Int J Clin Exp Med 2015;8(1):751-757

www.ijcem.com /ISSN:1940-5901/IJCEM0003706

Original Article
Child-Pugh versus MELD score for
predicting the in-hospital mortality of acute
upper gastrointestinal bleeding in liver cirrhosis
Ying Peng*, Xingshun Qi*, Junna Dai, Hongyu Li, Xiaozhong Guo

Liver Cirrhosis Study Group, Department of Gastroenterology, General Hospital of Shenyang Military Area, 83
Wenhua Road, Shenyang 110840, China. *Equal contributors.
Received November 8, 2014; Accepted January 9, 2015; Epub January 15, 2015; Published January 30, 2015

Abstract: A retrospective study was conducted to compare the performance of Child-Pugh and Model for End-Stage
Liver Diseases (MELD) scores for predicting the in-hospital mortality of acute upper gastrointestinal bleeding (UGIB)
in patients with liver cirrhosis. A total of 145 patients with a diagnosis of liver cirrhosis and acute UGIB between
July 2013 and June 2014 were retrospectively analyzed (male/female: 94/51; mean age: 56.77±11.33 years;
Child-Pugh class A/B/C: 46/64/35; mean Child-Pugh score: 7.88±2.17; mean MELD score: 7.86±7.22). The in-
hospital mortality was 8% (11/145). Areas under receiving-operator characteristics curve (AUROC) for predicting the
in-hospital mortality were compared between MELD and Child-Pugh scores. AUROCs for predicting the in-hospital
mortality for Child-Pugh and MELD scores were 0.796 (95% confidence interval [CI]: 0.721-0.858) and 0.810 (95%
CI: 0.736-0.870), respectively. The discriminative ability was not significant different between the two scoring sys-
tems (P=0.7241). In conclusion, Child-Pugh and MELD scores were similar for predicting the in-hospital mortality of
acute UGIB in cirrhotic patients.

Keywords: Liver cirrhosis, gastrointestinal bleeding, Child-Pugh, MELD, prognosis

Introduction rior to Child-Pugh class for the predicting the

Child-Pugh classification, including total biliru-
bin, albumin, international normalized ratio
(INR) or prothrombin time, hepatic encephalop-
athy, and ascites, is the most commonly used
scoring system for evaluating the prognosis of
liver cirrhosis [1]. Two of the five variables are
subjective, and the remaining three variables
are acquired from laboratory tests. In 2000,
Model for End-Stage Liver Disease (MELD)
score, a mathematical formula which is com-
posed of serum creatinine, total bilirubin, and
INR, is firstly introduced by the investigators
from Mayo Clinic to predict the mortality of
patients undergoing transjugular intrahepatic
portosystemic shunt (TIPS) insertions [2].
Notably, all of the variables in MELD scoring
system were objective [3]. Salerno et al. con-
firmed the superiority of MELD score over Child-
Pugh score in predicting the 3-month survival
in such patients [4]. However, Schepke et al.
showed that MELD score was only slightly supe-
long-term survival after TIPS [5]. At present,
MELD score is also widely used to prioritize the
organ allocation in candidates for liver trans-
plantation [3]. However, there is a debate about
whether MELD can replace Child-Pugh score for
predicting the survival in non-transplanted
patients with chronic liver disease [6, 7].
Acute upper gastrointestinal bleeding (UGIB) is
a lethal complication of liver cirrhosis [8, 9]. The
major predictors for early mortality of acute
UGIB in liver cirrhosis include hepatic encepha-
lopathy, Child-Pugh score or class, MELD score,
shock, renal failure, infection, hepatocellular
carcinoma, active bleeding, portal vein throm-
bosis, and hepatic venous pressure gradient [9,
10]. However, it remains unclear about whether
Child-Pugh or MELD score is better for predict-
ing the in-hospital mortality of acute UGIB in cir-
rhotic patients. Herein, we conducted a retro-
spective observational study to explore this
issue.
 Child-Pugh versus MELD
Figure 1. Patient selection.
Methods
All patients with a diagnosis of liver cirrhosis
who were admitted to the General Hospital of
Shenyang Military Region between July 2013
and June 2014 were retrospectively included in
the present study. Inclusion criteria were as fol-
lows. 1) Patients were diagnosed with liver cir-
rhosis based on the history of liver disease,
clinical manifestations, laboratory tests, imag-
ing tests, and liver biopsy, if necessary. 2)
Patents with hepatocellular carcinoma and
other malignancies were excluded by the dis-
ease history and imaging examinations. 3)
Patients presented with acute UGIB. The time
fame for the acute bleeding episodes should be
120 hours (5 days) according to the Baveno V
criteria [11]. 4) Source of acute UGIB was not
restricted. This was primarily because not all
patients underwent endoscopic examinations
at their emergent admissions. 5) Patients with
absence of complete laboratory tests were
excluded. The study protocol was approved by
the ethic committee of our hospital.
Clinical records were reviewed by two investiga-
tors (YP and JD), and checked by another inves-
tigator (XQ). The primary data collected at
admission were: the demographic data, causes
of liver diseases, severity of bleeding, vital
signs of hospitalized patients, laboratory data,
Child-Pugh score/class, and MELD score.
Additionally, we also collected endoscopic find-
752 Int J Clin Exp Med 2015;8(1):751-757
ings (i.e., location and grade of varices and red
color sign), treatment options (endoscopic liga-
tion or sclerotherapy, vasoactive drug, and/or
surgery, etc.), in-hospital death, and causes of
death.
Child-Pugh score was calculated based on the
severity of hepatic encephalopathy, ascites,
total bilirubin, albumin, and INR (1).
MELD score=9.57 × ln (creatinine [µmol/L] ×
0.01) + 3.78 × ln (bilirubin [µmol/L] × 0.05) +
11.2 × ln (INR) + 0.643 (3).
Statistical analysis
Categorical variables were reported as frequen-
cy (percentage) and continuous variables were
reported as mean ± standard deviations.
Receiving-operator characteristics (ROC) curve
analysis was performed to identify the discrimi-
native capacity of Child-Pugh and MELD scores
in predicting the risk of in-hospital death. A cut
off value of Child-Pugh score or MELD score
was chosen as both sensitivity and specificity
were optimal. Areas under the ROC curves
(AUROC) with 95% confidence intervals (CIs) for
these two scoring systems were also reported.
We compared the performance of the two scor-
ing systems by using the DeLong tests. P <
0.05 was considered statistically significant. All
statistical analyses were performed by using
the MedCalc software version 11.4.2.0.
Results
Overall, 849 patients with liver cirrhosis were
admitted to our hospital during the enrollment
period. Among them, 179 cirrhotic patients
without malignancy presented with acute UGIB.
Thirty-four patients were further excluded,
because some laboratory data for liver and
renal function were missing. Finally, 145
patients were included in the present study
(Figure 1).
Baseline characteristics at admission were
shown in Table 1. A majority of patients had a
history of viral hepatitis and alcohol abuse.
Endoscopic examinations were performed in
80% of patients. Child-Pugh and MELD scores
at admission were 7.88±2.17 and 7.86±7.22,
respectively.
Treatment options after admission were shown
in Table 2. Blood transfusion was given in 91
 Child-Pugh versus MELD

