#x201c Beyond MELD” – Emerging Stra
#x201c Beyond MELD” – Emerging Stra
#x201c Beyond MELD” – Emerging Stra
Summary
Keywords: MELD; In this review article, we discuss the model for end-stage liver disease (MELD) score and its dual purpose
Prognostication; Allocation;
in general and transplant hepatology. As the landscape of liver disease and transplantation has evolved
Frailty; Sarcopenia; EHR;
OMOP; Clinical considerably since the advent of the MELD score, we summarise emerging concepts, methodologies, and
Decision Support. technologies that may improve mortality prognostication in the future. Finally, we explore how these
novel concepts and technologies may be incorporated into clinical practice.
Received 2 December 2021;
received in revised form 24 © 2022 Published by Elsevier B.V. on behalf of European Association for the Study of the Liver.
February 2022; accepted 4
March 2022 Introduction
The deficit of available donor organs in relation to and recalibration of the MELD score, MELD 3.0, was
the number of patients in need of liver trans- recently published with the inclusion of sex and
plantation necessitates systems to allocate organs serum albumin.10 At the same time, a substantial
1
Division of Gastroenterology and
in an efficient yet equitable manner. The current proportion of liver transplants are allocated by
Hepatology, Department of principles of liver allocation in the United States,1 MELD “exception”, representing indications where
Medicine, University of California – the Eurotransplant region,2,3 and elsewhere the mortality risk and need for transplant are not
San Francisco, San Francisco, CA,
include determination of priority through objective well-represented by the MELD score.11
USA; 2Division of Gastroenterology
and Hepatology, Department of and measurable medical criteria, ordered from In addition, emerging technologies, new meth-
Medicine, Stanford University most to least medically urgent.1,4 Urgency has been odologies, and evolving conceptual frameworks for
School of Medicine, Stanford, represented primarily by the model for end-stage liver disease may improve clinicians’ ability to
CA, USA
liver disease (MELD) score, rather than the Child- prognosticate and manage patients with end-stage
* Corresponding author. Address: Pugh score, to avoid subjective variables such as liver disease. In this article, we present emerging
430 Broadway Street, Floor 3,
ascites and encephalopathy and to expand the tools and techniques “beyond MELD” for
Redwood City, CA 94063-3132,
USA; Tel.: +1-650-723-5135, fax: scale (to reduce the number of candidates with improvement in liver allocation, prognostication,
+1-650-723-5488. identical scores).5,6 and outcomes in patients with end-stage
E-mail address: wrkim@stanford. The MELD score, which is comprised of serum liver disease.
edu (W.R. Kim). bilirubin, creatinine, and the international nor-
https://doi.org/
malised ratio, has since served a dual purpose in Beyond MELD – for liver allocation
10.1016/j.jhep.2022.03.003 general and transplant hepatology. It effectively Improving the MELD score
predicts short-term (e.g., over 90 days) mortality Over the past two decades of MELD score-based
among patients with chronic liver disease, thereby liver allocation, the demographics of chronic liver
providing clinicians with a critical tool to prog- disease and indications for liver transplantation
nosticate liver-related and waitlist mortality. It has have changed dramatically worldwide. The wide-
been used to determine medical urgency (and spread availability of effective direct-acting anti-
hence priority) for liver transplant candidates since viral therapy for hepatitis C and the increasing
2002 in the United States and 2006 in the Euro- prevalence of alcohol-associated liver disease and
transplant region, making it an essential tool for non-alcoholic steatohepatitis has fundamentally
transparent and equitable organ allocation.7,8 changed the population of patients awaiting liver
The landscape of chronic liver disease and liver transplantation.11,12 Throughout these changes,
transplantation has evolved considerably in the last however, the MELD score has continued to provide
two decades. Both waitlist mortality prediction and robust predictions of short-term waitlist mortality
transplant organ allocation require ongoing re- that outperform most other clinical scoring sys-
evaluation to ensure accurate prognostication and tems, with c-statistics that exceed 0.80 in various
appropriate distribution of donor organs. In 2016, cohorts.9,10,13 Still, it has been perceived that the
the MELD score was updated to include serum predictive power of the MELD score may have
sodium, an objective biomarker that is often a diminished in recent years.14,15 The MELD score
surrogate indicator for ascites.9 A new update to may not represent mortality risk as accurately for
Table 1. Limitations of existing and proposed waitlist mortality risk scores to be used in liver allocation.
