Background
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 infection and hepatorenal syndrome, both of which significantly increase mortality [
2‐
4]. Untreated patients with severe AH, typically defined by a Maddrey discriminant function (DF) ≥32 and/or the presence of hepatic encephalopathy, have a particularly poor prognosis with one-month mortality rates ranging from 30 % to 50 % [
5,
6]. In patients with a mild presentation of AH, the risk of progression to cirrhosis is 50 %; this risk is highest in patients who continue to abuse alcohol [
7].
Population-based studies describing the epidemiology and outcomes of AH are limited. A Danish study by Sandahl et al. [
8] was the first population-based epidemiologic study of AH and demonstrated an increase in annual incidence among both Danish men and women between 1999 and 2008. Another study described the clinical characteristics and mortality of patients hospitalized in the United States for AH using the Nationwide Inpatient Sample (NIS) database [
9]. Both of these studies relied on International Classification of Diseases (ICD) diagnosis codes for AH in an administrative database for case identification. Although such codes have been validated for other non-hepatic and hepatic conditions (e.g. cirrhosis, viral hepatitis, autoimmune liver disease, drug hepatotoxicity) [
10‐
21], their validity for AH has yet to be confirmed. While the burden of alcohol-related liver disease is increasing in many regions, large-scale studies describing the epidemiology of AH in Canada have been hindered by the lack of a validated coding algorithm. Therefore, the primary objective of our study was to validate coding algorithms for the identification of patients hospitalized for AH using administrative data. Our study findings will inform researchers if administrative data can be used for epidemiological studies and surveillance of this condition.
Discussion
In this population-based study, we assessed the validity of ICD diagnosis coding algorithms for AH in a Canadian hospitalization database. Our finding of a low PPV for AH (54 %) suggests that AH was not accurately and completely coded in the administrative data and caution must be exercised if the DAD is used for large-scale epidemiologic studies of AH. Under the assumption that the PPV remains relatively constant over time (as observed in our study), the DAD could be used to assess temporal trends in AH admissions. However, the exact incidence for any given year based on this data is likely to be erroneous. Administrative data could also be used as a screening tool for potential cases with AH. In this situation, confirmation of the presence of AH requires additional clinical information such as that obtainable via a review of medical records. In light of our findings, the validity of previous studies that used administrative data to study the epidemiology and outcomes of AH should be scrutinized. In one study, Sandahl and colleagues used this ICD-10 code to identify AH cases from the Danish National Registry of Patients and describe its incidence and associated mortality [
8]. In total, 1,951 suspected cases of AH were identified between 1999 and 2008. The annual incidence increased from 37 to 46 per million population in men and from 24 to 34 per million in women. However, according to our results, close to half of the patients identified in this study may not have been true cases of AH. This study also reported 84-day mortality ranging from 14 % to 24 %, similar to the 90-day mortality rate of 17 % for confirmed AH in our study. In another study, Liangpunsakul used the Nationwide Inpatient Sample hospitalization database to report the outcomes of patients hospitalized for AH in the United States with case identification using the ICD-9 code for this condition (571.1) [
9]. The study showed that AH represented 0.71 % of all hospital admissions in 2007 and reported an in-hospital mortality rate of 6.8 %. Although our study validated only the ICD-10 coding classification (as ICD-9 codes were unavailable), there is no clear rationale for the accuracy to differ among the two classification systems both of which include a single code for AH. Moreover, a study by Quan and colleagues that used the DAD (as in our study) demonstrated that the coding accuracies between the ICD-9 and ICD-10 classifications for liver diseases are very similar (AH was not examined specifically in this study) [
37]. Therefore, our results should be generalizable to databases using either the ICD-9 or ICD-10 classifications.
The diagnosis field in which the code for AH was recorded in the administrative database (i.e. primary vs. secondary) had a significant impact on algorithm validity. Indeed, the majority (89 %) of confirmed AH cases had the AH code recorded as the primary diagnosis (vs. 50 % among non-AH hospitalizations). Accordingly, the PPV improved from 54 % to 67 % when the cohort was restricted to those with AH as the primary diagnosis compared with one of the secondary diagnosis fields. Patients with a severe presentation were also more likely to have AH recorded as the primary diagnosis. Specifically, 77 % of cases with hepatic encephalopathy and/or a Maddrey DF ≥32 had AH as a primary diagnosis compared with only 59 % of those with mild hepatic dysfunction. Algorithms that also included diagnosis codes for ascites or gastrointestinal hemorrhage had improved performance, particularly when restricted to cases with a primary diagnosis of AH in which PPVs of 76 % to 78 % were observed. The corollary is that studies focused on patients with severe AH may potentially identify these cases from administrative databases by restricting to cases with AH as the primary diagnosis plus these associated conditions. However, the prevalence of these cirrhosis-related complications was low enough (<40 %) in confirmed AH cases that any study employing this methodology will have reduced sensitivity for the identification of all relevant cases. Based on the sub-optimal accuracy observed in our study, we would advise that any AH case identified in this manner be confirmed via a review of medical records.
