Risk factors for mortality in patients with alcoholic hepatitis and assessment of prognostic models: A population-based study

The poor outcomes and high mortality risk associated with alcoholic hepatitis (AH) have prompted the search for predictive risk factors that could guide management and treatment, and facilitate risk stratification in clinical trials. Presently, several prognostic models for AH are available, all of which have helped physicians decide which therapies to initiate or to assess a given patient’s response to treatment. This retrospective study, conducted at a Canadian tertiary care centre, aimed to identify prognostic risk factors for hospitalized AH patients and to validate the predictive utility of well-known risk scores.

Abstract

BACKGROUND:

Severe alcoholic hepatitis (AH) is associated with a substantial risk for short-term mortality.

OBJECTIVES:

To identify prognostic factors and validate well-known prognostic models in a Canadian population of patients hospitalized for AH.

METHODS:

In the present retrospective study, patients hospitalized for AH in Calgary, Alberta, between January 2008 and August 2012 were included. Stepwise logistic regression models identified independent risk factors for 90-day mortality, and the discrimination of prognostic models (Model for End-stage Liver Disease [MELD] and Maddrey discriminant function [DF]) were examined using areas under the ROC curves.

RESULTS:

A total of 122 patients with AH were hospitalized during the study period; the median age was 49 years (interquartile range [IQR] 42 to 55 years) and 60% were men. Median MELD score and Maddrey DF on admission were 21 (IQR 18 to 24) and 45 (IQR 26 to 62), respectively. Seventy-three percent of patients received corticosteroids and/or pentoxifylline, and the 90-day mortality was 17%. Independent predictors of mortality included older age, female sex, international normalized ratio, MELD score and Maddrey DF (all P<0.05). For discrimination of 90-day mortality, the areas under the ROC curves of the prognostic models (MELD 0.64; Maddrey DF 0.68) were similar (P>0.05). At optimal cut-offs of ≥22 for MELD score and ≥37 for Maddrey DF, both models excluded death with high certainty (negative predictive values 90% and 96%, respectively).

CONCLUSIONS:

In patients hospitalized for AH, well-known prognostic models can be used to predict 90-day mortality, particularly to identify patients with a low risk for death.

Alcoholic hepatitis (AH) is a clinical syndrome characterized by hepatic dysfunction in the setting of heavy alcohol intake. Rapid onset of jaundice is a cardinal manifestation of AH; other common signs include fever, ascites, muscle wasting and hepatic encephalopathy (1). AH is often complicated by hepatorenal syndrome and infection, both of which significantly increase mortality (2–4). Untreated patients with severe AH have a particularly poor prognosis, with one-month mortality rates ranging from 30% to 50% (5,6). In patients with a mild presentation, the risk of progression to cirrhosis is 50%; this risk is highest in patients who continue to abuse alcohol (7).

With such a high risk for mortality, the identification of risk factors for poor outcomes due to AH can be beneficial to help guide patient management and for risk stratification in clinical trials. Several prognostic models have been developed for use in patients with AH (eg, the Maddrey discriminant function [DF], the Glasgow Alcoholic Hepatitis Score [GAHS] and the Lille model) or applied after derivation in other conditions (eg, the Model for End-stage Liver Disease [MELD]). In general, these models are used to help clinicians decide whether to initiate specific therapy (eg, with corticosteroids) or, in the case of the Lille model, to assess a patient’s response to corticosteroid treatment (6,8–10). The Maddrey DF, which has been in use the longest, is based on the prothrombin time (PT) and serum bilirubin concentration. A Maddrey DF >32 is typically used as the threshold for starting corticosteroid therapy (6). The MELD score, which was originally developed to predict mortality following transjugular intrahepatic portosystemic shunt insertion and widely used for liver allograft allocation, includes measurements of serum bilirubin, creatinine and the international normalized ratio (INR) (11). In one study, MELD was superior for the prediction of in-hospital mortality in patients with AH (12). Specifically, an admission MELD score ≥18 had similar sensitivity to a Maddrey DF ≥32 (85% versus 83%), but was more specific (84% versus 60%) and had a higher area under the ROC curve (AUROC; 0.89 versus 0.81) (12). The GAHS is a newer score that includes age, white blood cell (WBC) count, urea nitrogen, INR and bilirubin to identify patients at risk for mortality in the absence of treatment (9). One study has shown a higher 84-day survival rate of 59% in patients with a Maddrey DF ≥32 and GAHS ≥9 who were treated with corticosteroids compared with no treatment (38% survival) (13). Finally, the Lille model considers the change in serum bilirubin after a seven-day course of corticosteroids, among other factors, to assess treatment responsiveness. A score >0.45 indicates a lack of treatment response and the need to consider alternative therapies (10).

