A Cohort Study of the Incidence of Serious Acute Liver Injury in Diabetic Patients Treated With Hypoglycemic Agents FREE

CASE REPORTS of drug-induced liver injury frequently appear in the literature, and physicians are encouraged to report suspected cases to the manufacturers or the regulatory authorities.¹ These voluntary reports serve as signals to evaluate the safety of medications after marketing approval.² However, the limitations of spontaneous adverse drug reaction reports in postmarketing safety assessment are well recognized,^3- 4 and epidemiologic studies that estimate the risk of adverse events in large cohorts of drug users are needed to help interpret these reports.⁵

Troglitazone, the first of the thiazolidenediones approved for the clinical management of type 2 diabetes mellitus,⁶ was withdrawn from the market in the United Kingdom⁷ and the United States⁸ after case reports of acute liver failure associated with its use.^9- 12 Early clinical trials¹³ of troglitazone showed that 1.9% of study patients using this drug had elevations in serum alanine aminotransferase levels to more than 3 times the upper limit of normal vs a risk of 0.6% among those who received placebo. However, the trials were not large enough to precisely estimate the incidence of more severe forms of acute liver injury or to determine the relative rates of liver events in patients exposed to troglitazone and other hypoglycemic medications.¹⁴ Because diabetes mellitus can be associated with liver disease independent of medication use,^15- 19 it is important to determine the population-based incidence of acute liver injury in patients with diabetes mellitus and, particularly, the incidence of liver events without an apparent cause. This study was conducted in 5 geographically diverse health maintenance organizations (HMOs) to estimate the incidence of acute liver injury not clearly attributable to other known causes that resulted in hospitalization in a large cohort of patients who were treated with insulin or oral hypoglycemic drugs.

This is a retrospective cohort study that used automated health care information from HMOs to identify study patients and potential outcomes of interest, followed by expert review of hospital records of potential liver events. The study was approved by the institutional review board of each participating HMO.

The 5 HMOs that participated in the study are Fallon Community Health Plan (Worcester, Mass), HealthPartners (Minneapolis, Minn), Harvard Pilgrim Health Care, Kaiser Permanente of Northern California (Oakland), and Lovelace Health Plan (Albuquerque, NM). All 5 HMOs are members of the HMO Research Network and have established records of epidemiologic research.^20- 21 Automated pharmacy dispensing, demographic, and membership information at each HMO was used to identify a cohort of diabetic patients who had at least 1 dispensing of a hypoglycemic agent, including insulin (all preparations), any sulfonylurea, metformin, troglitazone, rosiglitazone, acarbose, miglitol, and repaglinide between April 1, 1997, and June 30, 1999. Eligible individuals were HMO members with prescription drug benefit for at least 60 days during the study period who were 19 years or older at the time of first dispensing of a hypoglycemic agent during the same period. Review of a random sample of 501 study patients #1approximately 100 patients for each HMO) showed that 99% had objective evidence of diabetes mellitus in their outpatient medical records. The study population comprised 171 264 men and women aged 19 to 102 years.

We identified patients with possible incident liver disease through an automated search of hospital discharge records of all study patients between April 1, 1995, and September 30, 1999. An extensive set of diagnosis and procedure codes that could plausibly be associated with, or result from, acute liver injury served as a highly sensitive measure for capturing potential cases #1Table 1). Hospitalizations with discharge diagnoses only for specific viral hepatitis or hepatic or biliary tract cancer were not sought, and neither were same-day hospitalizations. We also excluded patients with apparently prevalent liver disease, which was operationally defined as (1) patients with a single discharge diagnosis during the study period who had an identical discharge diagnosis within 2 years before April 1, 1997, for the following conditions: disorders of copper metabolism, acute or subacute necrosis of the liver, chronic liver disease and cirrhosis, hepatic coma, hepatorenal syndrome, unspecified hepatitis, hepatic infarction, other specified disorders of the liver, or unspecified disorders of the liver or (2) those who underwent liver transplantation during the 2 years before April 1, 1997.

