What Matters For Liver Transplant in Severe ACLF - Medpage Today Dec 28 2018-Merged
What Matters For Liver Transplant in Severe ACLF - Medpage Today Dec 28 2018-Merged
.
Therefore, the investigators sought to address the following questions:
.
What is the waitlist mortality for different grades of ACLF?
.
Is there a correlation between the aforementioned variable and increasing MELD scores?
What are the characteristics of recipients with ACLF-3 (failure of ≥3 organs) that are associated with greater post-
transplant survival?
From 2005 to 2016, Vinay Sundaram, MD, MSc of Cedars-Sinai Medical Center in Los
Angeles, and colleagues, identified 100,594 patients on the waitlist and 50,552 liver
transplant recipients from the United Network for Organ Sharing (UNOS) Database.
Patients with ACLF were identified using the European Association for the Study of the
Liver-chronic liver failure (EASL-CLIF) Consortium definition. According to a review by
Hernaez and colleagues, the "EASL-CLIF Acute-on-Chronic Liver Failure in Cirrhosis
(CANONIC) study is the most comprehensive registry to understand outcomes on
hospitalized patients with cirrhosis."
p p
Based on the number of organ failures (liver, renal, coagulation, cerebral, respiratory, and
circulatory), ACLF can be stratified into ACLF-1 (one renal failure or one non-renal organ
failure if associated to renal dysfunction and/or cerebral dysfunction), ACLF-2 (2 organ
failures), and ACLF-3.
A serum sodium (Na) level parameter was also taken into account to assess the
contribution of significant hyponatremia to the MELD score (MELD-Na).
In the current study, the number of eligible patients with ACLF-3, ACLF-2, ACLF-1, and no
ACLF were 6,381 (12.6%), 7,513 (14.9%), 7,375 (14.6%), and 29,283 (57.9%), respectively. Other
demographic criteria of patients with ACLF-3 included a low proportion of hepatitis-C-
virus-induced cirrhosis, a high percentage of patients with ALD (34.9%) and a high mean
MELD-Na score (37.4). Patients with ACLF-3 also received younger donor organs (mean age
38.7), many of whom had died from head trauma (38%). Additional descriptors included
fewer organs received from diabetic donors (8.8%), more organs received from donors
dying of head trauma (38.0%), and the smallest percentage of organs received from high-
risk donors with a donor risk index of 1.7 or more (22.9%).
Predefined 1-year waitlist mortality, 28-day, and 90-day survival outcomes were
compared among patients with different ACLF grades at listing across a range of MELD
score (estimation of the relative disease severity and likely survival of patients awaiting
liver transplantation) categories (i.e., <25, 25-29, 30-34, and ≥35). The primary outcome
was patient survival at 1 year.
One-year survival after transplant was lowest among patients with ACLF-3 (81.8%)
compared with other groups (88.1%-91.9%, P<0.001). Those who survived for 1 year had a
lower prevalence of mechanical ventilation use (33.1% vs 49.0%, P<0.001) and circulatory
failure (49.8% vs 60.6%, P<0.001).
The following variables were found to be independently associated with survival at 1 year
afte t ansplantation Mechanical entilation at li e t ansplantation (HR 1 49 95% CI 1 22
after transplantation: Mechanical ventilation at liver transplantation (HR 1.49, 95% CI 1.22-
1.84), donor risk index above 1.7 (HR 1.22, 95% CI 1.09-1.35), and liver transplantation within
30 days of listing (HR 0.89, 95% CI 0.81-0.98).
The investigators concluded that certain patients with ACLF-3 have poor outcomes
irrespective of their MELD-Na scores. Compared to other ACLF groups, patients with
ACLF-3 were more likely to be removed from the waitlist or die, irrespective of their
MELD-Na scores. This was particularly true for patients with ACLF-3 and MELD-Na score
<25 (43.8% at 28 days). For this group, liver transplantation increases the odds of survival,
especially if performed within 30 days of placement on the waitlist. The risk of death is
increased with the use of mechanical ventilation or marginal organs.
Because this was a retrospective analysis, there is the possibility of misclassification with
respect to diagnosis of patients with ACLF and the presence of organ failure. It is also
possible that bacterial infection rates could have been overlooked during the
investigation. Additionally, mechanical ventilation was used as a metric for respiratory
failure; however, it is possible that this method could have been employed for another
reason (e.g., for airway protection due to a patient's altered mental status). On the other
hand, patients with significant lung injury may not have been intubated at the time of
liver transplantation.
Source Reference: Gastroenterology online, Dec. 18, 2018; DOI: 10.1053/j.gastro.2018.12.007
A liver transplant is often the only way to treat patients with ACLF. But the role of
extrahepatic organ failures in prioritizing allocation for transplants is still a subject of
ongoing investigation. Moreover, there is a scarcity of data regarding actual outcomes of
liver transplant recipients.
Patients with ACLF-3 had the poorest outcomes (highest mortality and removal from the
waiting list) in this study irrespective of MELD score category. The investigators noted
that "nearly 44% of patients with ACLF-3 with a MELD score <25 would die or be
removed within 28 days of listing."
Caution should be exercised with intubated patients, as mechanical ventilation at liver
transplantation had the strongest link to lower post-transplant survival.