The objectives of the current study were to identify prognostic factors in patients hospitalized for AH and to validate these well-known prognostic scoring models in a Canadian, population-based setting. Our data fill a knowledge gap regarding the outcomes of patients with AH in Canada and confirm the external validity of these prognostic models for risk stratification.

METHODS

Study population and definition of AH

In the present retrospective study, the Discharge Abstract Database was queried to identify adults (≥18 years of age) hospitalized in Calgary, Alberta with a diagnosis of AH (International Classification of Diseases-10 diagnosis code, K70.1 [14]) between January 2008 and August 2012. Only the first hospitalization was considered among patients with multiple admissions for AH. The Discharge Abstract Database contains diagnosis, procedure and mortality information on all discharges from the three adult, acute care hospitals in Calgary (catchment population approximately 1.5 million individuals). Chart validation studies have shown rates of agreement >95% for demographics and 75% to 96% for most responsible diagnosis codes (15). This database has been used to examine the epidemiology (16,17), outcomes (18–22) and coding accuracy (16,23) of multiple medical conditions from a population-based perspective.

The diagnosis of AH was confirmed via a review of the paper and electronic medical records for each individual by one of three trained physicians (MB, ER, JP). AH was confirmed if all of the following criteria at hospital admission were satisfied: heavy alcohol consumption (>196 g/week or >56 g in any day among men, and >98 g/week or >42 g in any day among women) (24); elevated serum aspartate aminotransferase and/or alanine aminotransferase (ALT) concentration, but <300 IU/mL (to exclude other disorders associated with acute hepatic dysfunction including acetaminophen toxicity); serum bilirubin >34 μmol/L; elevated INR; and exclusion of other causes of acute hepatic dysfunction (eg, drug hepatotoxicity, autoimmune hepatitis, ischemic hepatitis, etc). A previous study showed excellent inter-rater agreement among these reviewers for this definition of AH (kappa 0.86; P<0.001). Although the reference standard for the diagnosis of AH includes liver biopsy (25), this procedure is not part of the routine clinical management of patients with suspected AH in Calgary hospitals. Self-reported average daily alcohol intake was recorded in a hierarchical fashion from medical records. Specifically, data were first taken from consultation notes by addictions specialists where available, followed by the admission history and, otherwise, from patient progress notes.

The Conjoint Health Research Ethics Board at the University of Calgary approved the study protocol.

Outcomes and predictor variables

The primary outcome measure was 90-day mortality as determined by a review of paper and electronic medical records. All survivors had ≥90 days of follow-up. Secondary outcomes included in-hospital mortality and length of hospital stay. The following clinical and laboratory characteristics were examined as potential predictors of 90-day mortality: age, sex, year and site of hospitalization, AH treatment (corticosteroids, pentoxifylline, both or neither), WBC count, hemoglobin, platelets, liver biochemistry (serum ALT, aspartate aminotransferase, alkaline phosphatase, gamma-glutamyl transferase [GGT] and total bilirubin), albumin, INR and serum creatinine at admission. Clinical evidence of hepatic encephalopathy or ascites, including that detectable only by abdominal imaging, were also recorded. Where data permitted, prognostic models MELD and Maddrey DF were calculated (Appendix 1). An early change in bilirubin level, defined as a reduction in serum bilirubin concentration from baseline to day 7, was recorded. Patients discharged alive from hospital within seven days who had a drop in bilirubin on day 3 were assumed to have an early change in bilirubin level. To identify comorbid conditions, administrative data from the Discharge Abstract Database were queried to identify the Elixhauser comorbidities, which have been associated with mortality in patients with hepatic (26,27) and nonhepatic disorders (28). Liver diseases and alcohol abuse were excluded from these analyses.

Statistical analyses

Patient characteristics and clinical data were descriptively summarized and are reported as medians (interquartile range [IQR]) and proportions. Between-groups comparisons were made using Fisher’s exact and χ² tests for categorical variables, and the Wilcoxon rank-sum test for continuous variables. Independent predictors of 90-day mortality (as outlined above) were identified using a series of three, stepwise-forward multivariate logistic regression models that included MELD (model 1), Maddrey DF and creatinine (model 2), and the individual components of MELD (INR, bilirubin and creatinine; model 3). Each model also included age, sex and variables that were significant (P<0.05) in univariate analyses. Variables included within specific prognostic scores were not included simultaneously in multivariate models including those scores (eg, PT or INR, which are included in MELD and Maddrey DF). A significance level of P<0.05 was used for retention of variables in the stepwise models.