A study patient could have 1 or more hospitalizations during the study period. Full-text medical records of all hospitalization episodes that fulfilled the selection criteria were requested for review. These potential cases underwent a 3-level review. First-level review was conducted by trained medical reviewers who classified the hospital event as (1) no evidence of liver disease (eg, coding errors, liver biopsy that yielded normal tissue findings, or nonhepatic causes of jaundice), (2) chronic liver disease (liver disease that clearly predated April 1, 1997), or (3) acute liver disease.

At the second level, one of us (J.V.S.) confirmed the status of individuals without liver disease and those with chronic liver disease. He reviewed the remaining potential cases and further excluded those who met predefined diagnostic criteria for 1 of 9 clinical conditions (primary hepatic carcinoma or metastatic cancer to the liver; viral hepatitis A, B, or C; alcoholic hepatitis; biliary tract obstruction; or hepatic injury due to hypovolemic, cardiogenic, or septic shock), those who had a chronic liver condition at the time of the first hospital admission with no evidence suggesting acute changes, and those with an apparent cause according to the attending physicians and who did not meet the diagnostic criteria for acute liver injury or failure defined in Table 2.

The third level of review was conducted by 3 hepatologists (J.E.E., R.H.M., and N.T.) who were masked to exposure status to hypoglycemic agents. Two hepatologists independently reviewed each potential case that was judged to have serious acute liver injury not attributable to a nondrug cause by the second-level physician reviewer. They categorized each hospitalization for 3 criteria: degree of hepatic injury, acuity of injury, and cause of injury #1Table 2). A patient was found to have probable acute liver failure if he or she had (1) hepatic encephalopathy, #12) coagulopathy (prothrombin time prolongation >3 seconds or an international normalized ratio >1.5), and (3) jaundice (total bilirubin level >3.0 mg/dL #1>51 µmol/L]); possible acute liver failure if he or she fulfilled 1 or 2 of the criteria; and acute liver injury if he or she did not meet any of the criteria for liver failure but had alanine transaminase or aspartate transaminase levels greater than 500 U/L. Causality assessment was based on all available information in the medical records, including tests for viral hepatitis, history of alcohol use, and diagnostic workup for possible malignancy. If the 2 reviewers did not reach the same conclusion, the third hepatologist was consulted. Consensus was reached in all cases after discussion among the hepatologists. The outcomes of interest were acute liver failure or acute liver injury without a probable etiology other than use of hypoglycemic agents.

Hypoglycemic agents were operationally classified into 4 groups: insulin #1all preparations), sulfonylureas (acetohexamide, chlorpropamide, glimepiride, glipizide, glyburide, tolazamide, and tolbutamide), metformin, and troglitazone. Acarbose, miglitol, repaglinide, and rosiglitazone were infrequently used during the study period, and their exposure data are not presented. Consecutive dispensings of drugs of the same group to the same person defined an exposure era, which started with the first dispensing and ended on the earliest of the following: (1) the date of the last dispensing plus the days-of-supply of the last dispensing plus 30 days, (2) termination of membership, or (3) June 30, 1999. During the study period, a patient could have 2 or more exposure eras to the same group of drugs or 2 or more groups of drugs with overlaps and intervening gaps. Total exposed person-time for insulin, sulfonylurea, metformin, or troglitazone was the summation of length of exposure eras to each drug group over all study patients. Patients with overlapping exposure eras to 2 or more groups of drugs contributed person-time to each group.

Numbers of cases of probable or possible acute liver failure or injury without a probable cause other than use of hypoglycemic agents were the numerators in the incidence estimates. For patients with multiple liver events, only the first event was counted. A liver event that occurred when the patient was exposed to multiple drug groups was counted toward the incidence estimate for each drug group. All incidence estimates and their 95% confidence intervals were standardized according to the age and sex distribution of the entire study population.