Score Components Strengths Limitations
Child-Pugh score139 Bilirubin, INR, albumin, asci- Established minimal listing criteria for Inclusion of potentially subjective vari-
tes, encephalopathy liver transplant candidates ables i.e. ascites and encephalopathy
MELD7 Bilirubin, INR, creatinine Adequate discriminative ability Underestimation of renal dysfunction in
Use of objective and widely available women compared to men
tests Does not accurately represent transplant
Improved waitlist mortality, equity in urgency for certain disease etiologies such
liver allocation as hepatocellular carcinoma
MELD-Na9 Bilirubin, INR, creatinine, Addition of sodium as a surrogate for May not accurately represent mortality
sodium ascites risk for complications such as hepatic en-
cephalopathy or acute-on-chronic liver
failure
MELD-Plus140 Bilirubin, INR, creatinine, so- Improved mortality prediction Only calculated after a cirrhosis-related
dium, albumin, total choles- compared to MELD-Na after hospital hospital admission
terol, WBC, age, length of stay admission
MELD-lactate141 Bilirubin, INR, creatinine, so- Improved in-hospital mortality pre- Only calculated during a hospital
dium, lactate diction compared to MELD or MELD- admission
Na in patients hospitalised for infec-
tion or MELD < −15
MELD-Na-MDRD18,22 Bilirubin, INR, creatinine, age, More accurate estimation of renal Did not improve mortality prediction
sex, race function accounting for potential dif-
ferences in muscle mass
MELD-GRAIL-Na23,24 Bilirubin, INR, creatinine, Estimation of renal function developed Inclusion of age and race could lead to bias
blood urea nitrogen, age, sex, for liver disease with better accuracy in allocation
race, albumin, sodium and precision compared to standard
eGFR calculations
Improved mortality prediction in
MELD >32
MELD-Cystatin C25,26 Bilirubin, INR, creatinine, cys- Biomarker of renal function less sus- Lack of clinical availability
tatin C ceptible to differences in muscle mass Mitigated sex differences but no
improvement in predictive power
MELD-Na-Shift28 Bilirubin, INR, creatinine, Adds 0-1 MELD points for women Addition of points for women at arbitrary
sodium Modelled to eliminate sex disparity in levels
transplant rates
MELD 3.010 Bilirubin, INR, creatinine, so- Addition of 1.33 points for women Calculation somewhat more complex
dium, sex, albumin Updated coefficients and interactions;
adjusted upper bound for serum
creatinine
Improved mortality prediction
compared to MELD-Na
eGFR, estimated glomerular filtration rate; GRAIL, glomerular filtration rate assessment in liver disease; INR, international normalised ratio; MDRD, modification of diet in
renal disease; MELD, model for end-stage liver disease; WBC, white blood cell.
same medical urgency should have an equal op- related factors and thereby improve access.38–40
portunity of receiving a liver transplant, yet this is Consensus processes, such as that described by
currently not the case. Upcoming changes in allo- the Italian liver transplant community, may help to
cation in the United States include not only opti- develop allocation policy that fairly balances the
misation of the MELD score but also continuous various priorities of liver transplantation, including
distribution, a composite point scoring system that urgency, utility, and transplant benefit.37
will enable the consideration of additional vari-
ables, including height, body surface area, blood Beyond MELD – For prognostication
type, geography, paediatric status, and travel effi- Muscle dysfunction as a clinical marker for
ciency, and indication for transplant (i.e. excep- assessing disease severity in patients
tions), to move closer to fair and equitable organ with cirrhosis
allocation. Under the proposed framework defined Emerging factors that have not classically been
Key point by the Organ Procurement and Transplantation reflected by the MELD score, such as malnutrition,
Network (OPTN) in the United States, continuous frailty, and sarcopenia, have improved our ability
Factors not traditionally to dynamically characterise the morbidity and
distribution will attempt to balance 5 goals: med-
reflected by the MELD
ical urgency, post-transplant survival, candidate mortality associated with cirrhosis.41 Malnutrition
score, such as malnutrition,
frailty, and sarcopenia, biology, patient access, and placement efficiency, represents a spectrum of nutritional deficiencies
have improved prognosti- although the specific attributes ultimately included that cause adverse effects on physiologic function
cation in patients with and their respective weighting will depend on or clinical outcomes.42 It contributes to and is
cirrhosis.