Discrimination of the prognostic models (MELD and Maddrey DF) for 90-day mortality was determined using AUROCs (29). We also derived novel prognostic scores based on the logistic regression analyses for models 1, 2 and 3 and assessed the discrimination of these models. Discrimination refers to a model’s ability to correctly distinguish between two outcomes; in this case, death or survival at 90 days. An AUROC of 1.0 indicates perfect discrimination, whereas an AUROC of 0.5 is equivalent to chance. AUROCs were compared using the method of DeLong et al (30). Sensitivity, specificity, accuracy, and positive (PPV) and negative predictive values (NPV) were calculated for both MELD and Maddrey DF at optimal cut-offs defined by the maximal sum of sensitivity and specificity. In addition, the performance at other model cut-offs previously reported in the literature was examined. All analyses were performed using Stata version 11.0 (StataCorp, USA). Two-sided P<0.05 was considered to be statistically significant.

RESULTS

Patient characteristics

A total of 228 patients were hospitalized in Calgary between January 2008 and August 2012 with a diagnosis code for AH in the Discharge Abstract Database. In 122 patients (54%), the diagnosis was confirmed based on a review of medical records. The characteristics of the 122 patients comprising the study cohort are outlined in Table 1. The median age was 49 years (IQR 42 to 55 years) and 60% were male. A similar proportion of patients were hospitalized at each of the three adult hospitals in Calgary (P=0.27). Twenty-seven percent of patients were untreated, whereas 42% received pentoxifylline, 16% received corticosteroids, and 15% received both pentoxifylline and corticosteroids. Patients who received therapy had greater liver disease severity as indicated by the Maddrey DF (50 versus 24; P<0.001) and MELD scores (23 versus 17; P<0.001) compared with untreated patients. Median length of hospital stay was 10 days (IQR 5 to 21 days) and in-hospital mortality was 7.4% (n=9).

TABLE 1.

Characteristics of the study population

	Entire cohort (n=122)	Mortality at 90 days (n=21)	Survival at 90 days (n=101)	P
Age, years	49 (42–55)	54 (48–59)	48 (40–55)	0.012
Male sex	73 (60)	9 (43)	64 (63)	0.092
Length of stay	10 (5–21)	18 (7–27)	9 (5–19)	0.088
Admission hospital
1	51 (42)	6 (29)	45 (44)	0.273
2	36 (29)	9 (42)	27 (27)
3	35 (29)	6 (29)	29 (29)
Year
2008	27 (22)	2 (10)	25 (24)	0.440
2009	18 (15)	2 (10)	16 (16)
2010	21 (17)	4 (19)	17 (17)
2011	26 (21)	6 (28)	20 (20)
2012	30 (25)	7 (33)	23 (23)
Treatment
None	33 (27)	4 (19)	29 (29)	0.188
Pentoxifylline only	51 (42)	11 (52)	40 (39)
Corticosteroids only	20 (16)	1 (5)	19 (19)
Pentoxifylline + corticosteroids	18 (15)	5 (24)	13 (13)
Admission laboratory values
Bilirubin, μmol/L	175 (106–272)	205 (115–265)	175 (106–273)	0.857
International normalized ratio	1.6 (1.3–1.9)	1.8 (1.6–2.3)	1.5 (1.3–1.8)	0.002
Creatinine, mmol/L	57 (46–78)	78 (53–89)	54 (46–71)	0.028
White blood cells, ×109/L	9.0 (5.9–13.5)	13.0 (8.8–16.7)	8.3 (5.7–13.3)	0.027
Hemoglobin, g/L	113 (97–133)	106 (88–133)	114 (98–132)	0.256
Platelets, ×109/L*	120 (68–184)	162 (99–187)	107 (62–182)	0.038
Albumin, g/L*	23 (20–26)	21 (18–25)	23 (21–26)	0.196
Alanine aminotransferase, U/L	50 (32–76)	55 (32–68)	48 (32–82)	0.981
Alkaline phosphatase, U/L*	234 (168–331)	232 (188–297)	236 (167–333)	0.699
Gamma-glutamyl transferase, U/L*	555 (300–955)	410 (228–588)	592 (317–1038)	0.044
Early change in bilirubin†	50 (50)	6 (32)	44 (54)	0.125
Prognostic models
Maddrey discriminant function	45 (26–62)	58 (45–82)	42 (25–60)	0.009
Model for End-stage Liver Disease	21 (18–24)	23 (21–27)	21 (17–24)	0.038
Severe alcoholic hepatitis‡	97 (80)	21 (100)	76 (75)	0.007
Complications
Hepatic encephalopathy	58 (48)	13 (62)	45 (45)	0.159
Ascites	87 (71)	17 (81)	70 (69)	0.427

Data presented as median (interquartile range) or n (%).