Table 1. Baseline characteristics of 145 patients patients. Somatosta-

Variables Values tin or its analogs were
prescribed in nearly
Sex (male/female) 94/51
all patients received.
Age (years) 56.77±11.33
Endoscopic therapy
Causes of liver diseases, n (%)
was performed in
Hepatitis B virus 46 (31.7) 104 patients. Splen-
Hepatitis C virus 11 (7.6) ectomy with devascu-
Alcohol 35 (24.1) larization was per-
Hepatitis B virus + Alcohol 3 (2.1) formed in 3 patients.
Hepatitis B virus + Hepatitis C virus 1 (0.7) Neither Sengstaken-
Unknown 35 (24.1) blakemore tube nor
Others 14 (9.7) TIPS was performed
Vital signs in any patients. The
Systolic blood pressure (mmHg) 118.83±20.12 in-hospital mortality
Diastolic blood pressure (mmHg) 67.34±11.26 was 8% (11/145).
Cause of death was
Heart rate (b.p.m.) 84.74±15.37
uncontrolled UGIB in
Interval between diagnosis of liver cirrhosis and admission (months) 55.33±71.60
all of the 11 patients.
Interval between bleeding and admission (hours) 32.54±31.88
Manifestation, n (%) In the ROC analysis,
Haematemesis 37 (25.5) Child-Pugh score had
Melena 54 (37.2) a cut-off value of 9
Haematemesis and melena 54 (37.2) with a specificity of
Diabetes (yes/no) 29/116 (20%) 63.6% and a sensitiv-
Laboratory tests ity of 79.1% (Figure
RBC (10*12/L) 2.65±0.70
2). AUROC was 0.796
(95% CI: 0.721-0.8-
Hb (g/L) 74.91±22.19
58). In the ROC analy-
WBC (10*12/L) 5.66±4.36
sis, MELD score had a
PLT (10*9/L) 66.00±62.58
cut-off value of 12
TBIL (umol/L) 28.18±25.58 with a specificity of
DBIL (umol/L) 14.27±18.40 83.6% and a sensitiv-
IBIL (umol/L) 13.87±11.08 ity of 72.7% (Figure
ALB (g/L) 30.35±6.79 3). AUROC was 0.810
ALT (U/L) 31.99±32.06 (95% CI: 0.736-
AST (U/L) 53.76±134.89 0.870). The discrimi-
ALP (U/L) 84.03±61.08 native ability was not
GGT (U/L) 65.21±93.64 significant different
BUN (mmol/L) 8.84±6.03 between the two scor-
CR (umol/L) 67.00±42.48 ing systems (Figure 4)
K (mmol/L) 4.07±0.52
(P=0.7241).
Na (mmol/L) 138.37±4.56 Discussion
Ca (mmol/L) 2.02±0.24
Blood ammonia (umol/L) 54.76±58.75 Numerous studies
PT (second) 18.08±6.24 have compared the
performance of Child-
APTT (second) 41.58±8.38
Pugh score with that
INR 1.54±0.79
of MELD score for the
Ascites, n (%) prognostic prediction
No 67 (46.2) in patients with liver
Mild 22 (15.2) diseases. As for the
Moderate and severe 56 (38.6) cirrhotic patients with