It is worth noting that the investigators assessed other predictive risk scores beyond the
MELD score and organ failure (i.e., Karnofsky's Performance Status Scale (KPS) and the
futility risk score). Neither score was associated with 1-year post-transplant mortality.
Possible reasons for these results include the fact that the futility risk score was validated
for five-year patient survival, whereas the authors evaluated shorter-term outcomes. In
addition, the proportion of patients with ACLF-3 and a high KPS score was very low
(2.0%). More studies are therefore needed to determine whether frailty, as measured by
KPS or other assessments, can provide extra information regarding post-liver
transplantation outcomes for patients with ACLF-3.
Overall, the findings supported the investigators' initial hypothesis that some patients
with ACLF-3 would be at high risk for premature death irrespective of MELD or MELD-Na
score. Thus, they would be candidates for timely intervention. Evidence from the
literature supported this notion. For instance, a recent registry analysis showed a 1-year
post transplant survival >80% for patients with ACLF-3. Similar trends were observed in
prior retrospective analyses of small numbers of patients with ACLF-3 (i.e., a 1-year post-
transplant survival of more than 80%).
Sundaram told MedPage Today. "Many of these patients will have significantly improved
survival if they undergo liver transplant, particularly if performed early after listing," he
added, calling for additional research into the prioritization of patients with severe ACLF
on waiting lists.
Diana Swift wrote the original story for MedPage Today.
Reviewed by Robert Jasmer, MD Associate Clinical Professor of
Medicine, University of California, San Francisco and Dorothy
Caputo, MA, BSN, RN, Nurse Planner
Primary Source
Gastroenterology
Source Reference: Sundaram V, et al "Factors associated with survival of patients with severe acute or chronic liver failure before and after liver transplantation"
Gastroenterol 2018; DOI: 10.1053/j.gastro.2018.12.007.
RESEARCH ARTICLE
Funding: The study was funded in part by grants improved the performance of MELD-Na by 11.4% (0.78 [95% CI, 0.75–0.81] versus 0.70
from the NIH K23 DK099422 (KEC), and NIH U54 [95% CI, 0.66–0.73]). In the MGH/BWH approximate 1-year model, MELD-Plus improved
LM008748 (SYS). NIH did not play a role in the
study design, data collection and analysis, decision
the performance of MELD-Na by 8.3% (0.78 [95% CI, 0.76–0.79] versus 0.72 [95% CI,
to publish, or preparation of the manuscript and 0.71–0.73]). Performance improvement was similar when the novel MELD-Plus risk score
only provided financial support in the form of was applied to an independent database; when considering 24,042 cirrhosis-related admis-
authors’ salaries and/or research materials. NIH
sions, MELD-Plus improved the performance of MELD-Na by 16.9% (0.69 [95% CI, 0.69–
provided support in the form of salaries for authors
UK, KEC, and SYS. IBM neither provided authors 0.70] versus 0.59 [95% CI, 0.58–0.60]).
[UK, KN] salaries related to the study nor played
any role in the study design, data collection and
analysis, decision to publish, or preparation of the
Conclusions
manuscript. We developed a new risk score, MELD-Plus that accurately stratifies the short-term mortal-
Competing interests: The authors have declared ity of patients with established cirrhosis, following a hospital admission. Our findings demon-
that no competing interests exist. UK and KN strate that using a small set of easily accessible structured variables can help identify novel
confirm that the commercial affiliation with IBM predictors of outcomes in cirrhosis patients and improve the performance of widely used tra-
does not alter their adherence to all PLOS ONE
policies on sharing data and materials.
ditional risk scores.
The aim of the present study was to develop a risk score to predict mortality following a cir-
rhosis-related admission. We demonstrated that a score composed of a small set of easily
accessible clinical variables improves the prediction performance of both the MELD and
MELD-Na scores. We further demonstrated the generalizability of our model through inde-
pendent validation in a large EMR-based data source.
Methods
Study population
We analyzed a previously defined cohort of 314,292 patients at increased risk for metabolic
disease who were admitted to Massachusetts General Hospital (MGH) or Brigham and Wom-
en’s Hospital (BWH) between 1992 and 2010 [13]. We identified an admission as cirrhosis-
related when the keyword “cirrhosis” was present in the discharge summary of the admission
and we observed at least one ICD-9 code (571.2, 571.5, or 571.6 as in [18]) within the 30 days
preceding the discharge date, including during the admission. This identification method was
validated by a physician (Dr. Kathleen Corey) chart review.
We excluded elective admissions if they included at least one diagnosis or procedure code
for liver biopsy, radiofrequency ablation, transarterial chemoembolization, hepatic resection,
or liver transplant. We included only patients 18 years of age or older at the time of the admis-
sion, and we tracked the records of all patients for 90 days after their discharge. We deter-
mined mortality through linkage to the social security master death index.
Prediction modeling
To predict mortality within 90 days, we developed a model that included a large set of struc-
tured variables extracted from the EMRs. In addition to variables available during the period
of admission, we considered variables available for the period of 12 months preceding the dis-
charge date (see Table 1).
The variables included demographics (e.g., gender, ethnicity, marital status), laboratory
measurements (e.g., albumin, sodium), and medications (e.g., anticoagulants, lipid lowering
agents). For laboratory variables, we used the most recent values found during admission
(when no value was found during admission, we considered the preceding 12 months). Typi-
cally, common laboratory measurements were available during the admission (as seen in
Table 1). We determined comorbidities from the number of diagnosis codes within the 12
months prior to the discharge date, and we determined medication count by recording the
number of prescriptions within the 12 months preceding the discharge date.