^*Data available for 121 patients for platelets, 103 patients for albumin, 121 patients for alkaline phosphatase, and 119 for gamma-glutamyl transferase;

^†Decreased serum bilirubin level on day 7. Patients admitted for <7 days who had a drop in bilirubin on day 3 and were discharged alive were assumed to have an early change in bilirubin level (data available for 100 patients);

^‡Defined as Maddrey discriminant function ≥32 and/or the presence of hepatic encephalopathy

Predictors of 90-day mortality

Overall, 21 patients (17%) died within 90 days of hospital admission. Patients who died were significantly older than survivors (median age, 54 versus 48 years; P=0.01), but length of stay, admission hospital, year of admission and the type of treatment did not differ significantly between groups (Table 1). Patients who died had significantly lower median GGT and higher INR, creatinine, WBC count and platelets (all P<0.05). However, alkaline phosphatase, bilirubin, albumin and hemoglobin did not differ compared with survivors. Patients who died within 90 days had a higher prevalence of hepatic encephalopathy (62% versus 45%; P=0.16) and ascites (81% versus 69%; P=0.43); however, these differences were not statistically significant. Comorbid conditions, as defined by the Elixhauser comorbidity index, were generally of low prevalence (0% to 20%) and not associated with 90-day mortality (Appendix 2).The median Maddrey DF (58 versus 42; P=0.009), and MELD (23 versus 21; P=0.04) scores were significantly higher among patients who died within 90 days compared with survivors (Table 1; Figure 1).

Figure 1).

Distributions of the Model for End-stage Liver Disease (MELD) score and Maddrey discriminant function (DF) according to mortality at 90 days following hospitalization for alcoholic hepatitis. Patients who died within 90 days had higher MELD (P=0.04) and Maddrey DF (P=0.009) compared with survivors. The horizontal line inside each box represents the median, and the lower and upper borders of the box represent the 25th and 75th percentiles, respectively. Vertical lines from the ends of the boxes encompass the adjacent values. Outliers are denoted as closed circles

Results of the three stepwise multivariate analyses evaluating independent predictors of 90-day mortality are shown in Table 2. Older age was an independent predictor of mortality in all three models, with ORs of 1.08 to 1.09 per year. Male sex was associated with a lower risk of mortality in models 1 (OR 0.30; 95% CI 0.11 to 0.87) and 2 (OR 0.34; 95% CI 0.12 to 0.98), and was of borderline significance in model 3 (OR 0.35; [95% CI 0.12 to 1.00]; P=0.051). MELD and Maddrey DF were both significant in their respective models (models 1 and 2). In model 3, which included the components of MELD instead of the MELD and Maddrey DF scores, only INR was significantly associated with mortality (OR 4.07; 95% CI 1.45 to 11.4). Other laboratory values including the serum bilirubin and creatinine concentrations were not statistically significant (Table 2).

TABLE 2.

Independent predictors of 90-day mortality in patients hospitalized for alcoholic hepatitis

Predictor	Model 1*		Model 2*		Model 3*
	OR (95% CI)	P	OR (95% CI)	P	OR (95% CI)	P
Age, per year	1.09 (1.02–1.15)	0.006	1.09 (1.03–1.16)	0.004	1.08 (1.02–1.14)	0.01
Male	0.30 (0.11–0.87)	0.027	0.34 (0.12–0.98)	0.045	0.35 (0.12–1.00)	0.051
White blood cells	NS		NS		NS
Platelets	NS		NS		NS
Gamma-glutamyl transferase	NS		NS		NS
Model for End-stage Liver Disease	1.10 (1.01–1.21)	0.033	–		–
Maddrey discriminant function	–		1.02 (1.00–1.04)	0.013	–
Creatinine	–		NS		NS
International normalized ratio	–		–		4.07 (1.45–11.4)	0.008
Bilirubin	–		–		NS

^*As defined in the Methods section. NS Not significant

Based on the lower risk of mortality observed among men versus women, an exploratory analysis of patient characteristics according to sex was conducted (Appendix 3). Men had lower median platelet counts and higher serum creatinine, hemoglobin, ALT and GGT concentrations. However, there were no significant differences between men and women with respect to age, treatment and disease severity as assessed by the prognostic models or the presence of ascites and hepatic encephalopathy.