753 Int J Clin Exp Med 2015;8(1):751-757

 Child-Pugh versus MELD

Hepatic encephalopathy, n (%) mortality was 0.80

No 132 (91.0) and 0.76, respective-
ly. AUROC for the
Grade I-II 7 (4.8)
MELD and Child-Pugh
Grade III-IV 6 (4.1)
scores for predicting
Endoscopy (yes/no) 116/29 (80%)
the 3-month mortality
Varices, n (%) was 0.79 and 0.76,
Mild-Moderate 13 (9.0) respectively. Cerquei-
Severe 103 (71.0) ra et al. included 102
NA 29 (20.0) cirrhotic patients con-
Location of varices, n (%) secutively admitted
No 1 (0.7) with oesophageal var-
Esophageal varices 63 (43.4) iceal bleeding [14].
Gastric varices 16 (11.0) AUROC for the MELD
Esophageal and gastric varices 35 (24.1) and Child-Pugh score
Unknown 1 (0.7) for predicting the in-
NA 29 (20.0)
hospital mortality was
0.760 (95% CI: 0.644-
Portal hypertensive gastropathy, n (%)
0.876) and 0.719
Yes 2 (1.4)
(95% CI: 0.585-0.8-
No 114 (78.6)
53), respectively. Mo-
NA 29 (20) re recently, Reverter
Erosive gastritis, n (%) 1 (0.7) et al. analyzed 178
Child-Pugh class, n (%) patients with cirrhosis
A 46 (31.7) and acute esophageal
B 64 (44.1) variceal bleeding [15].
C 35 (24.1) AUROC for the MELD
Child-Pugh score 7.88±2.17 and Child-Pugh scores
MELD score 7.86±7.22 for predicting the
Abbreviations: RBC, red blood cell; Hb, hemoglobin; WBC, white blood cell; PLT, platelet; TBIL, 6-week mortality was
total bilirubin; DBIL, direct bilirubin; IBIL, indirect; ALB, albumin; ALT, alanine aminotransferase; 0.79 and 0.74, respec-
AST, aspartate aminotransferase; ALP, alkaline phosphatase; GGT, gamma-glutamyltranspepti- tively (P=0.2179). The-
dase; BUN, blood urea nitrogen; CR, creatinine; NA, not available; MELD, model for end stage se studies by Amitr-
liver disease; PT, prothrombin time; APTT, activated partial thromboplastin time; INR, interna-
tional nomalized ratio. ano, Cerqueira, and
Reverter suggested
the superiority of ME-
acute variceal bleeding, their superiority LD score over Child-Pugh score [13-15].
remained controversial among studies. However, it should be noted that the difference
Chalasani et al. collected 239 cirrhotic patients was not statistically significant. Orloff et al.
with acute variceal bleeding from 4 large aca- enrolled 211 consecutive patients with liver cir-
demic hospitals, and compared the perfor- rhosis and esophageal variceal bleeding after
mance of the two scoring systems in predicting endoscopic sclerotherapy or emergency porta-
the in-hospital and 1-year mortality rates [12]. caval shunt [16]. The investigators found that
The MELD score was highly predictive of both Child-Pugh score was similar to MELD score in
in-hospital (AUROC=0.82) and 1-year predicting the survival, recurrent encephalopa-
(AUROC=0.75) mortality rates. But its advan- thy, and rebleeding. Additionally, Child-Pugh
tages over Child-Pugh score were not signifi- score was superior to MELD score in predicting
cant. Amitrano et al. retrospectively analyzed the hospital readmissions and readmission
the 6-week and 3-month mortality of 172 cir- days.
rhotic patients with the first episode of oesoph-
ageal variceal bleeding after drug and endo- As for the cirrhotic patients with unstable UGIB
scopic therapy [13]. AUROC for the MELD and (heart rate > 100 beats/minute or systolic
Child-Pugh scores for predicting the 6-week blood pressure < 100 mmHg), Hsu et al. retro-

754 Int J Clin Exp Med 2015;8(1):751-757

 Child-Pugh versus MELD

Table 2. Treatment in 145 patients

Treatment Values
Transfusion (yes/no), n (%) 91/54 (62.8/37.2)
Transfusion of RBC unit 4.40±4.04
Drugs, n (%)
Somatostatin (yes/no) 144/1 (99.3/0.7)
Proton pump inhibitor (yes/no) 145/0 (100/0)
Endosopic therapy, n (%)
None 12 (8.3)
Ligation 54 (37.2)
Sclerotherapy 3 (2.1)
Tissue adhesive 22 (15.2)
Ligation + sclerotherapy 1 (0.7)
Ligation + tissue adhesive 23 (15.9)
Sclerotherapy + tissue adhesive 1 (0.7)
NA 29 (20)
Surgery, n (%) Figure 3. ROC analysis of MELD scores for predict-
None 142 (97.9) ing the in-hospital mortality of acute UGIB in liver
Shunt 0 (0) cirrhosis.
Splenectomy + devascularization 3 (2.1)
Sengstaken-blakemore tube, n (%) 0 scores did not have any significant discrimina-
Abbreviations: RBC, red blood cell; NA, not available. tive ability for predicting the mortality
(AUROC=0.527, 95% CI: 0.393-0.661, P=
0.709; AUROC=0.591, 95% CI: 0.465-0.717,
P=0.208) [17].

Our target population has the following fea-

Figure 2. ROC analysis of Child-Pugh scores for pre-
dicting the in-hospital mortality of acute UGIB in liver
cirrhosis.
tures. 1) All patients had a diagnosis of liver cir-
rhosis. 2) All patients presented with acute
UGIB. Indeed, at the emergency admission for
UGIB, especially massive haematemesis, not
all patients had the opportunity to undergo the
endoscopic examinations to identify the sourc-
es of bleeding. 3) Child-Pugh and MELD scores,
two most important scoring systems for the
prognosis of liver cirrhosis, were compared in
our cohort. 4) The in-hospital mortality of acute
UGIB was the only endpoint of our study. We
found that both scoring systems had good dis-
criminative abilities for the in-hospital mortality
of acute UGIB in liver cirrhosis, and that the
AUROC for MELD score might be slightly supe-
rior to that for Child-Pugh score, but the differ-
ence was not statistically significant between
them.

The potential limitations of our study should be

spectively analyzed the performance of
Glasgow-Blatchford, Rockall, and MELD scores.
MELD scores had a significant discriminative
ability for predicting the mortality (AUR-
OC=0.736, 95% CI: 0.629-0.842, P=0.001). By
comparison, Glasgow-Blatchford and Rockall
clarified. First, the comparisons of long-term
follow-up outcome between the two scoring
systems were lacking. Second, 20% of included
patients did not undergo the endoscopic exami-
nation. Thus, we did not strictly limit the source
of UGIB (variceal or non-variceal). Third, none of