We conducted 3 types of auxiliary analyses to evaluate the robustness of results under different assumptions. First, a substantial number of liver events occurred during gaps of no exposure to any hypoglycemic agent when we defined exposure to be up to 30 days beyond the days-of-supply of a dispensing. We carried out analyses that allowed for longer periods (an additional 60 and 180 days) after each dispensing. Second, because risk for liver injury may be greater during the first few months of therapy, we repeated incidence calculations using only the first 6 months of exposure to each drug group for each person. In these analyses, persons without at least 4 months of HMO membership before the first dispensing were excluded to ensure the validity of incident drug use. Third, to assess whether the unmasked first- and second-level review of medical records would introduce bias and obscure an increase in troglitazone-related acute liver disease before the masked hepatologist review, we evaluated incidence for all acute liver events identified for the first-level review regardless of the severity and causality assessment by the hepatologists.

We used a modified version of the chronic disease score²² that was based on the use of drugs for chronic diseases before cohort inception to measure comorbidity among the study patients. Because this was a study of diabetic patients, we modified the calculation such that previous use of hypoglycemic agents did not contribute to the score. We used a proportional hazards model²³ with time to confirmed liver event as the dependent variable, 4 time-varying independent variables to represent exposure to the 4 drug groups, and age, sex, and chronic disease score as covariates. The hazard ratio for each drug group is interpreted as the ratio of the hazard during person-time exposed to the drug group to the hazard for all person-time while not exposed to that drug group, which includes person-time of exposure to agents of other groups and person-time while not exposed to any hypoglycemic agent.

A total of 171 264 adult members of the 5 participating HMOs received at least 1 oral hypoglycemic drug or insulin between April 1, 1997, and June 30, 1999. All but 121 received at least 1 dispensing of a sulfonylurea, metformin, troglitazone, or insulin during the study period. The mean ± SD age of the study cohort was 57.9 ± 14.1 years as of April 1, 1998, and 46.5% of them were female. Of all the patients, 120 855 (70.6%) had received a sulfonylurea, 57 413 (33.5%) had received insulin, 53 564 (31.3%) had received metformin, and 9604 (5.6%) had received troglitazone.

The study cohort received 2 559 448 dispensings of hypoglycemic agents during the study period, which corresponded to a mean ± SD of 14.9 ± 13.4 dispensings per patient, with a range of 1 to 177 dispensings per person during the 2½-year study period. Insulin was frequently used alone, with only 23% of all insulin-exposed person-time overlapping with oral hypoglycemic drug use. For sulfonylureas, 68% of exposed person-time was for sulfonylurea alone, and 25% of exposed person-time was for sulfonylurea and metformin concomitantly. On the other hand, only 23% of exposed person-time for metformin and 15% of exposed person-time for troglitazone were for use of either drug alone. Both drugs were commonly used with sulfonylureas (66% of metformin-exposed person-time and 32% of troglitazone-exposed person-time). Troglitazone exposure also overlapped frequently with insulin use (38% of exposed person-time).

We identified 1136 patients with 1 or more episodes of hospitalization longer than 1 day with at least 1 of the discharge diagnosis or procedure codes of interest during the study period. On the basis of automated data, we excluded from further review 250 patients who apparently had prevalent liver disease. Medical records of 1401 episodes of hospitalization were requested from the hospitals, and we received records for 1287 (92%) unique episodes. At the first-level review by medical record reviewers, no evidence of liver disease was found for 106 episodes, and clear evidence that the liver disease predated April 1, 1997, was found in 245 episodes. At the second-level, 172 episodes were found to be due to cancer, viral hepatitis, alcoholic liver disease, biliary tract obstruction, or shock. The physician reviewer found that 526 episodes had chronic hepatobiliary conditions, such as alcoholic liver disease or cryptogenic cirrhosis, without acute changes to suggest a new insult, and 126 episodes were acute events that did not meet criteria for acute liver failure or injury and that could be attributed to an apparent cause, such as fatty liver and congestive hepatomegaly. The remaining 112 hospitalizations of 91 persons underwent third-level review.