Discrimination of MELD and Maddrey DF for 90-day mortality

The AUROCs of the MELD and Maddrey DF scores for the prediction of 90-day mortality are shown in Table 3. Discrimination was moderate with AUROCs of 0.64 (95% CI 0.51 to 0.78) for MELD and 0.68 (95% CI 0.56 to 0.81) for Maddrey DF. Models derived in the present study cohort showed higher AUROCs, ranging from 0.74 (95% CI 0.63 to 0.85) for model 1 to 0.77 (95% CI 0.66 to 0.88) for model 3. However, pair-wise comparisons of the AUROCs between each model showed no statistically significant differences between any of the models (data not shown).

TABLE 3.

Discrimination of the prognostic models for 90-day mortality in patients hospitalized for alcoholic hepatitis

Prognostic model	AUROC (95% CI)	Standard error	P*
MELD (n=122)	0.64 (0.51–0.78)	0.07	–
Maddrey DF (n=122)	0.68 (0.56–0.81)	0.06	0.44
Model 1 (age, sex, MELD)	0.74 (0.63–0.85)	0.06	0.11
Model 2 (age, sex, Maddrey DF)	0.75 (0.64–0.86)	0.06	0.16
Model 3 (age, sex, INR)	0.77 (0.66–0.88)	0.06	0.11

^*Versus Model for End-stage Liver Disease (MELD). AUROC Area under the ROC curve; DF Discriminant function; INR International normalized ratio

Table 4 includes the operating characteristics of MELD and Maddrey DF for the prediction of 90-day mortality at various cut-offs. At an optimal cut-off of ≥22, MELD was 67% (95% CI 43% to 85%) sensitive and 59% (95% CI 49% to 69%) specific. While the PPV was suboptimal at 25% (95% CI 15% to 39%), the NPV was high at 90% (95% CI 80% to 96%). At an optimal Maddrey DF cut-off of ≥37, the sensitivity, specificity, NPV and PPV were 90% (95% CI 70% to 99%), 47% (95% CI 37% to 57%), 96% (95% CI 86% to 100%) and 26% (95% CI 16% to 38%), respectively. A cut-off value of ≥32 had identical sensitivity (90%), but was slightly less specific (41%). The sensitivity and NPV for 90-day mortality of severe AH, defined as a Maddrey DF ≥32 and/or the presence of hepatic encephalopathy, were both 100%, although the specificity (25%) and PPV (22%) were low.

TABLE 4.

Operating characteristics of Model for End-stage Liver Disease (MELD) and Maddrey discriminant function (DF) for the prediction of 90-day mortality in patients hospitalized for alcoholic hepatitis (AH)

Prognostic model	Cut-off	Sensitivity, % (95% CI)	Specificity, % (95% CI)	PPV, % (95% CI)	NPV, % (95% CI)	Accuracy, % (95% CI)
MELD	≥15	95 (76–100)	10 (5–17)	18 (11–26)	91 (59–100)	25 (17–33)
	≥18	86 (64–97)	27 (18–36)	20 (12–29)	90 (73–98)	37 (28–46)
	≥21	76 (53–92)	50 (39–60)	24 (14–36)	91 (80–97)	54 (45–63)
	≥22 (Optimal)	67 (43–85)	59 (49–69)	25 (15–39)	90 (80–96)	61 (51–69)
Maddrey DF	≥32	90 (70–99)	41 (31–51)	24 (15–35)	95 (84–99)	49 (40–58)
	≥37 (Optimal)	90 (70–99)	47 (37–57)	26 (16–38)	96 (86–100)	54 (45–63)
Severe AH	Maddrey DF ≥32 and/or HE	100 (84–100)	25 (17–34)	22 (14–31)	100 (86–100)	38 (29–47)

HE Hepatic encephalopathy; NPV Negative predictive value; PPV Positive predictive value

DISCUSSION

In the present population-based study from a large Canadian metropolitan centre, we describe prognostic factors and the predictive utility of well-known risk scores in patients hospitalized for AH. As expected, older age and indicators of greater hepatic and renal dysfunction were associated with 90-day mortality (9,31). In univariate analysis, patients who died had higher MELD and Maddrey DF scores, and greater INR and serum creatinine concentrations, supporting the prognostic significance of these factors. In addition, patients who died had higher baseline WBC and platelet counts compared with survivors, presumably a reflection of the systemic inflammatory response syndrome and/or sepsis (3,4), which are frequently associated with AH. We also observed a lower risk for death among men, a finding that has not previously been reported. Although we cannot exclude unmeasured confounding as an explanation for this finding, women are known to be twice as sensitive to alcohol-mediated hepatotoxicity (32). We did not observe any significant differences between men and women in terms of treatment or severity of AH that could clearly explain this novel finding. In multivariate analyses, independent predictors of mortality included older age, female sex, and higher INR, MELD and Maddrey DF scores. The latter findings support the prognostic utilities of these models (see below).