755 Int J Clin Exp Med 2015;8(1):751-757

 Child-Pugh versus MELD
Figure 4. Comparison of the performance of Child-
Pugh and MELD scores for predicting the in-hospital
mortality of acute UGIB in liver cirrhosis.
patients underwent TIPS for acute UGIB.
Indeed, a recent randomized controlled trial
suggested that early TIPS should be more
effective for improving the survival of acute
variceal bleeding in high-risk cirrhotic patients
[18]. This consideration is also supported by a
meta-analysis [19]. Thus, the mortality would
be lower in our patients, if TIPS was employed.
In conclusion, the discriminative ability for pre-
dicting the in-hospital mortality of acute UGIB
in liver cirrhosis was similar between Child-
Pugh and MELD scores.
Disclosure of conflict of interest
None.
Address correspondence to: Dr. Xiaozhong Guo or
Xingshun Qi, Department of Gastroenterology,
General Hospital of Shenyang Military Area, 83
Wenhua Road, Shenyang 110840, China. Tel: 86-24-
28897603; Fax: 86-24-28851113; E-mail: guo_
xiao_zhong@126.com (XZG); xingshunqi@126.com
(XSQ)
References
[1] Pugh RN, Murray-Lyon IM, Dawson JL, Pietroni
MC, Williams R. Transection of the oesophagus
for bleeding oesophageal varices. Br J Surg
1973; 60: 646-649.
[2] Malinchoc M, Kamath PS, Gordon FD, Peine
CJ, Rank J, ter Borg PC. A model to predict poor
756 Int J Clin Exp Med 2015;8(1):751-757
survival in patients undergoing transjugular in-
trahepatic portosystemic shunts. Hepatology
2000; 31: 864-871.
[3] Kamath PS, Kim WR. The model for end-stage
liver disease (MELD). Hepatology 2007; 45:
797-805.
[4] Salerno F, Merli M, Cazzaniga M, Valeriano V,
Rossi P, Lovaria A, Meregaglia D, Nicolini A, Lu-
batti L, Riggio O. MELD score is better than
Child-Pugh score in predicting 3-month surviv-
al of patients undergoing transjugular intrahe-
patic portosystemic shunt. J Hepatol 2002;
36: 494-500.
[5] Schepke M, Roth F, Fimmers R, Brensing KA,
Sudhop T, Schild HH, Sauerbruch T. Compari-
son of MELD, Child-Pugh, and Emory model for
the prediction of survival in patients undergo-
ing transjugular intrahepatic portosystemic
shunting. Am J Gastroenterol 2003; 98: 1167-
1174.
[6] Cholongitas E, Papatheodoridis GV, Vangeli M,
Terreni N, Patch D, Burroughs AK. Systematic
review: The model for end-stage liver disease-
-should it replace Child-Pugh’s classification
for assessing prognosis in cirrhosis? Aliment
Pharmacol Ther 2005; 22: 1079-1089.
[7] Durand F, Valla D. Assessment of the progno-
sis of cirrhosis: Child-Pugh versus MELD. J
Hepatol 2005; Suppl 42: S100-107.
[8] Cremers I, Ribeiro S. Management of variceal
and nonvariceal upper gastrointestinal bleed-
ing in patients with cirrhosis. Therap Adv Gas-
troenterol 2014; 7: 206-216.
[9] D’Amico G, de Franchis R. Upper digestive
bleeding in cirrhosis. Post-therapeutic out-
come and prognostic indicators. Hepatology
2003; 38: 599-612.
[10] Augustin S, Muntaner L, Altamirano JT, Gonza-
lez A, Saperas E, Dot J, Abu-Suboh M, Armen-
gol JR, Malagelada JR, Esteban R, Guardia J,
Genescà J. Predicting early mortality after
acute variceal hemorrhage based on classifi-
cation and regression tree analysis. Clin Gas-
troenterol Hepatol 2009; 7: 1347-1354.
[11] de Franchis R. Revising consensus in portal hy-
pertension: report of the Baveno V consensus
workshop on methodology of diagnosis and
therapy in portal hypertension. J Hepatol 2010;
53: 762-768.
[12] Chalasani N, Kahi C, Francois F, Pinto A,
Marathe A, Bini EJ, Pandya P, Sitaraman S,
Shen J. Model for end-stage liver disease
(MELD) for predicting mortality in patients with
acute variceal bleeding. Hepatology 2002; 35:
1282-1284.
[13] Amitrano L, Guardascione MA, Bennato R,
Manguso F, Balzano A. MELD score and hepa-
tocellular carcinoma identify patients at differ-
ent risk of short-term mortality among cirrhot-
 Child-Pugh versus MELD
ics bleeding from esophageal varices. J
Hepatol 2005; 42: 820-825.
[14] Cerqueira RM, Andrade L, Correia MR, Fer-
nandes CD, Manso MC. Risk factors for in-hos-
pital mortality in cirrhotic patients with oesoph-
ageal variceal bleeding. Eur J Gastroenterol
Hepatol 2012; 24: 551-557.
[15] Reverter E, Tandon P, Augustin S, Turon F, Casu
S, Bastiampillai R, Keough A, Llop E, González
A, Seijo S, Berzigotti A, Ma M, Genescà J,
Bosch J, García-Pagán JC, Abraldes JG. A
MELD-based model to determine risk of mor-
tality among patients with acute variceal
bleeding. Gastroenterology 2014; 146: 412-
419, e413.
[16] Orloff MJ, Vaida F, Isenberg JI, Wheeler HO,
Haynes KS, Jinich-Brook H, Rapier R, Hye RJ,
Orloff SL. Child-Turcotte score versus MELD for
prognosis in a randomized controlled trial of
emergency treatment of bleeding esophageal
varices in cirrhosis. J Surg Res 2012; 178:
139-146.
757 Int J Clin Exp Med 2015;8(1):751-757
[17] Hsu SC, Chen CY, Weng YM, Chen SY, Lin CC,
Chen JC. Comparison of 3 scoring systems to
predict mortality from unstable upper gastroin-
testinal bleeding in cirrhotic patients. Am J
Emerg Med 2014; 32: 417-420.
[18] Garcia-Pagan JC, Caca K, Bureau C, Laleman
W, Appenrodt B, Luca A, Abraldes JG, Nevens F,
Vinel JP, Mössner J, Bosch J; Early TIPS (Tran-
sjugular Intrahepatic Portosystemic Shunt) Co-
operative Study Group. Early use of TIPS in
patients with cirrhosis and variceal bleeding. N
Engl J Med 2010; 362: 2370-2379.
[19] Qi X, Jia J, Bai M, Guo X, Su C, Garcia-Pagan JC,
Han G, Fan D. Transjugular Intrahepatic Porto-
systemic Shunt for Acute Variceal Bleeding: A
Meta-analysis. J Clin Gastroenterol 2014;
[Epub ahead of print].
Medicine, Health Care and Philosophy
https://doi.org/10.1007/s11019-018-9845-y

SHORT COMMUNICATION

Toward an accelerated adoption of data-driven findings in medicine

Research, skepticism, and the need to speed up public visibility of data-driven findings

Uri Kartoun1

© Springer Nature B.V. 2018

Abstract
To accelerate the adoption of a new method with a high potential to replace or extend an existing, presumably less accurate,
medical scoring system, evaluation should begin days after the new concept is presented publicly, not years or even decades
later. Metaphorically speaking, as chameleons capable of quickly changing colors to help their bodies adjust to changes in
temperature or light, health-care decision makers should be capable of more quickly evaluating new data-driven insights and
tools and should integrate the highest performing ones into national and international care systems. Doing so is essential,
because it will truly save the lives of many individuals.