Of the 91 patients reviewed by the hepatologists, 71 had liver failure or injury, 19 did not meet the criteria for liver failure or injury, and 1 did not have enough information in the medical record (Table 2). Twelve patients with liver failure or injury had a chronic liver condition. Among the 59 persons with probable or possible acute liver failure or injury, the hepatologists identified a probable cause other than hypoglycemic agents in 24 cases. The remaining 35 patients (26 had acute liver failure and 9 had acute liver injury) formed the basis for incidence estimates; 18 were men, and the mean ± SD age was 62.8 ± 11.0 years (range, 44-86 years). Eight (23%) of the 35 patients died during hospitalization. Twenty-seven of the 35 cases had a possible alternative cause other than a hypoglycemic agent, including concomitant exposure to hepatotoxic drugs. Only 1 of the 35 cases had no apparent cause other than use of the hypoglycemic agent, which was tolazamide.

The overall age-and sex-standardized incidence in the study cohort, regardless of drug exposure, was 0.07 per 1000 person-years for probable or possible acute liver failure and 0.1 per 1000 person-years for acute liver failure or injury. Considering only cases of acute liver failure, age- and sex-standardized incidence is approximately 1 per 10 000 person-years or less for each drug exposure group (Table 3). The relative order of incidence among the different drug exposure groups did not change substantially when events and exposure time beyond 30 days after a given dispensing were included. Combining cases of acute liver failure and acute liver failure of lesser severity showed similar patterns of relative rank of incidence.

We reexamined the incidence of acute liver failure and acute liver injury by restricting follow-up to the first 6 months of exposure to each drug group #1Table 4). The incidence was larger, indicating increased risk during the early course of therapy, but again similar across the 4 drug groups. An analysis of all hospitalized acute liver events, regardless of their disposition (n = 936), did not materially affect the relative order of the incidence (data not shown).

We analyzed the 130 905 study patients (76.4% of the full cohort) who had at least 1 year of health plan membership before the first dispensing of a hypoglycemic agent during the study. Twenty-seven of the 35 cases of outcomes of interest were included in this subset of data. We calculated the baseline modified chronic disease score using pharmacy dispensing records for these patients, and we estimated the hazard ratios for each drug group in proportional hazards models, with exposure to each drug group as time-varying predictors. With adjustment for age, sex, and comorbidity, the hazard ratios did not differ substantially across the 4 drug groups (Table 5). The hazard ratio for insulin was largest (2.41; 95% confidence interval, 0.98-5.94), whereas the hazard ratios for the other 3 drug groups were nearly identical. We then fitted these models without the chronic disease score so that all study patients and liver events could be included. Results did not change substantially except that the hazard ratios for insulin and metformin increased and the 95% confidence intervals excluded the null value.

The occurrence of acute liver injury in persons with chronic illnesses such as diabetes mellitus treated with multiple drugs has not been well quantified. Reports of adverse events during exposure to specific agents leave open the questions of the extent of ascertainment of events (the numerator) and of total exposure (the denominator). Observation of entire populations over time is a nearly ideal approach provided that those populations are large enough to generate precise estimates for what are expected to be rare events. We applied a systematic ascertainment and review protocol to a population-based cohort of more than 170 000 patients with diabetes mellitus. We estimate that acute liver failure and acute hospitalized liver injury not attributable to a known cause other than use of diabetic medications occurs with a frequency of approximately 1 per 10 000 person-years in those with diabetes mellitus. The incidence in patients exposed to different oral hypoglycemic drugs seemed to be similar, and patients with concomitant chronic diseases were at higher risk of developing an acute liver event. This is crucial information in the risk-benefit profiles of hypoglycemic agents that will help physicians select specific drugs to treat patients with diabetes mellitus.