Additional variables included body mass index, NAFLD fibrosis score (Eq 1), and the
MELD score (Eq 2). Missing values were imputed with the mean of the available data for each
variable. We randomly selected two thirds of the admissions to serve as a derivation set,
whereas the remaining one third served as a validation set. A complete list of the variables we
used is available in S1 Table, and all diagnoses and procedure definitions used in this study are
available in S2 Table.
NAFLD Fibrosis Score ¼ 1:675 þ 0:037 Age þ 0:094 BMI
þ1:13 IFG=Diabetes ðyes ¼ 1; no ¼ 0Þ ð1Þ
þ0:99 AST=ALT ratio 0:013 Platelet 0:66 Albumin
Table 1. Baseline characteristics. All values extracted during the 12 months preceding discharge date. For
laboratory variables, values are the most recent. Comorbidity calculations count the number of diagnosis
codes. Prevalence calculations consider admissions with at least one measurement for laboratories and at
least one diagnosis code for comorbidities.
Variable and category Cirrhosis-related admissions (n = 4,781)
Age (years); Mean (SD) 60.0 (13.7)
Gender (%)
Male 64.4
Female 35.6
Ethnicity (%)
Caucasian 77.4
African American 6.8
Other 2.3
Unknown 13.5
Insurance Type (could be 1 types per patient) (%)
Medicaid 5.6
Medicare 60.0
Other 98.6
BMI (kg/m2); Mean (SD) 28.7 (8.2)
Laboratory values; Mean (SD) / Prevalence (%)
Sodium (mmol/l) 136.5 (5.8) / 99.7
eGFR (ml/min/1.73m2) 59.6 (33.7) / 31.9
WBC (th/cumm) 6.7 (3.8) / 99.8
Platelets (th/cumm) 141.3 (100.0) / 99.8
Prothrombin time (INR) 1.5 (0.5) / 91.3
Albumin (g/dl) 2.9 (0.7) / 98.7
Total Bilirubin (mg/dl) 2.5 (4.3) / 98.7
Transaminase SGOT (u/l) 60.2 (76.6) / 98.8
Transaminase SGPT (u/l) 36.1 (38.2) / 96.7
GGT (u/l) 249.4 (346.6) / 6.8
MELD score 14.21 (6.1) / 84.8
NAFLD Fibrosis score 1.70 (2.1) / 13.7
Comorbidities; Mean (SD) / Prevalence (%)
Variceal hemorrhage / Gastrointestinal bleed 0.8 (2.1) / 24.2
Spontaneous bacterial peritonitis 0.1 (0.6) / 3.4
Hepatocellular carcinoma 0.7 (5.5) / 4.3
Hepatorenal syndrome 0.1 (0.5) / 3.2
Hepatic encephalopathy 1.4 (3.6) / 28.6
Ascites 2.4 (5.7) / 37.9
Renal failure 1.8 (8.4) / 13.1
Cerebrovascular disease 0.5 (2.4) / 10.8
Congestive heart failure 2.7 (7.9) / 36.7
Hypertension 3.1 (7.2) / 58.9
Acute myocardial infarction 0.5 (2.1) / 13.4
Ischemic heart disease 0.3 (1.2) / 12.4
Peripheral vascular disease 0.5 (2.2) / 13.0
Diabetes 5.9 (10.3) / 57.0
https://doi.org/10.1371/journal.pone.0186301.t001
To select the most informative variables, we applied feature selection on the derivation set.
We used logistic regression with the adaptive least absolute shrinkage and selection operator
(LASSO) algorithm [19] because it is considered an efficient algorithm for parsimoniously
ranking variables in clinical predictive modeling [20, 21]. We considered all variables that
were statistically significantly different from a univariate analysis (P < 0.05) as in [22]. The
generalized linear model (GLM) equations used to calculate prediction risk at MGH/BWH
and at the independent EMR source are presented in Eqs 3 and 4.
L¼
þ11:794383
þ2:076192 Log10ð1 þ Total BilirubinÞ
þ2:494291 Log10ð1 þ CreatinineÞ
6:049540 Log10ð1 þ AlbuminÞ
þ2:525904 Log10ð1 þ INRÞ
ð3Þ
þ1:911856 Log10ð1 þ WBCÞ
þ0:015411 Length of stay
ðnumber of nights the patient spent in the hospital during the cirrhosis related admissionÞ
þ0:041047 Age ðyearsÞ
6:625270 Log10ð1 þ SodiumÞ
1:445666 Log10ð1 þ Total CholesterolÞ
expðLÞ
MELD Plus ¼ pð90 day mortalityÞ ¼ ð4Þ
1 þ expðLÞ
To calculate 95% confidence intervals, we applied the bootstrap procedure with 1,000
replicates. We calculated the area under the receiver operating characteristic curves
(AUROC) to measure the model’s accuracy in the validation set. Additionally, we evaluated
for overfitting by comparing the AUROC in the validation set to an average AUROC value
for 100 permutations of randomly selected derivation and validation sets (each including
two thirds and one third of the derivation set’s cirrhosis-related admissions, respectively).