feedback from the transplant community and interdependent with measurable clinical
modelling and analysis. The system is envisioned to manifestations of muscle dysfunction: frailty
provide a more dynamic reflection of patient- and sarcopenia.41
European Union.64,77 While the trend towards longitudinal analyses of EHR data to predict out-
common data models and centralised EHR data for comes of cirrhosis.92
observational research had already been underway, In liver transplant, ML methodologies have been
the COVID-19 pandemic drastically accelerated this used to explore waitlist mortality and organ allo-
movement with the creation of the National COVID cation.87,88,92–96 One of the first ML models in
Cohort Collaborative (N3C).65,78 transplant hepatology developed in 2003 was an
N3C is a novel, centralised, and harmonised ANN model to predict 1-year mortality in a cohort
repository of EHR data from more than 64 sites of 92 patients. While limited in scale, this ANN
from across the United States built on the OMOP model outperformed logistic regression and the
platform, formed in response to the need for rapid Child-Pugh score.93 Similarly, an ANN-based mor-
accrual and analyses of clinical data during the tality model derived from patients awaiting liver
COVID-19 pandemic.65,78 Its effective use has transplantation in Italy and validated in the United
allowed for the rapid generation of insights into Kingdom showed better predictive ability than the
the mortality risk of SARS-CoV-2 infection among original MELD score.94 The optimised prediction of
patients with cirrhosis.79 The work highlights the mortality model – developed in 2019 and trained
prospect of transplant hepatology-specific multi- on OPTN data using ML optimal classification trees
centre EHR collaboratives with deep clinical con- – demonstrated superior mortality prediction vs.
tent expertise, which may accelerate the the MELD score, and led to decreased mortality and
development of comprehensive models for mor- increased survival benefit across all candidate de-
tality prediction in patients with end-stage mographics, diagnoses, and geographic regions in
liver disease. liver simulated allocation model simulations.97
Despite these encouraging results, ML models
Novel modelling methodologies for mortality for waitlist mortality have several limitations,
risk prediction including interoperability and complexity. In
While high-dimension multicentre EHR data has addition, many early applications of ML method-
tremendous potential, their “big data” nature may ologies have only considered binary outcomes
require the use of novel analytical techniques.80,81 rather than a time-dependent survival function
“Big data” is an amorphous term that is classically which is key in the accurate determination of
defined in terms of the 5 “Vs” (volume, velocity, transplant urgency and waitlist priority. Due to
variety, veracity, and value) to describe large these limitations and challenges in practical
datasets that may be more effectively analysed implementation, waitlist mortality models based
using 82,83 artificial intelligence-based methods, on ML have yet to gain much traction in organ
such as machine learning (ML), which permit data- allocation.98,99 ML models have the potential to
driven rather than hypothesis-driven discov- better predict post-transplant outcomes through
ery.84,85 The most prevalent ML algorithms are the real-time considerations of longitudinal
divided into supervised (classification) and unsu- candidate variables, donor variables, and the
pervised (sorting) methods (Table 2).84,86,87 interaction of donor-candidate matching, which
There is often some overlap between traditional may play a role in continuous distribution.38,39
statistical and ML approaches: Logistic regression
is such an example of a methodology common to Potential pitfalls of algorithms for
both. In general, classification trees and neural clinical prediction
network-based methods have generally been the While there is substantial potential for ML to in-
predominant ML algorithms applied to contempo- fluence clinical practice in transplant hepatology
rary hepatology research. The cirrhosis mortality and potentially improve patient outcomes, limita-
model, developed from the United States Veterans tions of these technologies must be recognised.85
Affairs Corporate Data Warehouse (VHACDW) us- First, additional complexity may not improve pre-
ing a combination of gradient boosting and logistic dictive performance if underlying data and vari-
regression methods, offered significantly improved ables are the same. When comparing the ability of
discrimination compared to the MELD score.88 Of ML models (support vector classification and
particular interest are artificial neural networks random forest) vs. logistic regression to predict
(ANNs), which are learning algorithms that can be readmission and death in 2,179 North American
employed for both supervised and unsupervised patients with ACLF, ML model accuracies were
tasks. Neural networks are inspired by neuro- equivalent to models generated using only the
anatomy – each neuron is a computing unit, and all MELD score. The performance of future ML
neurons are connected to build a network. Signals modelling may improve if higher density data
travel from input layer to the output layer going incorporating novel variables, such as sarcopenia
through multiple hidden layers – which represent and frailty, are available.100
higher complexity.89–91 Deep neural networks, Second, despite harmonisation and ration-
characterised by multiple layers between the input alisation of different ontologies and semantics, data
and output layers,91 have been utilised for quality, shift, and reproducibility are still ongoing
issues in the modelling of EHR data.80,101 Dataset and allocation for kidney transplant.29,107 In
shift describes the changes in model performance transplant hepatology, eGFR has been avoided in
due to temporal or spatial shifts between the clinical prognostication modelling due to its po-
population used for training and the population tential for exacerbating race- and sex-based dis-
upon which the algorithm is deployed.102,103 One parities. Human intelligence, in addition to
prominent recent example is the University of artificial intelligence, remains critically important
Michigan’s deactivation of a proprietary sepsis- for the thoughtful and deliberate selection of data
alert model due to shifts in patient populations features, variables and analytic methodologies.
