-4
(n=86), further application of model on required to achieve similar specificity and sensitivity
two HCV datasets (n=254, 230)

This article is protected by copyright. All rights reserved

 Search Terms: “machine learning” [All Fields] OR “artificial neural networks” [All Fields] OR (predictive[All Fields] AND
Accepted Article modeling/model[All Fields] AND (”Model Driven Eng Lang Syst”[Journal] OR “models”[All Fields])) OR “natural
language processing”[All Fields] OR “deep learning”[All Fields] AND (”liver transplantation”[MeSH Terms] OR
(”liver[All Fields] AND “transplantation” [All Fields]) OR “liver transplantation”[All Fields] OR (”liver [All Fields]
Database: MEDLINE (PubMed) and Embase (Elsevier)

With a combination of various liver disease search terms: Hepatitis B/C,

Primary Sclerosing Cholangitis, Liver Disease, NASH, NAFLD, Hepatocellular
Carcinoma, Alcoholic Liver Disease, Viral Hepatitis
(n = 487)

Publications identifies after duplicates removed:

(n = 182)

Records Screened:
(n = 182) Records Excluded:
(n = 122)
- No relation to liver disease, or minor/no
application of machine learning techniques
by title/abstract (n = 121)
Full test articles assessed for eligibility:
Excluded: (n = 60)
(n = 17)
- Not specific to liver patient
populations (n = 17)
Studies included:
(n = 43)

Excluded:
Manuscript review and application of inclusion criteria
(n = 3)
- Abstract only (n = 1)
- No ML technique used (n = 1)
- Mice models (n = 1)
Studies included:
(n = 40)

hep_31103_f1.eps

This article is protected by copyright. All rights reserved

 MACHINE LEARNING in LIVER DISEASE

A Naïve Bayes Classifier & the B Gradient Boosting Machines To Predict Primary C Generative Adversarial Model
Accepted Article
risk of Cirrhosis in Viral Hepatitis

Posterior =
Prior
* Likelihood
Evidence
Sclerosing Cholangitis Survival

Low Survival
Real
Data

High Survival
Cirrhosis No Cirrhosis
Noise
Synthetic
Generator Data Real or Synthetic
Discriminator
Data?

D Support Vector Machine to Classify Shear Wave E Neural Networks & Random Forest Predict PTDM Risk in Transplant Population
Elastography (SWE) Images Hidden Layer

Output Layer
Dataset
SWE Chronic Liver Disease

Tree 1 Tree 2 Tree n

Healthy
Risk of
Input Layer
PTDM
Mean
Neural Networks
Random Forest
Prediction

F K-Nearest Neighbors G Variational Auto-Encoder Model

Sample
Real Data Generated Data
Class A
Class B σ

Class ? μ

hep_31103_f2.eps

This article is protected by copyright. All rights reserved