Keywords  Clinical informatics · Prediction modeling · Electronic medical records · Machine-learning · Data-mining ·
Cirrhosis · Liver transplantation

Throughout history, skepticism has played an important role
in evaluating a variety of phenomena. In medicine, some
scientists have occasionally been dismissed as irrational only
to be proven right many years later. For instance, Galen, a
second-century philosopher and physician, believed that the
liver was the source of all veins and the principle organ for
blood production (ElMaghawry et al. 2014). Though most
of Galen’s writings were incorrect, people still held strong
to his beliefs even 1500 years later. Dr. William Harvey was
the first to describe blood circulation to the heart, brain, and
body in detail. In 1628, in his book, De Motu Cordis (On the
Motion of the Heart and Blood), describing the structure of
the heart and arteries, he posited for the first time that blood
passed through the heart, not the liver as previously believed.
Harvey’s findings were ridiculed, and many doctors in the
seventeenth century noted that they would “rather err with
Galen than proclaim the truth with Harvey.” (Bushak 2015).
Another example of skepticism in medicine concerns non-
alcoholic fatty liver disease (NAFLD). Until a few decades
ago, the scientific community was undecided about whether
* Uri Kartoun
uri.kartoun@ibm.com
1 patient’s rank on the organ allocation waiting list; notably,
Center for Computational Health, IBM Research,
Cambridge, MA, USA
NAFLD is actually a clinical condition. An NAFLD diag-
nosis has important health and clinical implications because
it is a risk factor for the development of diseases such as
type 2 diabetes mellitus and an independent risk factor for
cardiovascular-related mortality and all-cause mortality
(Musso et al. 2011; Byrne and Targher 2015). Nonalcoholic
steatohepatitis, the progressive form of NAFLD, can result
in cirrhosis and hepatocellular carcinoma and is estimated
to become the leading indication for liver transplant in the
United States by 2020 (Charlton 2008).
Recent remarkable advancements in computer hardware
and software and the growing accessibility of electronic
medical records (EMRs) have accelerated research on pre-
dicting patient outcomes. Such advances have allowed the
rapid development of massive-scale predictive models—
powerful resources to study disease complications at the
population level. Such models have proved highly useful
to discovering or confirming disease correlations, sub-cat-
egories of diseases, and adverse drug events. The model of
the end-stage liver disease (MELD) risk score, for instance,
is one of the most important and widely used risk predic-
tion scores in medicine. Unlike in the case of other scores,
a patient’s MELD score may indicate the likelihood of a
major clinical event for the patient. MELD determines the
since 2002, MELD has played a crucial role in determining