during the COVID-19 pandemic.104 Dataset shift is Fourth, structured data, which forms the basis
not exclusive to ML algorithms but also to other for most classical models and ML algorithms at this
clinical prediction scoring systems. Periodic audits time, are limited by coding. For example, efforts to
and updating of scoring systems, such as the up- diagnose Fontan-associated liver disease were
date of MELD to MELD 3.0,10 are necessary to adapt limited by the lack of specific structured diagnostic
our clinical tools to changing conditions. codes across multiple clinical databases.108 The
Third, underlying bias can be amplified by volume of unstructured data far exceeds structured
clinical prediction and ML-based algorithms.105,106 data, with an estimated 90% of digital data in
The most prominent example in transplantation healthcare being unstructured. Incorporating or
is the incorporation of race in estimated glomer- converting unstructured data elements in the EHR,
ular filtration rate (eGFR) calculations, which have such as imaging reports, pathology reports, and
disadvantaged racial minorities in listing practices clinical documentation, into structured or tagged
features remains challenging. Transformation of design, and post hoc systems, which provide local
such data into structured data requires substantial and reproducible explanations for algorithm out-
cleaning, splitting, merging, validating, and sorting, puts, are now commonly utilised to enable greater
but does improve clinical representation in pre- trust in ML algorithms.116,120 Similarly, active
dictive analytics.109 incorporation of human knowledge, or expert-
Key point
Finally, algorithms are not anticipated to augmentation, in the algorithm construction pro-
There is an increasing push completely replace the “subjective” judgment of cess is another strategy to improve “explain-
to develop data-driven clinicians involved in the care of the peri- ability.”121 To begin to address these concerns, the
machine learning-based
transplant patient.110 For instance, significant development of standardised tools and evaluations
algorithms to further
improve outcome predic- technical expertise is required to conduct split liver on transplant reporting and assessments of bias in
tion in patients with liver transplantation,111 to use donor organs with tech- applied ML techniques is currently underway.102,122
disease. nical variants or higher risk features,112 or to suc-
cessfully transplant patients with complex surgical Beyond MELD – for improvement in
histories.113 These institution- and clinician- patient outcomes
specific knowledge and skills are often ill- Emerging technologies to actively manage
captured and ill-evaluated by algorithms. waitlist mortality risk
For these reasons, the application of ML-based One technology to overcome issues with unstruc-
artificial intelligence has received a mixed recep- tured data is natural language processing (NLP),
tion from both clinicians and the general popula- which is a suite of related techniques to convert
tion.114–116 Among clinicians, there are latent fears unstructured or narrative text into tagged or
that algorithms may ultimately replace their skills structured elements for analysis.123,124 There has
or functions.116,117 In addition, many clinicians are been particular interest in utilising NLP for the
uncomfortable with “black box” ML tools, even diagnosis of non-alcoholic fatty liver disease as this
though examples of similar opacity abound in condition is poorly documented in structured EHR
other diagnostic and therapeutic areas of clinical data.125,126 NLP has been used on abdominal ul-
medicine.118 Among providers and patients, there trasound, computerised tomography, and magnetic
is a concern about the loss of patient-provider re- resonance imaging reports from the VHACDW to
lationships, privacy in data use, and accountability rapidly screen patients with radiographic evidence
– namely who is responsible for adverse outcomes of fatty liver disease.126 In an analysis of clinical
due to clinical decisions influenced or augmented notes available for 38,575 patients enrolled in the
by artificial intelligence.114,115,119 There is an Mount Sinai BioMe cohort, NLP methods out-
increasing recognition that transparency, inter- performed ICD codes and text search.125
Key point
pretability, and explainability are necessary for Real-time clinical decision support (CDS) and
Clinical decision support long-term acceptance of artificial intelligence tools. prospective risk modelling are also emerging areas
and prospective risk Ante hoc systems, which are interpretable by of research/implementation in the management of
modelling are emerging
areas of research that are
hoped to lead to improve-
ments in the management
of patients with cirrhosis Feedback to
and those on the liver clinical decision
transplant waiting list. support
Clinical decision
Intervention
support
Usual care
Feedback to
computational
assessment
Fig. 1. Rapid-cycle testing in ‘electronic’ randomised controlled trials. Schematic of rapid-cycle ‘electronic’ randomised
controlled trials could be implemented using CDS systems: computational risk assessment allows a patient to be randomised for
an intervention associated with a CDS, the results of which could then be used to iteratively modify the risk stratification al-
gorithm or the CDS system. CDS, clinical decision support.
Authors’ contributions
Supplementary data
Authorship was determined using ICMJE recom-
Supplementary data to this article can be found on-
mendations. Ge: Drafting of manuscript; critical
line at https://doi.org/10.1016/j.jhep.2022.03.003.
revision of the manuscript for important intellec-
tual content. Kim: Drafting of manuscript; critical
References [16] Hernaez R, Liu Y, Kramer JR, Rana A, El-Serag HB, Kanwal F. Model for
end-stage liver disease-sodium underestimates 90-day mortality risk in
Author names in bold designate shared co-first authorship
patients with acute-on-chronic liver failure. J Hepatol 2020;73:1425–
1433. https://doi.org/10.1016/j.jhep.2020.06.005.