We compared categorical variables using a chi-squared test, and we compared the differ-
ences in the means of continuous variables using a t-test or Wilcoxon rank sum test, as
appropriate. We further compared the differences in standard deviations by using an F-test.
All statistical tests were two-sided, with Bonferroni corrections for the 68 comparisons, and
the adjusted P value was 7.410−4 for each comparison. We performed all programming
using the R statistical language [23].
Independent validation
We were granted access to a data source of 18,345,793 individuals, pooled from multiple differ-
ent health-care systems with EMRs (“The IBM Explorys Network”) [24]. The data were stan-
dardized and normalized using common ontologies, searchable through a HIPAA-enabled,
de-identified cloud-computing platform. Patients were seen in multiple health-care systems
between 2010 and 2015, with a combination of data from clinical EMRs, outgoing health-care
system bills, and adjudicated payor claims.
We first identified all cirrhosis-related admissions in this database, and then we extracted
values for the selected variables of our MGH/BWH 90-day model. We further deployed the
GLM equations on the independent source (Eqs 3 and 4). Missing values were imputed based
on the mean values of the MGH/BWH 4,781 cirrhosis-related admissions: total bilirubin
(2.486604493), creatinine (1.375274633), albumin (2.888940902), INR (1.499619615), WBC
(6.673836966), sodium (136.5401552), and total cholesterol (133.6246914).
Because death dates were not available for patients in the IBM Explorys Network, we used
the year of death to determine the outcome. We were able then only to use a 1-year estimated
death for this population (e.g., for a patient discharged on October 13, 2010, we could only
determine if the patient died either in 2010 or 2011, or survived after that). To compare the
performance of the IBM Explorys approximate one-year prediction model, we used the origi-
nal 314,292-patient population at MGH/BWH and applied the same approximate one-year
mortality outcome identification method. We calculated AUROCs for MELD, MELD-Na (Eq
5), and for our risk score (Eqs 3 and 4).
MELD Na ¼ MELD þ 1:59 ð135 NaÞ ð5Þ
The institutional review board of Partners HealthCare and IBM approved this study and all
its methods, including the EMR cohort assembly, data extraction, and analyses.
Results
Univariate analysis
We identified a total of 4,781 admissions as cirrhosis-related, of which 778 resulted in death
within 90 days of the discharge date (16.3%). In a sample of 50 randomly selected patients,
64% were admitted primarily for cirrhosis, for instance, due to the presence of ascites or spon-
taneous bacterial peritonitis, and the rest had the comorbidity of cirrhosis but were admitted
primarily for different reasons such as heart failure or chronic obstructive pulmonary disease
(COPD). Individuals who died within the 90-day period after discharge were older in compari-
son with those who survived (64.1 years versus 59.2 years, P = 4.2210−17); however, the two
populations did not differ by gender (65.0% male versus 64.0% female, P = 1.0) or ethnicity
(77.0% Caucasian for both, P = 1.0).
We calculated event ratios by dividing the values ascertained for the two populations (e.g.,
the mean MELD scores were 18.5 and 13.3 for admissions that resulted in death and survival,
respectively, yielding a ratio of 1.4, P = 4.4510−65). Individuals who died within the 90-day
period after discharge had higher ratios of liver-related comorbidities than those who survived,
and these comorbidities included hepatorenal syndrome (ratio = 5.1, P = 4.6010−21), hepatocel-
lular carcinoma (ratio = 5.0, P = 1.4010−16), and ascites (ratio = 2.1, P = 3.5510−24). Laboratory
measurements also significantly differentiated the two populations. For instance, albumin was
lower in those who died within the 90-day period (2.60 g/dl versus 2.95 g/dl, P = 5.6410−35),
and the total bilirubin (4.87 mg/dl versus 2.02 mg/dl, P = 1.8210−41), INR (1.70 versus 1.46,
P = 5.6310−30), and creatinine (1.80 mg/dl versus 1.29 mg/dl, P = 1.2410−36) were higher in
those who died within the 90-day period.
No difference was found in the prevalence of COPD, cerebrovascular disease, diabetes, cor-
onary artery disease, peripheral vascular disease, pneumonia, or sleep apnea between the popu-
lations. The complete list of variables, indicating the differences between the surviving and the
deceased populations, is presented in S3 Table.
Fig 1. AUROCs using differing variable combinations in a 90-day mortality prediction model at MGH/
BWH.
https://doi.org/10.1371/journal.pone.0186301.g001
laboratory values and other trivial variables (e.g., age, length of stay), we decided to follow the
model that comprised the 9 readily available clinical variables. To evaluate the contribution of
the MELD score to the 90-day mortality prediction, we evaluated the performance of MELD
and MELD-Na scores alone. Considering the 4,781 admissions, using the MELD score alone
to predict the 90-day mortality resulted in an AUROC value of 0.69. An additional model
using the MELD-Na score alone yielded an AUROC value of 0.70.
Each of the MELD-Na components were associated with an increased mortality, including
INR (OR, 1.58; 95% CI, 1.30–1.96), creatinine (OR, 1.25; 95% CI, 1.16–1.34), total bilirubin
(OR, 1.11; 95% CI, 1.08–1.14), and sodium (OR, 0.97; 95% CI, 0.95–0.99). Other laboratory
measurements associated with mortality included WBC (OR, 1.10; 95% CI, 1.07–1.13), total
cholesterol (OR, 0.996; 95% CI, 0.993–0.999), and albumin (OR, 0.45; 95% CI, 0.37–0.52).