Surprisingly, several admission characteristics did not differ significantly between survivors and those who died within 90 days. Importantly, after adjustment for other confounders, serum bilirubin and creatinine concentrations – components of MELD – were not independently associated with mortality. We speculate that this finding relates to the limited power of our study because only 21 patients (17% of the cohort) died within 90 days of hospitalization. Moreover, the limited ranges of these variables may have precluded finding a significant association with mortality. Specifically, >75% of the cohort had serum bilirubin concentrations exceeding 100 μmol/L while a similar proportion had normal serum creatinine concentrations. The frequent use of pentoxifylline treatment in our cohort may also have contributed to this finding with respect to serum creatinine because the major benefit of this therapy is the prevention of hepatorenal syndrome (2). Another noteworthy finding of our study is that 90-day mortality did not differ significantly between untreated (12%) and treated patients (19%), nor among those who received pentoxifylline (22%), corticosteroids (5%), or a combination of agents (28%). In this regard, a large randomized trial (the STOPAH trial [33]) comparing prednisolone, pentoxifylline, both or neither agent in patients with severe AH recently reported a benefit of corticosteroids only for 28-day mortality, but not beyond. Moreover, patients in our study who received therapy had greater liver disease severity (eg, higher MELD and Maddrey DF scores) compared with untreated patients, further hindering any conclusions.

The second major objective of our study was to examine the prognostic utilities of well-described risk scores in our population-based cohort of patients hospitalized for AH. As shown in Table 3, the AUROCs for 90-day mortality of MELD and Maddrey DF were similar (0.64 versus 0.68). Novel models derived in our study cohort based on age and sex combined with MELD (model 1), Maddrey DF (model 2) and INR (model 3) had slightly improved performance; however, differences were not statistically significant. These AUROC values are generally considered to be indicative of adequate test performance. Previous studies have demonstrated similar findings regarding both one- and three-month mortality (34,35). For example, in a study by Ali et al (34) that included 82 patients with AH, an AUROC for predicting 28-day mortality of 0.73 for Maddrey DF was reported. In a larger cohort of 274 patients, Sandahl et al (35) reported an AUROC of 0.74 for MELD for the same outcome measure. For 84-day mortality, an AUROC of 0.70 for MELD was reported. Discrepancies in the AUROCs observed among studies likely reflect differences in study populations and methodology (eg, treatment assignments). Other prognostic models (GAHS and Lille) have been previously validated, but we have insufficient data in our study cohort to perform any meaningful analyses on these scores. Based on existing literature, these models perform similarly to MELD and Maddrey DF with AUROCs for 28-day mortality of 0.75 for GAHS and 0.78 for the Lille model, and AUROCs for 84-day mortality of 0.72 and 0.77, respectively (35). From a practical perspective, MELD has distinct advantages compared with the other models including its relative simplicity, its familiarity among clinicians (eg, due to its widespread use in risk stratification among cirrhotic patients) and the accessibility of MELD calculators (eg, online and on hand-held devices). Moreover, whereas MELD includes INR, Maddrey DF includes PT, which can vary markedly among laboratories due to the differences in sensitivities of the thromboplastin reagents used. On the other hand, the calculation of INR accounts for this factor and, therefore, allows direct comparisons of results among laboratories (36).

The optimal cut-off values of the prognostic indices observed in our study for predicting 90-day mortality (MELD ≥22 and Maddrey DF ≥37) are similar to previous reports (8). At the respective optimal cutoffs, Maddrey DF was highly sensitive (91%) but lacked specificity (47%), whereas MELD had a better balance between sensitivity (67%) and specificity (59%). As illustrated in Table 4, the major role of MELD and Maddrey DF appears to be ruling out 90-day mortality. For example, MELD scores <22 and Maddrey DF <37 excluded death in 90% to 96% of cases. On the contrary, scores exceeding these cut-offs were poorly predictive of death based on PPVs of only approximately 25%. However, because PPVs are prevalence-dependent, the low 90-day mortality rate observed in our cohort (17%) likely contributed to these findings. In a cohort of patients with more severe AH, the accuracy for identifying patients at risk of death should improve. Noteworthy is the observation that the traditional definition of severe AH (Maddrey DF ≥32 and/or hepatic encephalopathy) had a 100% NPV, because no individuals with mild AH died within 90 days of admission. Importantly, however, one cannot conclude based on this finding that mild AH does not require treatment because the majority of our cohort (73% overall and 40% [10 of 25] with mild AH) received pentoxifylline and/or corticosteroid therapy. A study by Kulkarni et al (37) observed a 28-day mortality rate of 17% among 48 untreated patients with Maddrey DF <32. Even though this is significantly lower than the 39% mortality rate among the 41 untreated patients with Maddrey DF >32 in that study, it clearly supports the practice observed in our region of also treating some patients who have a mild presentation. Although the Maddrey DF has typically been used to predict 30-day mortality due to AH (6), our study also provides evidence supporting its use to assess the risk of death at 90 days. We focused on 90-day mortality due to the low number of deaths within one month (n=9 [7.4%]), which limited the power of related analyses.

