Combining K-72 Hepatic Failure with 15 Individual T-Codes to Identify Patients with Idiosyncratic Drug-Induced Liver Injury in the Electronic Medical Record

Abstract

Background

The aim of this study was to determine the utility of combining three K72 codes (hepatic failure) with 15 individual T-Codes (drug toxicity/poisoning) to identify potential DILI cases.

Methods

The EMR was searched for encounters that had a K72 code combined with a T-code that also met minimal liver injury laboratory criteria between 10/1/15 and 9/30/18. After manual chart review, a DILIN expert opinion causality score (1–5) was assigned to each case.

Results

Among the 345 patient encounters identified, mean age was 57 years, 53% were male, and 89% Caucasian. Thirty-seven cases (10.7%) were adjudicated as probable DILI with antibiotics being the most frequently identified suspect drugs. Of the 308 non-DILI cases, liver injury was most commonly due to congestive hepatopathy (38%) and hepatic metastases (15%). The probable-DILI cases were significantly more likely to have hepatocellular liver injury (57% vs 32.5%, p = 0.01), higher total bilirubin levels (7.7 vs 4.6 mg/dl, p = 0.03), and more severe liver injury scores (p < 0.01). The K72.0 (acute/subacute hepatic failure) yielded the most DILI cases (29) compared to K72.9 (13) and K72.1 (0). The positive predictive value of the searching algorithm was 10.7% and improved to 15% when using only the K72.0 codes.

Conclusions

K72 codes combined with drug poisoning T-codes had a low positive predictive value in identifying patients with idiosyncratic DILI. These data support further refinement of ICD-10-based algorithms to detect DILI cases in the EMR.

Introduction

Idiosyncratic drug-induced liver injury (DILI) is an important cause of liver injury and liver-related hospitalizations [1]. Although the incidence of DILI is low in the general population with an estimated incidence of 10 to 19 cases per 100,000 patient years, severe injury from idiosyncratic drug hepatotoxicity accounts for up to 13% of acute liver failure (ALF) cases [2, 3]. Furthermore, hepatotoxicity is a leading reason for the removal of medications from the marketplace [4]. Establishing a diagnosis of idiosyncratic DILI is challenging due to the lack of a consistent relationship between drug dose, route, or duration with DILI onset, and the need to exclude other more common causes of liver injury [1]. Adding to the diagnostic difficulty is the complex interplay between drug, environmental, and other host factors that may influence interindividual susceptibility to DILI [5, 6].

Timely and accurate identification of DILI patients is not only important for clinical care but also for pharmacoepidemiology and DILI pathogenesis studies [7]. We previously illustrated that an algorithm combining K71 codes (toxic liver disease) with T-codes for drug toxicity/poisoning and a liver chemistry laboratory filter was able to identify DILI cases from the electronic medical record (EMR) with a positive predictive value of 66.5% [8]. However, it remained uncertain the extent to which the algorithm was potentially missing additional DILI cases in the EMR. Specifically, we hypothesized that an algorithm directed toward patients with a diagnosis of hepatic failure (K72 code) may help identify additional DILI cases. Therefore, the aim of the current study was to determine if combining K72 codes with T-codes could further improve the ability to identify DILI cases in the EMR.

Methods

Patient Selection

A waiver to conduct this retrospective study was approved by the University of Michigan Medical School Institutional Review Board. The EMR (Epic Systems, Madison, WI) at the University of Michigan Health Systems was searched for DILI using a computerized algorithm. Initial search criteria included patients who had an inpatient or outpatient encounter with one of three ICD-10 [9] K72 codes (K72.0, acute and subacute hepatic failure; K72.1, chronic hepatic failure; K72.9, hepatic failure unspecified, Supplemental Table 1) and at least one of fifteen ICD-10 T-codes between October 1, 2015, and September 30, 2018 (Table 2). After identifying these patients, a filter was then applied to select patients that met at least one laboratory criteria for clinically significant liver injury. Of note, we did not exclude individual cases that may have add another ICD-10 code assigned for the liver injury episode such as alcoholic liver disease or congestive hepatopathy since our intent was to be inclusive in finding a rare entity like acute DILI. Consistent with Drug-Induced Liver Injury Network (DILIN) laboratory criteria, clinically significant liver injury was defined as aspartate aminotransferase (AST) ≥ 5 × upper limit of normal (ULN), alanine aminotransferase (ALT) ≥ 5 × upper limit of normal (ULN), alkaline phosphatase (ALP) ≥ 2 × ULN, total bilirubin ≥ 2.5 mg/dL or international normalized ratio (INR) ≥ 1.5 [10, 11]. Following this, patients were excluded if they had an ICD-10 code corresponding to Acetaminophen hepatotoxicity (T39.1), prior bone marrow transplantation (Z94.81), or prior liver transplantation (Z94.4).