[1] Organ Procurement and Transplantation Network (OPTN). Final Rule as [17] Allen AM, Heimbach JK, Larson JJ, Mara KC, Kim WR, Kamath PS, et al.
revised by amendments. 1999. Reduced access to liver transplantation in women: role of height,
[2] Jochmans I, van Rosmalen M, Pirenne J, Samuel U. Adult liver allocation MELD exception scores, and renal function underestimation.
in eurotransplant. Transplantation 2017;101:1542–1550. https://doi.org/ Transplantation 2018;102:1710–1716. https://doi.org/10.1097/TP.
10.1097/TP.0000000000001631. 0000000000002196.
[3] Goudsmit BFJ, Putter H, Tushuizen ME, Vogelaar S, Pirenne J, Alwayn IPJ, [18] Myers RP, Shaheen AAM, Aspinall AI, Quinn RR, Burak KW. Gender, renal
et al. Refitting the model for end-stage liver disease for the eurotrans- function, and outcomes on the liver transplant waiting list: assessment
plant region. Hepatology 2021;74:351–363. https://doi.org/10.1002/hep. of revised MELD including estimated glomerular filtration rate. J Hepatol
31677. 2011;54:462–470. https://doi.org/10.1016/j.jhep.2010.07.015.
[4] Trotter JF. Liver transplantation around the world. Curr Opin Organ [19] Mathur AK, Schaubel DE, Gong Q, Guidinger MK, Merion RM. Sex-
Transpl 2017;22:123–127. https://doi.org/10.1097/MOT. based disparities in liver transplant rates in the United States. Am J
0000000000000392. Transpl 2011;11:1435–1443. https://doi.org/10.1111/j.1600-6143.2011.
[5] Wiesner R, Edwards E, Freeman R, Harper A, Kim R, Kamath P, et al. 03498.x.
Model for end-stage liver disease (MELD) and allocation of donor livers. [20] Verna EC, Lai JC. Time for action to address the persistent sex-based
Gastroenterology 2003;124:91–96. https://doi.org/10.1053/gast.2003. disparity in liver transplant access. JAMA Surg 2020;155:545–547.
50016. https://doi.org/10.1001/jamasurg.2020.1126.
[6] Freeman RB, Wiesner RH, Harper A, McDiarmid SV, Lake J, Edwards E, [21] Cholongitas E, Marelli L, Kerry A, Goodier DW, Nair D, Thomas M, et al.
et al. The new liver allocation system: moving toward evidence-based Female liver transplant recipients with the same GFR as male recipients
transplantation policy. Liver Transpl 2002;8:851–858. https://doi.org/ have lower MELD scores–a systematic bias. Am J Transpl 2007;7:685–
10.1053/jlts.2002.35927. 692. https://doi.org/10.1111/j.1600-6143.2007.01666.x.
[7] Kamath PS, Wiesner RH, Malinchoc M, Kremers W, Therneau TM, [22] Leithead JA, MacKenzie SM, Ferguson JW, Hayes PC. Is estimated
Kosberg CL, et al. A model to predict survival in patients with end-stage glomerular filtration rate superior to serum creatinine in predicting
liver disease. Hepatology 2001;33:464–470. https://doi.org/10.1053/ mortality on the waiting list for liver transplantation? Transpl Int
jhep.2001.22172. 2011;24:482–488. https://doi.org/10.1111/j.1432-2277.2011.01231.x.
[8] Quante M, Benckert C, Thelen A, Jonas S. Experience since MELD [23] Asrani SK, Jennings LW, Kim WR, Kamath PS, Levitsky J, Nadim MK, et al.
implementation: how does the new system deliver? Int J Hepatol MELD-GRAIL-Na: glomerular filtration rate and mortality on liver-
2012;2012:264015. https://doi.org/10.1155/2012/264015. transplant waiting list. Hepatology 2020;71:1766–1774. https://doi.org/
[9] Kim WR, Biggins SW, Kremers WK, Wiesner RH, Kamath PS, Benson JT, 10.1002/hep.30932.
et al. Hyponatremia and mortality among patients on the liver- [24] Asrani SK, Jennings LW, Trotter JF, Levitsky J, Nadim MK, Kim WR, et al.
transplant waiting list. N Engl J Med 2008;359:1018–1026. https://doi. A model for glomerular filtration rate assessment in liver disease (GRAIL)
org/10.1056/NEJMoa0801209. in the presence of renal dysfunction. Hepatology 2019;69:1219–1230.