Our study has several limitations that warrant discussion. First, the relatively small size and low 90-day mortality rate limited our ability to identify independent predictors of death and to show significant differences in the discrimination of the AH risk scores. For example, it is conceivable that differences in treatment, both specific for AH and general management, differed over time and among hospitals; however, we could not confidently exclude these effects due to limited power. In addition, due to the retrospective nature of our study, we lacked sufficient data to calculate the GAHS and Lille model for the entire cohort. Without complete data, the assessment of these scores would be restricted to only a fraction of the cohort, which would have introduced selection bias. Similarly, we used a clinical definition of AH; liver histology was not available in our cohort. A recent study (25) showed that a composite score including the degree of fibrosis, neutro-phil infiltration, type of bilirubinostasis and presence of megamitochondria (combined in the Alcoholic Hepatitis Histologic Score) was predictive of 90-day mortality (AUROC 0.77), with similar discrimination to MELD (AUROC 0.70). By combining the Alcoholic Hepatitis Histologic Score with clinical variables (age, bilirubin and INR), an improvement in discrimination was reported (AUROC 0.89) (25). Also, we did not have any data on postdischarge alcohol consumption, which may have had an important impact on mortality. Finally, we assessed all-cause, rather than liver-related mortality because underlying causes of death are not available in the data sources that we utilized. Because the calculations for these predictive models are based largely on liver-related parameters, inclusion of non-liver-related mortality might have led to an underestimation of the predictive utility of the models. Nevertheless, because most deaths of patients with AH are liver-related, particularly in severe cases, the impact of this limitation is likely small (37).

In summary, in this population-based, Canadian study of patients hospitalized for AH, older age and greater liver disease severity as assessed by MELD, Maddrey DF and INR, were associated with an increased risk of 90-day mortality. Men had a lower risk of death, but this novel finding warrants confirmation. Well-known prognostic models, such as MELD and Maddrey DF, have similar discrimination and can be used to predict 90-day mortality, particularly to identify patients with a low risk of death.

APPENDIX 1. Formulas for the prognostic models

Model	Formula
MELD	9.57 × ln(creatinine [mg/dL]) + 3.78 × ln(bilirubin [mg/dL]) + 11.20 × ln(INR) + 6.43
Maddrey DF	(4.6 × [PT - control PT (s)]) + bilirubin (mg/dL)

DF Discriminant function; INR International normalized ratio; MELD Model for End-stage Liver Disease; PT Prothrombin time

APPENDIX 2. Comparison of the prevalence of the Elixhauser comorbidities in the administrative data according to mortality at 90 days after hospitalization for alcoholic hepatitis