Additional predictors included age at time of the admission (OR, 1.04; 95% CI, 1.03–1.05) and
length of stay (OR, 1.02; 95% CI, 1.005–1.03).
Because total cholesterol and hospital length of stay are typically not uniform factors
across different hospitals and may vary in different countries, we evaluated an additional
model that included only 7 of the 9 variables. This yielded an AUROC of 0.77 and
resulted in the following associations with increased mortality: INR (OR, 1.66; 95% CI,
1.38–2.05), creatinine (OR, 1.25; 95% CI, 1.17–1.35), total bilirubin (OR, 1.11; 95% CI,
1.08–1.14), sodium (OR, 0.97; 95% CI, 0.95–0.98), WBC (OR, 1.10; 95% CI, 1.07–1.13),
albumin (OR, 0.43; 95% CI, 0.36–0.51), and age (OR, 1.04; 95% CI, 1.03–1.05). We pres-
ent the GLM equations used to calculate prediction performance at MGH/BWH in Eqs 6
and 7.
L¼
þ8:53499496
þ2:06503238 Log10ð1 þ Total BilirubinÞ
þ2:59679650 Log10ð1 þ CreatinineÞ
6:34990436 Log10ð1 þ AlbuminÞ ð6Þ
þ2:99724802 Log10ð1 þ INRÞ
þ1:92811726 Log10ð1 þ WBCÞ
þ0:04070442 Age ðyearsÞ
6:47834101 Log10ð1 þ SodiumÞ
Generalization evaluation
Applying our 9-variable risk score (the MELD-Plus score) demonstrated an improvement
over MELD and MELD-Na scores in all prediction models, as shown in Fig 3. On the MGH/
BWH 90-day model, MELD-Plus improved the performance of MELD-Na by 11.4% (0.78
[95% CI, 0.75–0.81] versus 0.70 [95% CI, 0.66–0.73]). On the MGH/BWH approximate 1-year
model, MELD-Plus improved the performance of MELD-Na by 8.3% (0.78 [95% CI, 0.76–
0.79] versus 0.72 [95% CI, 0.71–0.73]). On the IBM Explorys Network model used for external
validation, MELD-Plus improved the performance of MELD-Na by 16.9% (0.69 [95% CI,
0.69–0.70] versus 0.59 [95% CI, 0.58–0.60]).
It is notable that the performance of MELD-Plus on the IBM Explorys data was lower in
comparison with both MGH/BWH models (0.69 versus 0.78). Consistent with MELD-Plus,
the performance of MELD and MELD-Na were also much lower on the IBM Explorys data in
comparison with MGH/BWH. A potential reason for this is that the IBM Explorys Network
population was relatively healthier. Patients in the IBM Explorys network had lower severity of
liver disease in comparison with the corresponding MGH/BWH 1-year prediction model
Fig 2. Predicted versus observed 90-day mortality within each risk quintile. (A) Entire cohort of 4,781
cirrhosis-related admissions. (B) Derivation set of 3,187 cirrhosis-related admissions. (C) Validation set of
1,594 cirrhosis-related admissions.
https://doi.org/10.1371/journal.pone.0186301.g002
(mean MELD: 9.4 versus 16.8; P < 0.0001, mean MELD-Na: 11.4 versus 18.1; P < 0.0001).
There may be other differences in the data or populations in the independent systems; the
Partners HealthCare Research Patient Data Registry collected the MGH/BWH data, whereas
dozens of distinct data aggregation mechanisms collected the data for the IBM Explorys Net-
work. Furthermore, the variability in the levels of prediction performance might be influenced
by the variability in the data; prediction performance might be higher when there is more vari-
ability in the data source (i.e., the population comprising patients with a broad spectrum of
Fig 3. Prediction performance across different cirrhosis populations. (A) MGH/BWH 90-day mortality
(4,781 cirrhosis-related admissions). (B) The IBM Explorys Network approximate 1-year mortality (24,042
cirrhosis-related admissions). (C) MGH/BWH approximate 1-year mortality (4,680 cirrhosis-related
admissions).
https://doi.org/10.1371/journal.pone.0186301.g003
levels of cirrhosis severity). In the other direction, when the data is more uniform (e.g., most
patients have just been diagnosed with cirrhosis for the first time, and only a minority suffers
from an advanced cirrhosis), then prediction accuracy is lower. This hypothesis was confirmed
because the IBM Explorys network had a statistically significant lower standard deviation of
severity of liver disease in comparison with the MGH/BWH 1-year population (STD MELD:
1.8 versus 8.2; P < 0.0001, STD MELD-Na: 3.6 versus 8.2; P < 0.0001).