[10] Kim WR, Mannalithara A, Heimbach JK, Kamath PS, Asrani SK, https://doi.org/10.1002/hep.30321.
Biggins SW, et al. MELD 3.0: the model for end-stage liver disease [25] Finkenstedt A, Dorn L, Edlinger M, Prokop W, Risch L, Griesmacher A,
updated for the modern era. Gastroenterology 2021. https://doi.org/10. et al. Cystatin C is a strong predictor of survival in patients with cirrhosis:
1053/j.gastro.2021.08.050. is a cystatin C-based MELD better? Liver Int 2012;32:1211–1216. https://
[11] Kwong AJ, Kim WR, Lake JR, Smith JM, Schladt DP, Skeans MA, et al. doi.org/10.1111/j.1478-3231.2012.02766.x.
OPTN/SRTR 2019 annual data report: liver. Am J Transpl 2021;21(Suppl [26] De Souza V, Hadj-Aissa A, Dolomanova O, Rabilloud M, Rognant N,
2):208–315. https://doi.org/10.1111/ajt.16494. Lemoine S, et al. Creatinine- versus cystatine C-based equations in
[12] Younossi ZM, Stepanova M, Younossi Y, Golabi P, Mishra A, Rafiq N, et al. assessing the renal function of candidates for liver transplantation with
Epidemiology of chronic liver diseases in the USA in the past three decades. cirrhosis. Hepatology 2014;59:1522–1531. https://doi.org/10.1002/hep.
Gut 2020;69:564–568. https://doi.org/10.1136/gutjnl-2019-318813. 26886.
[13] Leise MD, Kim WR, Kremers WK, Larson JJ, Benson JT, Therneau TM. [27] Nephew LD, Goldberg DS, Lewis JD, Abt P, Bryan M, Forde KA. Exception
A revised model for end-stage liver disease optimizes prediction of points and body size contribute to gender disparity in liver trans-
mortality among patients awaiting liver transplantation. Gastroenter- plantation. Clin Gastroenterol Hepatol 2017;15:1286–1293.e2. https://
ology 2011;140:1952–1960. https://doi.org/10.1053/j.gastro.2011.02.017. doi.org/10.1016/j.cgh.2017.02.033.
[14] Godfrey EL, Malik TH, Lai JC, Mindikoglu AL, Galván NTN, Cotton RT, et al. [28] Wood NL, VanDerwerken D, Segev DL, Gentry SE. Correcting the sex
The decreasing predictive power of MELD in an era of changing etiology disparity in MELD-Na. Am J Transpl 2021;21:3296–3304. https://doi.org/
of liver disease. Am J Transpl 2019;19:3299–3307. https://doi.org/10. 10.1111/ajt.16731.
1111/ajt.15559. [29] Vyas DA, Eisenstein LG, Jones DS. Hidden in plain sight - reconsidering
[15] Kwong A, Mannalithara A, Kim WR. Reply to: “The decreasing predictive the use of race correction in clinical algorithms. N Engl J Med
power of MELD in an era of changing etiology of liver disease”. Am J 2020;383:874–882. https://doi.org/10.1056/NEJMms2004740.
Transpl 2020;20:901–902. https://doi.org/10.1111/ajt.15733.
[70] Chua H-R, Zheng K, Vathsala A, Ngiam K-Y, Yap H-K, Lu L, et al. Health [92] Guo A, Mazumder NR, Ladner DP, Foraker RE. Predicting mortality
care analytics with time-invariant and time-variant feature importance among patients with liver cirrhosis in electronic health records with
to predict hospital-acquired acute kidney injury: observational longi- machine learning. PLoS One 2021;16:e0256428. https://doi.org/10.1371/
tudinal study. J Med Internet Res 2021;23:e30805. https://doi.org/10. journal.pone.0256428.
2196/30805. [93] Banerjee R, Das A, Ghoshal UC, Sinha M. Predicting mortality in patients
[71] Weisenthal SJ, Quill C, Farooq S, Kautz H, Zand MS. Predicting acute with cirrhosis of liver with application of neural network technology.
kidney injury at hospital re-entry using high-dimensional electronic J Gastroenterol Hepatol 2003;18:1054–1060. https://doi.org/10.1046/j.
health record data. PLoS One 2018;13:e0204920. https://doi.org/10.1371/ 1440-1746.2003.03123.x.
journal.pone.0204920. [94] Cucchetti A, Vivarelli M, Heaton ND, Phillips S, Piscaglia F, Bolondi L, et al.
[72] Rajkomar A, Oren E, Chen K, Dai AM, Hajaj N, Hardt M, et al. Scalable and Artificial neural network is superior to MELD in predicting mortality of
accurate deep learning with electronic health records. Npj Digital Med patients with end-stage liver disease. Gut 2007;56:253–258. https://doi.