13
Vol.:(0123456789)
 U. Kartoun
which patient on a waiting list will be the next to receive a
liver transplant (Kamath and Kim 2007).
Combining the ability to store and rapidly process the
records of millions of individuals by accessing the reposi-
tories of Massachusetts General Hospital (MGH), Brigham
and Women’s Hospital (BWH), and the IBM Explorys Plat-
form using machine-learning algorithms has helped us cre-
ate a new and highly accurate score to predict short-term
mortality in cirrhosis patients (Kartoun et al. 2017). We
took an unbiased approach to the discovery of biomarkers.
In this approach, we filtered a large collection of medical
records through a feature-selection algorithm and identi-
fied a small set of variables that could serve as the most
efficient predictors for a given medical outcome. We used
the traditional supervised-learning paradigm to assess accu-
racy and applied standard statistical methods to assess the
validity of our approach. We realized that combining the
components of MELD with several easily accessible vari-
ables would enable us to construct a new score that would
be approximately 10% more accurate. We named our new
score MELD-Plus. MELD-Plus is an attempt to create a new
mortality prediction risk score in cirrhosis. Our unbiased
data-driven approach, which involves the use of an algo-
rithm to select predicting variables as well as the large and
independent databases used for validation, makes our score
a useful tool that could truly save lives. Furthermore, the
fact that MELD-Plus’s variables are available for any patient
(including total bilirubin, creatinine, albumin, INR, WBC,
sodium, total cholesterol, length of stay, and age) makes it
easy to calculate the patient’s mortality risk using Excel or
to deploy on any digital health repository.
Our preceding manuscript drafts, in which we outlined a
better scoring system than MELD, raised significant skepti-
cism from reviewers and editors. Although we were invited
to present our earlier findings at a medical informatics
conference (Kartoun et al. 2016), leading medical journals
repeatedly criticized our work. The criticism always had a
reasonable rationale, but our findings and the proposition for
an alternative score did not change throughout our resubmis-
sions and were, therefore, kept out of the public eye. Eventu-
ally we successfully published our study in October 2017 in
PLoS ONE, a peer-reviewed journal.
The main criticism of our initial manuscript was valid:
until 2016 we had access to only one source of data (MGH/
BWH), and our claim of generalizability was indeed weak.
Another criticism was that the interest in such scores might
be limited to individual clinicians who are making decisions.
This claim, however, may rule out the usefulness of any
other type of risk score as well. Another concern was that
our predictive model contained too many variables, reducing
its practicability in day-to-day use. Such criticism was valid
if powerful computers capable of instantaneously processing
tremendous collections of EMRs were not in such broad use,
13
as they are today. Future risk scores will likely be composed
of tens of thousands of patient characteristics and be calcu-
lated automatically as an integrated component of an EMR
system to provide real-time decision support to monitor a
disease or to prioritize organ transplant candidacy.
Finally, we faced criticism that several of the variables
that we used (all selected by a feature-selection algorithm)
were associated with cardiovascular risk rather than liver-
related mortality. Strikingly, researchers from the Cleveland
Clinic validated another of our liver-related studies in which
we strengthened the existing knowledge and discovered new
biomarkers regarding the interplay between cardiovascular
risk and liver disease. Both studies were published in The
American Journal of Gastroenterology (Corey et al. 2016;
Mehta et al. 2016). Medical publications that describe unbi-
ased approaches to feature selection for developing new
scores or to classifying diseases more accurately are rare. A
few, however, have been published, including, for instance,
a prediction model for 30-day readmission for heart fail-
ure patients (Kartoun et al. 2015) and models to classify
rheumatoid arthritis (Liao et al. 2010), Crohn’s disease, and
ulcerative colitis (Ananthakrishnan et al. 2013). Although
we were not criticized explicitly for favoring a data-driven
unbiased approach rather than relying on domain expertise,
it could have been our use of an approach not yet broadly
accepted that have raised further criticism.
Furthermore, our approach also relied on a new text-
processing method that we developed to accurately extract
concepts from clinical narrative notes. The method, text
nailing (TN), raised skepticism in reviewers of medical
informatics journals who claimed that TN “relies on simple
tricks to simplify the text,” and “leans heavily on human
annotation.” TN indeed may seem just like a trick of the
light at first glance, but it is actually a fairly sophisticated
method that finally caught the attention of more adventurous
reviewers and editors who ultimately accepted it for publica-
tion (Kartoun 2017a, b). We found TN to be highly accurate,
outperforming traditional machine-learning algorithms in
multiple scenarios, such as extracting family history of coro-
nary artery disease (Corey et al. 2016), classifying patients
with sleep disorders (Beam et al. 2017; Kartoun et al. 2018),
and improving the accuracy of the Framingham risk score
for patients with NAFLD (Simon et al. 2017).
As for the historical trajectory of adopting liver alloca-
tion scores, in 1998, the Committee on Organ Procurement
and Transplantation Policy of the Institute of Medicine (cur-
rently called “The National Academy of Medicine”) pub-
lished “The Final Rule,” calling for “…standardized medi-
cal criteria to be used to determine the status of a person’s
illness and when that person can be placed on a waiting
list” and further stating “…Minimum listing criteria for
including transplant candidates on the national list shall
be standardized and, to the extent possible, shall contain
Toward an accelerated adoption of data-driven findings in medicine
explicit thresholds for listing a patient and be expressed
through objective and measurable medical criteria” (Insti-
tute of Medicine 1999). Independently, scientists reported
in 2000 on a well-validated model and on the creation of a
new equation to calculate survival probabilities for patients
following a transjugular intrahepatic portosystemic shunt
placement (Malinchoc et al. 2000). In February 27, 2002,
this equation was selected to serve as the basis for the new
allocation policy (Freeman et al. 2002). The equation, form-
ing MELD, has become the standard by which priorities
are determined in donor liver allocation, and as expected,
implementation of MELD led to an immediate reduction in
liver transplant waiting list registrations for the first time in
the history of liver transplantation (with a 12% decrease in
2002) (Kamath and Kim 2007). In subsequent years, multi-
ple studies proposed that the incorporation of sodium into
the original MELD equation could significantly improve
prediction accuracy for liver disease. For instance, a study
published in the New England Journal of Medicine in 2008