Jick et al²⁴ reported an incidence of possibly drug-induced liver disease of 5 per 10 000 person-years in a cohort of 44 406 type 2 diabetic patients from the United Kingdom treated with oral hypoglycemic agents. Taking into account that in the United Kingdom study 29 of the 57 identified cases were described as asymptomatic or having mild disease, the reported rate from the United Kingdom is compatible with the rates we found in US HMOs. In the study by Jick et al, only 2 of the 57 cases did not have a possible cause of the acute liver disease other than oral hypoglycemic drugs. In our study, 1 of 35 cases, who was exposed to tolazamide, was definitely ascribed to use of a hypoglycemic agent. For the other 34 cases, there was insufficient evidence to ascribe a specific cause other than use of a hypoglycemic agent, and 27 could possibly be caused by concomitant drug use (such as diclofenac sodium or lovastatin) or concomitant medical conditions. Development of acute liver disease in these individuals was likely to be a result of multiple contributing factors.²⁵

As in other epidemiologic studies^26- 28 of acute liver injury, we used a set of sensitive diagnosis and procedure codes to identify all potential liver events. As expected, most of these events were due to chronic rather than acute liver disease, and nearly all could be attributed to one of the common causes of liver injury. We used a structured medical record review instrument with explicit criteria for many of the exclusions and a masked adjudication process to eliminate bias. Nevertheless, the physician reviewer was responsible for excluding many cases judged to be due to chronic liver disease or definable acute causes that did not meet the study criteria for acute liver injury. Concern about the possibility that bias may have affected our findings led us to compare the occurrence of all possible acute events. As in the more restricted analyses, we again found similar rates among patients exposed to different groups of hypoglycemic agents.

Although the safety of troglitazone therapy prompted initiation of this study, we recognized in advance that the number of troglitazone users was not large enough to provide precise incidence estimates of rare events. The incidence of acute liver failure or injury was similar among the various oral hypoglycemic agents, but the few confirmed liver events precluded precise relative risk estimates among individual drugs. The incidence of acute liver failure or injury seemed to be higher in patients using insulin. In an extensive review of the literature, insulin was not found to be hepatotoxic,²⁹ although hepatotoxicity attributed to insulin has been reported in isolated cases in Japanese patients.³⁰ The increased incidence among insulin-exposed person-time was probably due to the longer duration and severity of diabetes mellitus among patients receiving insulin.

According to the Food and Drug Administration, the incidence of acute liver failure among troglitazone users was 1 in 8000 to 1 in 20 000.³¹ Two cases of acute liver failure were reported in 2 clinical trials of troglitazone.³¹ One case of acute liver failure was found in a cohort of 9369 troglitazone users with 4873 person-years of exposure in an HMO population.³¹ In our study, on the basis of 1 case of acute liver failure that was possibly due to troglitazone therapy, we estimated this risk to be 1 in 10 000 person-years per year, with a 95% confidence interval compatible with rates ranging from 1 per 500 000 to 1 per 2000 person-years. This patient also had concurrent use of amiodarone, a known cause of liver injury,³² and acute cholangitis, exemplifying the problem of attributing cause and the value of an epidemiologic (rather than a notification) approach to assessing risk. The 3-fold higher rates of acute liver events during the first 6 months of exposure to any oral hypoglycemic drug suggested increased risk early in the course of therapy. A postmarketing study³³ in Japan found 45 cases of abnormal liver function test results among 1052 troglitazone users but no acute liver failure. Although our study has the largest reported cohort of troglitazone-exposed persons, cohorts with even more troglitazone users are needed to more precisely estimate the incidence of these rare events and the relative risk of troglitazone use compared with the use of other hypoglycemic agents.

Corresponding author and reprints: Joe V. Selby, MD, MPH, Division of Research, Kaiser Permanente, 2000 Broadway, Oakland, CA 94612.

Accepted for publication June 26, 2002.

This study was supported by a research grant from Parke-Davis Pharmaceutical, Ann Arbor, Mich. Parke-Davis is the US manufacturer of troglitazone, and Parke-Davis scientists provided input into the study design and reviewed the manuscript. The authors had final authority with regard to content and wording of the manuscript.

This study was presented at the 17th International Conference on Pharmacoepidemiology, Toronto, Ontario, August 25, 2001.

We thank Claire Canning, Jackie Cernieux, Marlie Jacobs, Susan Paine, Melissa Parker, Hans Petersen, Parker Pettus, Nicholas Strub, and Bix Swain for technical support.