In addition to examining the impacts of coding details and disease severity on the validity of an AH code in the administrative data, we studied the effects of year and hospital of admission. Despite the presence of different health records coders at the three hospitals in our region, coding validity was similar supporting the generalizability of our findings. Moreover, the PPV of an AH code did not differ across study years, potentially supporting the temporal trends in disease incidence suggested by the study of Sandahl and colleagues [
8]. In this regard, we did not observe an increase in the number of admissions over time although the sample size of our study was limited.
While previous studies have confirmed the validity of administrative data for the identification of patients with various liver disorders [
19,
38,
39], our results suggest that caution is needed when using these data sources to study AH. However, the major challenge with our analysis relates to the diagnostic definition for AH. Specifically, AH represents a spectrum ranging from mild abnormalities in liver biochemistry to life-threatening liver failure due to abusive alcohol consumption when other causes of liver disease (e.g. viral hepatitis and drug hepatotoxicity) have been excluded. Although AH has characteristic histological findings including steatosis, ballooned hepatocytes, Mallory bodies, lobular neutrophilc inflammation, and centrizonal fibrosis [
40], liver biopsy is not routinely performed in our region, nor are these findings specific (e.g. they may be seen in patients with non-alcoholic steatohepatitis). The diagnostic criteria suggested by Lucey et al. (i.e. elevated AST but <300 IU/L, AST to ALT ratio >2, serum bilirubin >86 μmol/L, elevated INR, and neutrophilia in patients with ascites and a history of heavy alcohol use) [
1] are very stringent and would mostly capture only patients with severe AH. In light of this fact, we utilized a less rigorous definition of AH that allowed us to identify a wider spectrum of suspected AH cases. Moreover, since AST is not part of standard liver biochemical profiles in our region and due to the high correlation between AST and ALT, we considered ALT in place of AST when necessary. Nevertheless, since such a biochemically focused, even more lenient, case definition is generally not strictly followed in clinical practice, patients who do not meet all of these criteria may still be diagnosed with AH by their physician thereby contributing to the sub-optimal accuracy of the diagnosis code observed in our study.
Among the 106 patients who did not fulfill our diagnostic criteria for confirmed AH, 31 patients clearly did not have AH upon medical records review. The diagnosis for these patients includes cirrhosis, sepsis, recent history of AH, pancreatitis, acetaminophen toxicity, cocaine use, heart failure and alcohol intoxication.
Taking into consideration that our definition of AH may not be strictly followed by physicians, we performed a sensitivity analysis using a loosened definition of AH that confirmed diagnosis of AH if patients fulfilled both of the following criteria: heavy alcohol consumption (>196 g/week or >56 g in any day among males, and >98 g/week or >42 g in any day among females) and exclusion of other causes of acute hepatic dysfunction (e.g. drug hepatotoxicity, autoimmune hepatitis, ischemic hepatitis, etc.). With this loosened diagnostic criteria, PPV improved to 73 % (95 % CI 67-79 %), up from 54 %. This improved PPV remains sub-optimal, thus further supporting our initial findings that AH was not accurately and completely coded in the administrative data. Realizing the limitations of both sets of diagnostic criteria, one could infer that the true PPV for the AH diagnosis code lies within 54 % to 73 %.
Our study has several limitations that warrant discussion. First, we lack data for a control group representative of the general hospitalized population. Without controls, the sensitivity, specificity and NPVs of the diagnosis code for AH (and related algorithms) cannot be calculated. Second, since this is a retrospective study, the reliability of patients’ self-reported alcohol intake is questionable. On numerous occasions, multiple descriptions of alcohol consumption were recorded in the medical record. Anecdotally, patients tended to admit to greater alcohol intake when questioned by addictions specialists. As a result, we utilized a hierarchical approach to record alcohol consumption (see Methods). Since the amount of alcohol intake is a vital criterion in the reference standard for an AH diagnosis, underreporting may have led to an underestimation of AH in our study.
Competing interests
Dr. Myers was supported by a salary support award from the Canadian Institutes for Health Research (CIHR). Dr. Kaplan is supported by salary support awards from CIHR and Alberta Innovates-Health Solutions (AIHS). Dr. Swain is supported by the Cal Wenzel Family Foundation Chair in Hepatology. Dr. Quan is supported by a salary support award from AIHS. Dr. Borman is supported by a Canadian Association for the Study of the Liver/Vertex Clinical Hepatology Fellowship. Dr. Heitman was supported by an award from the Noel Hershfield Professorship in Therapeutic Endoscopy.
Authors’ contributions
JP was involved in the design of the study, data extraction, data analysis, and writing of the manuscript. RM, ER, MB and MS were involved in the conception and design of the study. ER, MB, SZ assisted in data collection for the study. GK, SH, HQ and KB provided assistance for the analysis of this study. All authors participated in the final development of the manuscript and approved of its contents.