estimated that using an extended version of MELD, one that
incorporated serum sodium levels, would save 90 lives in the
period from 2005 to 2006 (Kim et al. 2008). Additional stud-
ies supported the usefulness of sodium to improve prediction
performance for liver disease (Ruf et al. 2005; Londoño et al.
2007; Luca et al. 2007). The MELD-Na score, an equation
that incorporates sodium into MELD, was finally adopted in
2014 (Mulligan and Hirose 2014).
Why did it take many years to adopt MELD-Na, a score
that was created by using a data-driven approach, instead
of starting to use it, say, in 2008, right after multiple stud-
ies demonstrated the advantage of using sodium to improve
the prediction accuracy of MELD? The lives of hundreds
would have been saved if MELD-Na was in use starting in
2008 rather than in 2014. The reason for the delay was most
likely to let the scientific community assess and discuss
further the combination’s potential usefulness as well as its
drawbacks, a consideration undertaken by a large number
of independent investigators and through the use of patient
data captured at multiple health systems. Only after broad
scientific evidence had been accumulated, was the United
Network for Organ Sharing (UNOS) convinced to extend
MELD to MELD-Na. Furthermore, UNOS estimated that
MELD-Na was expected to save between 50 and 60 lives per
year (Mulligan and Hirose 2014), and relevant to MELD-
Plus, our experiments demonstrate that while MELD-Na
performed slightly better than MELD, MELD-Plus per-
formed significantly better than MELD (> 10% better) (Kar-
toun et al. 2017). Thus, MELD-Plus, if incorporated into
hospital systems, could save hundreds of patients every year
in the United States alone. Furthermore, as an encourag-
ing first step toward adoption, a very recent study reported
that MELD-Plus plays a predictive role in the occurrence of
post-liver transplantation acute kidney injury, proposing a
broader usefulness beyond mortality prediction (Tudoroiu
et al. 2018).
The adoption of MELD-Na would had been faster if the
scientific community had been able to publish convinc-
ing studies earlier to assess the contribution of sodium to
MELD. Organizations such as The American Medical Infor-
mation Association (AMIA) have encouraged universities as
well as commercial companies to form “Challenges,” such
as the de-identification and the smoking status challenges
(Uzuner et al. 2007, 2008). Such challenges have resulted
in a variety of high-impact papers that have significantly
enhanced the medical informatics subdomain, as well as the
entire health-care domain. If UNOS had worked more col-
laboratively with AMIA, as well as with The Institute of
Electrical and Electronics Engineers (IEEE)’s Engineering
in Medicine and Biology Society, new challenges could have
been formed, with titles such as “The MELD-Plus Chal-
lenge” or “The Liver Disease Challenge,” inviting investiga-
tors from all around the globe to assess current scores and
propose new scores that might even outperform MELD-Plus.
Additional associations, such as the Association for Comput-
ing Machinery (ACM), might be encouraged to be involved
in such efforts focused on computational assessments of
health. Such initiatives could help accelerate the adoption of
health-related data-driven findings, as these challenges are
expected to produce scientific papers faster and thus support
or rule out the usefulness of the newest findings.
In a desirable future scenario, UNOS may decide to
replace MELD (or its subsequent score, MELD-Na) with
MELD-Plus or even with more advanced futuristic scores
that may be developed by other researchers that incorpo-
rate, for instance, additional behavioral and genetic aspects.
Hypothetically, we can imagine a patient a decade from now
who is in need of a liver replacement. That patient might feel
encouraged if MELD-Plus was in use, determining more
accurately his or her rank on the waiting list. MELD-Plus
will not cure that patient, of course, but its ability to assess
the severity of a condition more precisely could mean that
the patient might wait 2 months less for a new liver than if
the original MELD was in use. MELD-Plus, therefore, could
save the patient’s life.
On the one hand, skeptics are often proven wrong as
science advances. For instance, it took years for the main-
stream scientific community to accept Harvey’s contri-
butions over Galen’s. On the other hand, skepticism in
medicine is essential, especially regarding questionable
treatments and methods and the potential effects of using
new medicines. Advances in medicine that have raised
significant skepticism include, for example, a human
head transplant operation proposed by neurosurgeon Dr.
Sergio Canavero (former director of the Turin Advanced
Neuromodulation Group, Italy) or a new approach to slow
the progression of Alzheimer’s disease proposed by Dr.
13
 U. Kartoun
Dale Bredesen (University of California, Los Angeles).
The development of new risk scores, by contrast, and
especially those that are based on components of similar
widely used scores, such as MELD, should not be inter-
preted as questionable and thus should be expected to
raise only minor levels of skepticism. Regardless of any
specific disease, when a new score is introduced publicly
in a peer-reviewed scientific journal, the scientific com-
munity would benefit from the availability of mechanisms
that could evaluate the scores more rapidly, considering
data derived from multiple health-care systems. Espe-
cially, official organizations that rely on the scores (e.g.,
the American Diabetes Association, the American Heart
Association, UNOS) would benefit from such mechanisms.
Combining the most advanced data-driven algorithms with
human expertise is expected to achieve a desirable increase
in knowledge, and this could potentially affect decisions to
either replace or extend current scoring methods. Incorpo-
rating data-driven scores adjusted by human expertise is
essential, especially within the context of medical ethics,
and will help in deciding which research findings may be
worth accelerating and what safeguards need to be pro-
vided. For instance, a better scoring system might result
in some patients being pushed up the queue for transplan-
tation, but it would result in other patients being pushed
down. Additional components coming into play regarding
medical scoring systems must consider the ability to assess
the true risk of learning algorithms (Kartoun 2018) and
the potential for sociological biases at the individual level,
as well as at the social level. An algorithm favoring a can-
didate for transplant based on his or her political views,
family status, or sexual preference are just a few examples
for such potential biases.
To accelerate the adoption of a new method with a high
potential to replace or extend an existing, presumably less
accurate, medical scoring system, evaluation should begin
days after the new concept is presented publicly, not years
or even decades later. Metaphorically speaking, as cha-
meleons capable of quickly changing colors to help their
bodies adjust to changes in temperature or light, health-
care decision makers should be capable of more quickly
evaluating new data-driven insights and tools and should
integrate the highest performing ones into national and