Discussion
In this study, we used accessible EMR variables to develop a highly accurate, predictive model
of 90-day post-discharge mortality in individuals with cirrhosis. We identified 9 variables that
accurately predicted 90-day mortality with an AUROC of 0.78. Our risk score improved the
performance of MELD and MELD-Na scores in multiple, independent patient populations,
and this also held true in a large external validation patient cohort. Furthermore, our model’s
calculated 90-day mortality risk was highly correlated with the observed mortality rate across
all five risk quintiles. In particular, the model’s performance on the highest-risk quintile (the
calculated and observed 90-day mortality was 31.6% and 31.2%, respectively) suggests that
high-risk patients can be accurately identified. An additional model that included only 7 of the
9 variables and excluded length of stay and total cholesterol yielded an AUROC of 0.77 [95%
CI, 0.74–0.80]. Although the 7-variable model demonstrated improved identification ability
compared to MELD or MELD-Na, the improved prediction performance achieved by includ-
ing total cholesterol in MELD-Plus suggests that it may be beneficial for cholesterol labs to be
routinely collected in cirrhosis admission order sets.
The MELD score has been used extensively to predict patient outcomes, mortality, and
readmission rates in individuals with cirrhosis [25, 26, 4]. Furthermore, although MELD-Na
[17] was superior to MELD, the MELD-Plus score yielded improved levels of discrimination
consistently in all prediction models, with AUROCs that significantly outperformed the tradi-
tional scores. These findings suggest that new types of cirrhosis-related risk indexes utilizing
novel risk indicators may improve prognostication in this high-risk population.
MELD-Plus includes all MELD-Na’s components, as well as additional variables (albumin,
total cholesterol, WBC, age, and length of stay). It is logical that a predication model that has all
the MELD-Na model variables and additional ones would perform better, as was observed by
MELD-Plus. Not only that, but many of the variables have physiological plausibility for inclu-
sion in a prediction model. Decreased albumin correlated with worse outcomes in our model,
which may be the result of decreased albumin marking decreased liver function in cirrhosis
patients [25, 5]. Increasing age and length of hospital stay helped predict worse outcomes as well
as could be expected. Along with that, higher WBC was correlated with a worse prognosis,
potentially indicating poorer patient status (e.g., infection) at time of score calculation. Although
patients may have multiple WBC measurements during admission, our model is both internally
and externally valid because it uses the most recent WBC lab value. We chose the most recent
WBC during model development because the last available set of labs is more reflective of the
current health of patients than older measurements. Surprisingly, increased total cholesterol pre-
dicted a more favorable prognosis. Although unintuitive at first, this aligns with previous reports
that claim cholesterol levels become less of a risk factor or even an inverse risk factor for mortal-
ity because serious diseases may lower cholesterol soon before death occurs [27].
Although our study describes analyses of retrospective medical databases, the proposed
score could be used to identify patients that are at a high-risk of mortality in real time and thus
may inform risk-stratification and therapeutic decision-making. In a desirable scenario, our
score could be calculated automatically as an integrated component of an EMR system; the
clinician would see a risk score (probability) or a risk quantile (highest, lowest, or in between)
associated with the discharged patient, and this could be used to guide outpatient monitoring
strategies. With further validation, the MELD-Plus score could also be used longitudinally in
outpatients to monitor disease progression and/or responses to therapy.
Our study has limitations. First, it is a retrospective analysis limited to two academic, ter-
tiary-care hospitals. Even though we validated our model on a large external patient cohort,
subsequent studies must further assess the validity of our model in the external population and
consider different age ranges, coding systems, and data-collection methods. Second, the cir-
rhosis populations may vary at different centers—for example, alcohol use might significantly
vary between patients residing in the Boston area versus patients residing in other states [28].
Furthermore, although MGH and BWH are urban care facilities, the high prevalence of rural
populations at the IBM Explorys Network might affect prediction performance. Third,
although mortality was recorded, either through linkage to the social security master death
index as in the MGH/BWH models or through using EMR or billing/claims in the IBM Exp-
lorys model, such death indications may under-represent the true mortality rates. To minimize
this potential under-representation, we considered only patients who survived the study fol-
low-up. All patients had EMR data entries (such as laboratory measurements) after the study
follow-up, indicating survival, or had a recorded indication of death during the study follow-
up, with no EMR data entries found afterward.
Another limitation of MELD-Plus is that it did not specifically consider which procedures
patients underwent during the cirrhosis-related admissions. Furthermore, all the patients con-
sidered in our models survived the admission, but neither MGH/BWH’s nor IBM’s databases
contained information on post-discharge cause of death. To further assess MELD-Plus’s appli-
cability in clinical practice, future analyses should consider subgroups of patients to determine
linkages between invasive inpatient procedures and causes of mortality. Regardless of this limi-
tation, however, our MELD-Plus displayed validity in predicting overall mortality, which is
clinically applicable, because it provides clinicians with information on populations of patients
who need more intense or closer care.
Although we excluded elective admissions for liver biopsy, radiofrequency ablation, trans-
arterial chemoembolization, hepatic resection, or liver transplant, these criteria might exclude
patients with early and intermediate hepatocellular carcinoma (HCC), but not patients with
advanced HCC who underwent medical treatments only. Liver cancer can lead to early mortal-
ity, even in patients with mild liver cirrhosis, and, as such, our exclusion criteria may reduce
the applicability of our risk score when applying it to patients with more advanced HCC. Fur-
thermore, because we excluded admissions associated with a liver transplant, mortality risk
may decrease after a cirrhosis-related admission if patients successfully underwent a transplant
in a preceding admission.