2018;1:18. https://doi.org/10.1038/s41746-018-0029-1. org/10.1136/gut.2005.084434.
[73] Ge J, Najafi N, Zhao W, Somsouk M, Fang M, Lai JC. A methodology to [95] Ioannou GN, Tang W, Beste LA, Tincopa MA, Su GL, Van T, et al. Assess-
generate longitudinally updated acute-on-chronic liver failure prog- ment of a deep learning model to predict hepatocellular carcinoma in
nostication scores from electronic health record data. Hepatol Commun patients with hepatitis C cirrhosis. JAMA Netw Open 2020;3:e2015626.
2021;5:1069–1080. https://doi.org/10.1002/hep4.1690. https://doi.org/10.1001/jamanetworkopen.2020.15626.
[74] Haendel MA, Chute CG, Robinson PN. Classification, ontology, and pre- [96] Ferrarese A, Sartori G, Orrù G, Frigo AC, Pelizzaro F, Burra P, et al.
cision medicine. N Engl J Med 2018;379:1452–1462. https://doi.org/10. Machine learning in liver transplantation: a tool for some unsolved
1056/NEJMra1615014. questions? Transpl Int 2021;34:398–411. https://doi.org/10.1111/tri.
[75] Atiemo K, Skaro A, Maddur H, Zhao L, Montag S, VanWagner L, et al. 13818.
Mortality risk factors among patients with cirrhosis and a low model for [97] Bertsimas D, Kung J, Trichakis N, Wang Y, Hirose R, Vagefi PA. Develop-
End-Stage Liver Disease Sodium score (< −15): an analysis of liver trans- ment and validation of an optimized prediction of mortality for candi-
plant allocation policy using aggregated electronic health record data. dates awaiting liver transplantation. Am J Transpl 2019;19:1109–1118.
Am J Transpl 2017;17:2410–2419. https://doi.org/10.1111/ajt.14239. https://doi.org/10.1111/ajt.15172.
[76] Health Level Seven International - Homepage | HL7 International n.d. [98] Kwong AJ, Asrani SK. Artificial neural networks and liver trans-
https://www.hl7.org/ (accessed October 3, 2021). plantation: are we ready for self-driving cars? Liver Transpl
[77] European Health Data Evidence Network – ehden.eu n.d. https://www. 2018;24:161–163. https://doi.org/10.1002/lt.24993.
ehden.eu/ (accessed November 20, 2021). [99] Miller PE, Pawar S, Vaccaro B, McCullough M, Rao P, Ghosh R, et al.
[78] Bennett TD, Moffitt RA, Hajagos JG, Amor B, Anand A, Bissell MM, et al. Predictive abilities of machine learning techniques may be limited by
The national COVID cohort collaborative: clinical characterization and dataset characteristics: insights from the UNOS database. J Card Fail
early severity prediction. medRxiv 2021. https://doi.org/10.1101/2021.01. 2019;25:479–483. https://doi.org/10.1016/j.cardfail.2019.01.018.
12.21249511. [100] Hu C, Anjur V, Saboo K, Reddy KR, OʼLeary J, Tandon P, et al. Low predictability
[79] Ge J, Pletcher MJ, Lai JC, N3C Consortium. Outcomes of SARS-CoV-2 of readmissions and death using machine learning in cirrhosis. Am J Gas-
infection in patients with chronic liver disease and cirrhosis: a na- troenterol 2020. https://doi.org/10.14309/ajg.0000000000000971.
tional COVID cohort collaborative study. Gastroenterology 2021. https:// [101] Raghupathi W, Raghupathi V. Big data analytics in healthcare: promise
doi.org/10.1053/j.gastro.2021.07.010. and potential. Health Inf Sci Syst 2014;2:3. https://doi.org/10.1186/2047-
[80] Rumsfeld JS, Joynt KE, Maddox TM. Big data analytics to improve car- 2501-2-3.
diovascular care: promise and challenges. Nat Rev Cardiol 2016;13:350– [102] Bates DW, Auerbach A, Schulam P, Wright A, Saria S. Reporting and
359. https://doi.org/10.1038/nrcardio.2016.42. implementing interventions involving machine learning and artificial
[81] Genta RM, Sonnenberg A. Big data in gastroenterology research. Nat Rev intelligence. Ann Intern Med 2020;172:S137–S144. https://doi.org/10.
Gastroenterol Hepatol 2014;11:386–390. https://doi.org/10.1038/nrgas- 7326/M19-0872.
tro.2014.18. [103] Finlayson SG, Subbaswamy A, Singh K, Bowers J, Kupke A, Zittrain J, et al.