international care systems. Doing so is essential, because
it will truly save the lives of many individuals.
Funding  The author received honoraria and travel funding from The
American Association for the Study of Liver Diseases (October 2017).
Compliance with ethical standards  on Translational Science, San Francisco, CA.
Conflict of interest  The author has declared that no competing inter-
ests exist. The author confirms that the commercial affiliation with
13
IBM does not alter his adherence to all Medicine, Health Care and
Philosophy policies.
References
Ananthakrishnan, A. N., T. Cai, G. Savova, S. C. Cheng, P. Chen, R.
G. Perez, V. S. Gainer, S. N. Murphy, P. Szolovits, Z. Xia, S.
Shaw, S. Churchill, E. W. Karlson, I. Kohane, R. M. Plenge, and
K. P. Liao. 2013. Improving case definition of Crohn’s disease
and ulcerative colitis in electronic medical records using natural
language processing: A novel informatics approach. Inflammatory
Bowel Diseases 19 (7): 1411–1420.
Beam, A. L., U. Kartoun, J. K. Pai, A. K. Chatterjee, T. P. Fitzgerald,
S. Y. Shaw, and I. S. Kohane. 2017. Predictive modeling of phy-
sician-patient dynamics that influence sleep medication prescrip-
tions and clinical decision-making. Scientific Reports 7: 42282.
Bushak, L. 2015. Mad Scientist: 6 scientists who were dismissed as
crazy, only to be proven right years later. Medical Daily, Nov
18, 2015.
Byrne, C. D., and G. Targher. 2015. NAFLD: A multisystem disease.
Journal of Hepatology 62: S47–S64.
Charlton, M. 2008. Cirrhosis and liver failure in nonalcoholic fatty
liver disease: Molehill or mountain? Hepatology 47: 1431–1433.
Corey, K. E., U. Kartoun, H. Zheng, R. T. Chung, and S. Y. Shaw.
2016. Using an electronic medical records database to identify
nontraditional cardiovascular risk factors in nonalcoholic fatty
liver disease. The American Journal of Gastroenterology 111 (5):
671–676.
ElMaghawry, M., A. Zanatta, and F. Zampieri. 2014. The discovery of
pulmonary circulation: From Imhotep to William Harvey. Global
Cardiology Science and Practice 2014(2):103–116.
Freeman, R. B. Jr, R. H. Wiesner, A. Harper, S. V. McDiarmid, J. Lake,
E. Edwards, R. Merion, R. Wolfe, J. Turcotte, and L. Teperman.
2002. UNOS/OPTN Liver Disease Severity Score, UNOS/OPTN
Liver and Intestine, and UNOS/OPTN Pediatric Transplantation
Committees. The new liver allocation system: Moving toward
evidence-based transplantation policy. Liver Transplantation 8
(9): 851–858.
Institute of Medicine. 1999. Analysis of waiting times. In Organ pro-
curement and transplantation: Assessing current policies and the
potential impact of the DHHS final rule, ed. Committee on Organ
Procurement, and Transplantation Policy, 57–58. Washington,
D.C.: National Academy Press.
Kamath, P. S., and W. R. Kim. 2007. The model for end-stage liver
disease (MELD). Hepatology 45 (3): 797–805.
Kartoun, U. 2017a. Text nailing: An efficient human-in-the-loop text-
processing method 2017. ACM Interactions 24 (6): 44–49.
Kartoun, U. 2017b. Beyond brute force. Communications of the ACM
60 (10): 8–9.
Kartoun, U. 2018. Toward an equation that anticipates AI risks. Com-
munications of the ACM 61 (2): 8–9.
Kartoun, U., V. Kumar, A. Brettman, S. Yu, K. Liao, E. Karlson, A.
Ananthakrishnan, Z. Xia, V. Gainer, A. Cagan, S. Murphy, S.
Churchill, I. Kohane, P. Szolovits, T. Cai, and S. Y. Shaw. 2015. A
risk model for 30-day heart failure re-admission using electronic
medical records. 2015 Joint Summits on Translational Science,
San Francisco, CA.
Kartoun, U., K. Corey, H. Zheng, and S. Shaw. 2016. A prediction
model to assess mortality risk in cirrhosis. 2016 Joint Summits
Kartoun, U., K. Corey, T. Simon, H. Zheng, R. Aggarwal, K. Ng, and
S. Shaw. 2017. The MELD-Plus: A generalizable prediction risk
score in cirrhosis. PLoS ONE 12 (10): e0186301. https​://doi.
org/10.1371/journ​al.pone.01863​01. (eCollection 2017).
Toward an accelerated adoption of data-driven findings in medicine
Kartoun, U., R. Aggarwal, A. Beam, J. Pai, A. Chatterjee, T. Fitzgerald,
I. Kohane, and S. Shaw. 2018. Development of an algorithm to
identify patients with physician-documented insomnia. Scientific
Reports 8: 7862.
Kim, W. R., S. W. Biggins, W. K. Kremers, R. H. Wiesner, P. S.
Kamath, J. T. Benson, E. Edwards, and T. M. Therneau. 2008.
Hyponatremia and mortality among patients on the liver-trans-
plant waiting list. New England Journal of Medicine 359 (10):
1018–1026.
Liao, K. P., T. Cai, V. Gainer, S. Goryachev, Q. Zeng-treitler, S. Ray-
chaudhuri, P. Szolovits, S. Churchill, S. Murphy, I. Kohane, E.
W. Karlson, and R. M. Plenge. 2010. Electronic medical records
for discovery research in rheumatoid arthritis. Arthritis Care &
Research 62 (8): 1120–1127.
Londoño, M. C., A. Cárdenas, M. Guevara, L. Quintó, D. de Las Heras,
M. Navasa, A. Rimola, J. C. Garcia-Valdecasas, V. Arroyo, and P.
Ginès. 2007. MELD score and serum sodium in the prediction of
survival of patients with cirrhosis awaiting liver transplantation.
Gut 56 (9): 1283–1290.
Luca, A., B. Angermayr, G. Bertolini, F. Koenig, G. Vizzini, M. Ploner,
M. Peck-Radosavljevic, B. Gridelli, and J. Bosch. 2007. An inte-
grated MELD model including serum sodium and age improves
the prediction of early mortality in patients with cirrhosis. Liver
Transplantation 13 (8): 1174–1180.
Malinchoc, M., P. S. Kamath, F. D. Gordon, C. J. Peine, J. Rank, and
P. C. ter Borg. 2000. A model to predict poor survival in patients
undergoing transjugular intrahepatic portosystemic shunts. Hepa-
tology 31 (4): 864–871.
Mehta, N., T. Singh, R. Lopez, and N. Alkhouri. 2016. The heart age
is increased in patients with nonalcoholic fatty liver disease and
correlates with fibrosis and hepatocyte ballooning. The American
Journal of Gastroenterology 111 (12): 1853–1854.
Mulligan, D. C., and R. Hirose. 2014. OPTN/UNOS, Liver and Intes-
tinal Organ Transplantation Committee. Report to the Board of
Directors. June 23–24, 2014, Richmond, Virginia.
Musso, G., R. Gambino, M. Cassader, and G. Pagano. 2011. Meta-
analysis: Natural history of non-alcoholic fatty liver disease
(NAFLD) and diagnostic accuracy of noninvasive tests for liver
disease severity. Annals of Medicine 43: 617–649.
Ruf, A. E., W. K. Kremers, L. L. Chavez, V. I. Descalzi, L. G. Podesta,
and F. G. Villamil. 2005. Addition of serum sodium into the
MELD score predicts waiting list mortality better than MELD
alone. Liver Transplantation 11 (3): 336–343.
Simon, T. G., U. Kartoun, H. Zheng, A. T. Chan, R. T. Chung, S. Shaw,
and K. E. Corey. 2017. MELD-Na score predicts incident major
cardiovascular events, in patients with nonalcoholic fatty liver
disease. Hepatology Communications 1 (5): 429–438.
Tudoroiu, M. I., G. Constantin, L. Pâslaru, S. Iacob, C. Gheorghe, I.
Popescu, D. Tomescu, and L. S. Gheorghe. 2018. The combina-
tion of serum cystatin C, urinary kidney injury molecule-1 and
MELD-Plus score predicts early acute kidney injury after liver
transplantation. Surgery, Gastroenterology and Oncology 23 (2):
121–126.
Uzuner, O., Y. Luo, and P. Szolovits. 2007. Evaluating the state-of-
the-art in automatic de-identification. Journal of the American
Medical Informatics Association 14 (5): 550–563.
Uzuner, O., I. Goldstein, Y. Luo, and I. Kohane. 2008. Identifying
patient smoking status from medical discharge records. Journal
of the American Medical Informatics Association 15 (1): 14–24.
13