Comorbidity	Entire cohort (n=122)	Mortality at 90 days (n=21)	Survival at 90 days (n=101)	P
Congestive heart failure	5 (4)	1 (5)	4 (4)	1.00
Cardiac arrhythmia	5 (4)	2 (10)	3 (3)	0.20
Valvular disease	1 (1)	0 (0)	1 (1)	1.00
Pulmonary circulation disorders	1 (1)	1 (5)	0 (0)	0.17
Peripheral vascular disorders	0 (0)	0 (0)	0 (0)	–
Hypertension	12 (10)	3 (14)	9 (9)	0.43
Hypertension, uncomplicated	12 (10)	3 (14)	9 (9)	0.43
Hypertension, complicated	0 (0)	0 (0)	0 (0)	–
Paralysis	0 (0)	0 (0)	0 (0)	–
Other neurological disorders	15 (12)	2 (10)	13 (13)	1.00
Chronic pulmonary disease	6 (5)	1 (5)	5 (5)	1.00
Diabetes, uncomplicated	5 (4)	0 (0)	5 (5)	0.59
Diabetes, complicated	1 (1)	1 (5)	0 (0)	0.17
Hypothyroidism	4 (3)	0 (0)	4 (4)	1.00
Renal failure	0 (0)	0 (0)	0 (0)	–
Peptic ulcer disease excluding bleeding	4 (3)	2 (10)	2 (2)	0.14
AIDS/HIV	2 (2)	0 (0)	2 (2)	1.00
Lymphoma	0 (0)	0 (0)	0 (0)	–
Metastatic cancer	0 (0)	0 (0)	0 (0)	–
Solid tumour without metastasis	0 (0)	0 (0)	0 (0)	–
Rheumatoid arthritis	1 (1)	1 (5)	0 (0)	0.17
Coagulopathy	13 (11)	2 (10)	11 (11)	1.00
Obesity	2 (2)	0 (0)	2 (2)	1.00
Weight loss	5 (4)	0 (0)	5 (5)	0.59
Fluid and electrolyte disorders	25 (20)	3 (14)	22 (22)	0.56
Blood loss anemia	5 (4)	0 (0)	5 (5)	0.59
Deficiency anemia	15 (12)	5 (24)	10 (10)	0.14
Drug abuse	2 (2)	0 (0)	2 (2)	1.00
Psychosis	3 (2)	0 (0)	3 (3)	1.00
Depression	8 (7)	1 (5)	7 (7)	1.00

Data presented as n (%) unless otherwise indicated

APPENDIX 3. Comparison of patient characteristics according to sex

	Entire cohort (n=122)	Male (n=73)	Female (n=49)	P
Age, years	49 (42–55)	50 (42–55)	47 (43–54)	0.49
Length of stay	10 (5–21)	10 (5–19)	9 (6–24)	0.91
Admission hospital
1	51 (42)	32 (44)	19 (39)	0.72
2	36 (29)	22 (30)	14 (29)
3	35 (29)	19 (26)	16 (33)
Year
2008	27 (22)	13 (18)	14 (29)	0.16
2009	18 (15)	11 (15)	7 (14)
2010	21 (17)	17 (23)	4 (8)
2011	26 (21)	13 (18)	13 (27)
2012	30 (25)	19 (26)	11 (22)
Treatment
None	33 (27)	17 (23)	16 (33)	0.54
Pentoxifylline only	51 (42)	34 (47)	17 (35)
Corticosteroids only	20 (16)	11 (15)	9 (18)
Pentoxifylline + corticosteroids	18 (15)	11 (15)	7 (14)
Admission laboratory values
Bilirubin, μmol/L	175 (106–272)	200 (113–294)	171 (98–256)	0.40
International normalized ratio	1.6 (1.3–1.9)	1.6 (1.3–1.8)	1.6 (1.3–2.0)	0.76
Creatinine, mmol/L	57 (46–78)	59 (52–78)	53 (38–71)	0.02
White blood cells, ×109/L	9.0 (5.9–13.5)	8.7 (6.2–13.3)	9.4 (5.7–14.1)	0.99
Hemoglobin, g/L	113 (97–133)	119 (102–134)	106 (91–119)	0.01
Platelets, ×109/L*	120 (68–184)	106 (62–160)	154 (89–213)	0.02
Albumin, g/L*	23 (20–26)	23 (21–25)	23 (19–26)	0.91
Alanine aminotransferase, U/L	50 (32–76)	57 (35–82)	41 (26–57)	0.03
Alkaline phosphatase, U/L*	234 (168–331)	241 (174–354)	224 (165–286)	0.33
Gamma-glutamyl transferase, U/L*	555 (300–955)	626 (344–1083)	440 (237–838)	0.046
Early change in bilirubin†	50 (50)	32 (52)	18 (47)	0.84
Prognostic models
Maddrey discriminant function	45 (26–62)	45 (27–59)	45 (26–62)	0.88
Model for End-stage Liver Disease	21 (18–24)	21 (18–24)	21 (17–24)	0.52
Severe alcoholic hepatitis‡	97 (80)	60 (82)	37 (76)	0.37
Complications
Hepatic encephalopathy	58 (48)	36 (49)	22 (45)	0.71
Ascites	87 (71)	54 (74)	33 (67)	0.54

Data presented as median (interquartile range) or n (%).

^*Data available on 121 patients for platelets, 103 patients for albumin, 121 patients for alkaline phosphatase and 119 patients for gamma-glutamyl transferase;

^†Decreased serum bilirubin on day 7. Patients admitted for <7 days who had a drop in bilirubin on day 3 and were discharged alive were assumed to have an early change in bilirubin level (data available for 100 patients);

^‡Defined as Maddrey discriminant function ≥32 and/or the presence of hepatic encephalopathy