Another limitation of our study is algorithmic. The adaptive LASSO method identified 9
predictors and left out variables that may also be correlated with predicting death. Feature
selection algorithms are known to be blind to the clinical importance of variables, and when
highly correlated predictors are identified, the algorithm randomly selects one. On the one
hand, important variables such as ascites, hepatocellular carcinoma, and diuretic medications
were not selected as predictors. On the other hand, the feature selection algorithm assures that
a minimal set of covariates produce a high level of prediction accuracy. Furthermore, we con-
ducted our model performance evaluation on a held-out data set not used for training.
Although a prediction model’s error usually decreases when more variables are included, this
is not always the case. This is true when performance is evaluated on the training set (due to
overfitting) but not the case when performance is evaluated on a held-out test dataset, as was
used across all our models.
Supporting information
S1 Table. Summary of variables.
(DOCX)
S2 Table. Billing codes used to define conditions.
(DOCX)
S3 Table. Comparison of variables in patients who died vs. survived 90-days after dis-
charge.
(DOCX)
S4 Table. Comparison between the highest-risk (1st) and the lowest-risk (5th) quintiles.
(DOCX)
Acknowledgments
We acknowledge Chin Hur MD MPH (Director, GI Health Outcomes Research, GI Unit, Mas-
sachusetts General Hospital / Harvard Medical School) for his critical review of this manuscript.
Author Contributions
Conceptualization: Uri Kartoun, Kathleen E. Corey, Stanley Y. Shaw.
Data curation: Uri Kartoun, Rahul Aggarwal.
Formal analysis: Uri Kartoun, Rahul Aggarwal, Kenney Ng, Stanley Y. Shaw.
Funding acquisition: Kathleen E. Corey, Stanley Y. Shaw.
Investigation: Stanley Y. Shaw.
Methodology: Uri Kartoun, Kathleen E. Corey, Kenney Ng, Stanley Y. Shaw.
Resources: Stanley Y. Shaw.
Software: Uri Kartoun, Kenney Ng.
Supervision: Kathleen E. Corey, Stanley Y. Shaw.
Validation: Uri Kartoun, Rahul Aggarwal, Kenney Ng.
Writing – original draft: Uri Kartoun.
Writing – review & editing: Kathleen E. Corey, Tracey G. Simon, Hui Zheng, Rahul Aggar-
wal, Kenney Ng, Stanley Y. Shaw.
References
1. Kim WR, Brown RS Jr, Terrault NA, El-Serag H. Burden of liver disease in the United States: summary
of a workshop. Hepatology 2002; 36:227–42. https://doi.org/10.1053/jhep.2002.34734 PMID:
12085369
2. Blachier M, Leleu H, Peck-Radosavljevic M, Valla DC, Roudot-Thoraval F. The burden of liver disease
in Europe: a review of available epidemiological data. J Hepatol 2013; 58(3):593–608. https://doi.org/
10.1016/j.jhep.2012.12.005 PMID: 23419824
3. Rakoski MO, McCammon RJ, Piette JD, Iwashyna TJ, Marrero JA, Lok AS, et al. Burden of cirrhosis on
older Americans and their families: analysis of the health and retirement study. Hepatology 2012; 55
(1):184–91. https://doi.org/10.1002/hep.24616 PMID: 21858847
4. Johnson KB, Campbell EJ, Chi H, Zheng H, King LY, Wu Y, et al. Advanced disease, diuretic use, and
marital status predict hospital admissions in an ambulatory cirrhosis cohort. Dig Dis Sci 2014; 59
(1):174–82. https://doi.org/10.1007/s10620-013-2832-5 PMID: 23990035
5. Younossi ZM, Henry L, Stepanova M. A new comorbidity model for predicting mortality in patients with
cirrhosis: does it work? Gastroenterology 2014; 146(1):19–24. https://doi.org/10.1053/j.gastro.2013.11.
026 PMID: 24287302
6. Kamath PS, Kim WR. The model for end-stage liver disease (MELD). Hepatology 2007; 45(3):797–805.
https://doi.org/10.1002/hep.21563 PMID: 17326206
7. Gotthardt D, Weiss KH, Baumgärtner M, Zahn A, Stremmel W, Schmidt J, et al. Limitations of the
MELD score in predicting mortality or need for removal from waiting list in patients awaiting liver trans-
plantation. BMC Gastroenterol 2009; 9(72):1–7.
8. Huo TI, Lin HC, Wu JC, Hou MC, Lee FY, Lee PC, et al. Limitation of the model for end-stage liver dis-
ease for outcome prediction in patients with cirrhosis-related complications. Clin Transplant 2006; 20
(2):188–94. https://doi.org/10.1111/j.1399-0012.2005.00463.x PMID: 16640525
9. Montano-Loza AJ, Duarte-Rojo A, Meza-Junco J, Baracos VE, Sawyer MB, Pang JX, et al. Inclusion of
Sarcopenia within MELD (MELD-Sarcopenia) and the prediction of mortality in patients with cirrhosis.
Clin Transl Gastroenterol 2015; 6(e102):1–8.