[82] Favaretto M, De Clercq E, Schneble CO, Elger BS. What is your definition The clinician and dataset shift in artificial intelligence. N Engl J Med
of Big Data? Researchers’ understanding of the phenomenon of the 2021;385:283–286. https://doi.org/10.1056/NEJMc2104626.
decade. PLoS One 2020;15:e0228987. https://doi.org/10.1371/journal. [104] Wong A, Otles E, Donnelly JP, Krumm A, McCullough J, DeTroyer-Cooley O,
pone.0228987. et al. External validation of a widely implemented proprietary sepsis
[83] Asri H, Mousannif H, Al Moatassime H, Noel T. Big data in healthcare: prediction model in hospitalized patients. JAMA Intern Med
Challenges and opportunities. In: 2015 International Conference on 2021;181:1065–1070. https://doi.org/10.1001/jamainternmed.2021.2626.
Cloud Technologies and Applications (CloudTech). IEEE; 2015. p. 1–7. [105] What Do We Do About the Biases in AI? n.d. https://hbr.org/2019/10/
https://doi.org/10.1109/CloudTech.2015.7337020. what-do-we-do-about-the-biases-in-ai (accessed October 3, 2021).
[84] Wiens J, Shenoy ES. Machine learning for healthcare: on the verge of a [106] Gianfrancesco MA, Tamang S, Yazdany J, Schmajuk G. Potential biases in
major shift in healthcare epidemiology. Clin Infect Dis 2018;66:149–153. machine learning algorithms using electronic health record data. JAMA
https://doi.org/10.1093/cid/cix731. Intern Med 2018;178:1544–1547. https://doi.org/10.1001/jamainternmed.
[85] Obermeyer Z, Weinstein JN. Adoption of artificial intelligence and ma- 2018.3763.
chine learning is increasing, but irrational exuberance remains. NEJM [107] Kuppachi S, Norman SP, Lentine KL, Axelrod DA. Using race to estimate
Catal 2020;1. https://doi.org/10.1056/CAT.19.1090. glomerular filtration and its impact in kidney transplantation. Clin
[86] Rajkomar A, Dean J, Kohane I. Machine learning in medicine. N Engl J Transpl 2021;35:e14136. https://doi.org/10.1111/ctr.14136.
Med 2019;380:1347–1358. https://doi.org/10.1056/NEJMra1814259. [108] Kim MH, Nguyen A, Lo M, Kumar SR, Bucuvalas J, Glynn EF, et al. Big data
[87] Spann A, Yasodhara A, Kang J, Watt K, Wang B, Goldenberg A, et al. in transplantation practice-the devil is in the detail-Fontan-associated
Applying machine learning in liver disease and transplantation: a liver disease. Transplantation 2021;105:18–22. https://doi.org/10.1097/
comprehensive review. Hepatology 2020;71:1093–1105. https://doi.org/ TP.0000000000003308.
10.1002/hep.31103. [109] Adnan K, Akbar R, Khor SW, Ali ABA. Role and challenges of unstructured
[88] Kanwal F, Taylor TJ, Kramer JR, Cao Y, Smith D, Gifford AL, et al. Devel- big data in healthcare. In: Sharma N, Chakrabarti A, Balas VE, editors.
opment, validation, and evaluation of a simple machine learning model Data management, analytics and innovation: proceedings of ICDMAI
to predict cirrhosis mortality. JAMA Netw Open 2020;3:e2023780. 2019, volume 1, vol. 1042. Singapore: Springer Singapore; 2020. p. 301–
https://doi.org/10.1001/jamanetworkopen.2020.23780. 323. https://doi.org/10.1007/978-981-32-9949-8_22.
[89] Jain AK, Mao Jianchang, Mohiuddin KM. Artificial neural networks: a [110] Kuo RYL, Harrison CJ, Jones BE, Geoghegan L, Furniss D. Perspectives: a
tutorial. Computer (Long Beach Calif) 1996;29:31–44. https://doi.org/10. surgeon’s guide to machine learning. Int J Surg 2021;94:106133. https://
1109/2.485891. doi.org/10.1016/j.ijsu.2021.106133.
[90] Schuster M, Paliwal KK. Bidirectional recurrent neural networks. IEEE Trans [111] Cauley RP, Vakili K, Fullington N, Potanos K, Graham DA, Finkelstein JA,
Signal Process 1997;45:2673–2681. https://doi.org/10.1109/78.650093. et al. Deceased-donor split-liver transplantation in adult recipients: is
[91] LeCun Y, Bengio Y, Hinton G. Deep learning. Nature 2015;521:436–444. the learning curve over? J Am Coll Surg 2013;217:672–684.e1. https://doi.
https://doi.org/10.1038/nature14539. org/10.1016/j.jamcollsurg.2013.06.005.