10. Kohane IS, Drazen JM, Campion EW. A glimpse of the next 100 years in medicine. N Engl J Med 2012;
367:2538–39. https://doi.org/10.1056/NEJMe1213371 PMID: 23268669
11. Liao KP, Kurreeman F, Li G, Duclos G, Murphy S, Guzman R, et al. Associations of autoantibodies,
autoimmune risk alleles, and clinical diagnoses from the electronic medical records in rheumatoid arthri-
tis cases and non-rheumatoid arthritis controls. Arthritis Rheum 2013; 65(3):571–81. https://doi.org/10.
1002/art.37801 PMID: 23233247
12. Doshi-Velez F, Ge Y, Kohane I. Comorbidity clusters in autism spectrum disorders: an electronic health
record time-series analysis. Pediatrics 2014; 133(1):e54–e63. https://doi.org/10.1542/peds.2013-0819
PMID: 24323995
13. Liao KP, Cai T, Savova GK, Murphy SN, Karlson EW, Ananthakrishnan AN, et al. Development of phe-
notype algorithms using electronic medical records and incorporating natural language processing.
BMJ 2015; 350:h1885. https://doi.org/10.1136/bmj.h1885 PMID: 25911572
14. Beam AL, Kartoun U, Pai JK, Chatterjee AK, Fitzgerald TP, Shaw SY, Kohane IS. Predictive modeling
of physician-patient dynamics that influence sleep medication prescriptions and clinical decision-mak-
ing. Sci Rep 2017; 7(42282); https://doi.org/10.1038/srep42282 PMID: 28181568
15. Malinchoc M, Kamath PS, Gordon FD, Peine CJ, Rank J, ter Borg PC. A model to predict poor survival
in patients undergoing transjugular intrahepatic portosystemic shunts. Hepatology 2000; 31(4):864–71.
https://doi.org/10.1053/he.2000.5852 PMID: 10733541
16. Kamath PS, Wiesner RH, Malinchoc M, Kremers W, Therneau TM, Kosberg CL, et al. A model to pre-
dict survival in patients with end-stage liver disease. Hepatology 2001; 33(2):464–70. https://doi.org/10.
1053/jhep.2001.22172 PMID: 11172350
17. Mulligan DC, Hirose R. OPTN / UNOS, Liver and Intestinal Organ Transplantation Committee. Report
to the Board of Directors. June 23–24, 2014. Richmond, Virginia.
18. Myers RP, Quan H, Hubbard JN, Shaheen AA, Kaplan GG. Predicting in-hospital mortality in patients
with cirrhosis: results differ across risk adjustment methods. Hepatology 2009; 49(2):568–77. https://
doi.org/10.1002/hep.22676 PMID: 19085957
19. Zou H. The adaptive lasso and its oracle properties. J Am Stat Assoc 2006; 101:1418–29.
20. Liao KP, Cai T, Gainer V, Goryachev S, Zeng-treitler Q, Raychaudhuri S, et al. Electronic medical rec-
ords for discovery research in rheumatoid arthritis. Arthritis Care Res (Hoboken) 2010; 62(8):1120–7.
21. Ananthakrishnan AN, Cai T, Savova G, Cheng SC, Chen P, Perez RG, et al. Improving case definition
of Crohn’s disease and ulcerative colitis in electronic medical records using natural language process-
ing: a novel informatics approach. Inflamm Bowel Dis 2013; 19(7):1411–20. https://doi.org/10.1097/
MIB.0b013e31828133fd PMID: 23567779
22. Corey KE, Kartoun U, Zheng H, Shaw SY. Development and validation of an algorithm to identify nonal-
coholic fatty liver disease in the electronic medical record. Dig Dis Sci 2016; 61(3):913–9. https://doi.
org/10.1007/s10620-015-3952-x PMID: 26537487
23. Ihaka R. & Gentleman R. R: a language for data analysis and graphics. J Comput Graph Stat 1996;
5:299–314.
24. The IBM Explorys Platform / Solution Brief. IBM Watson Health. IBM Corporation 2016.
25. Singal AG, Rahimi RS, Clark C, Ma Y, Cuthbert JA, Rockey DC, et al. An automated model using elec-
tronic medical record data identifies patients with cirrhosis at high risk for readmission. Clin Gastroen-
terol Hepatol 2013; 11(10):1335–41. https://doi.org/10.1016/j.cgh.2013.03.022 PMID: 23591286
26. Volk ML, Tocco RS, Bazick J, Rakoski MO, Lok AS. Hospital readmissions among patients with decom-
pensated cirrhosis. Am J Gastroenterol 2012; 107(2):247–52. https://doi.org/10.1038/ajg.2011.314
PMID: 21931378
27. Ravnskov U, Diamond DM, Hama R, Hamazaki T, Hammarskjöld B, Hynes N, Kendrick M, Langsjoen
PH, Malhotra A, Mascitelli L, McCully KS, Ogushi Y, Okuyama H, Rosch PJ, Schersten T, Sultan S,
Sundberg R. Lack of an association or an inverse association between low-density-lipoprotein choles-
terol and mortality in the elderly: a systematic review. BMJ Open 2016; 6(6):e010401. https://doi.org/10.
1136/bmjopen-2015-010401 PMID: 27292972
28. Sacks JJ, Gonzales KR, Bouchery EE, Tomedi LE, Brewer RD. 2010 National and State Costs of
Excessive Alcohol Consumption. Am J Prev Med 2015; 49(5):e73–9. https://doi.org/10.1016/j.amepre.
2015.05.031 PMID: 26477807