Table 2. Individual T-codes associated with the probable-DILI and non-DILI cases.

From: Combining K-72 Hepatic Failure with 15 Individual T-Codes to Identify Patients with Idiosyncratic Drug-Induced Liver Injury in the Electronic Medical Record

ICD-10 T-codes	Probable-DILI N = 37 (%)	Non-DILI N = 308
T36-systemic antibiotics	11 (29.7)	28 (9.1)
T37-systemic anti-infectives and anti-parasitics	6 (16.2)	5 (1.6)
T38-hormones, synthetic substitutes and antagonists	5 (13.5)	33 (10.7)
T39-non-opioid analgesicsa	1 (2.7)	7 (2.3)
T40-narcotics and psychodysleptics, hallucinogens	0	54 (17.5)
T41-anesthetics and therapeutic gases	0	7 (2.3)
T42-antiepileptic, sedative/hypnotic	3 (8.1)	7 (2.3)
T43-psychotropic drugs, NOS	2 (5.4)	7 (2.3)
T44-drugs affecting autonomic nervous system	0	5 (1.6)
T45-hematologic drugs	2 (5.4)	81 (26.3)
T46-cardiovascular drugs	1 (2.7)	12 (3.9)
T47-drugs affect GI system	0	12 (3.9)
T48-drugs affect muscle, respiratory	0	0
T50-diuretics and other drugs, NOS	5 (13.5)	49 (15.9)
T65-Toxic effect of unspecified drugs	1 (2.7)	1 (0.3)

DILI drug-induced liver injury, NOS not otherwise stated, GI gastrointestinal

^aAll acetaminophen hepatotoxicity cases were excluded using ICD-10 code T39.1 for 4-aminophenol derivatives and on manual chart review

Data Extraction

Using manual chart review, baseline demographic information including age, sex, race, and ethnicity were obtained for each case. In addition, data regarding the dose and duration of the suspect drug, liver biochemistries at the time of liver injury, laboratory testing for alternative etiologies of liver injury (hepatitis A IgM, hepatitis B surface antigen, hepatitis B core IgM, hepatitis C antibody, anti-nuclear antibody, and anti-smooth muscle antibody), liver imaging at the time of injury, and liver biopsy results (if performed) were also collected.

Assessment of Liver Injury Pattern and Severity

The pattern of liver injury was defined using the R ratio, which is calculated as (ALT/ALT ULN)/(ALP/ALP ULN). R ratio values of ≥ 5, ≤ 2, and < 5 but > 2, were characterized as hepatocellular, cholestatic, or mixed, respectively [11]. DILI case severity was classified on a scale from 1 to 5: (1) mild, elevated ALT and/or ALP levels but bilirubin < 2.5 mg/dL and INR < 1.5; (2) moderate, elevated ALT and/or ALP levels and bilirubin ≥ 2.5 mg/dL or INR ≥ 1.5; (3) moderate-severe, elevated ALT, ALP, bilirubin and/or INR levels and the patient was hospitalized or an ongoing hospitalization was prolonged in the setting of DILI; (4) severe, elevated ALT and/or ALP levels and bilirubin ≥ 2.5 mg/dL and at least one of the following occurred: (a) hepatic failure (INR ≥ 1.5, ascites or encephalopathy), (b) other organ failure determined secondary to DILI; and (5) fatal, the patient died or underwent liver transplantation due to DILI [11]. For comparison, non-DILI cases were also given a DILIN case severity score.

Each case was assigned a DILIN expert opinion causality score, which ranges from 1 (definite) to 5 (unlikely): (1) definite ≥ 95% likelihood, (2) highly likely = 75–95% likelihood, (3) probable = 50–74% likelihood, (4) possible = 25–49% likelihood, and (5) unlikely = < 25% likelihood [12]. Cases were deemed “probable DILI” when their respective DILIN causality score was either definite (1), highly likely (2), or probable (3). Cases with liver injury from a non-DILI etiology were classified as “non-DILI” and had an expert opinion causality score of possible (4) or unlikely (5).

Statistical Analysis

Descriptive data were presented as number (percent) for categorical data and mean (standard deviation, SD) or median (interquartile range, IQR) for continuous data. Differences between categorical data and continuous data were assessed using the Chi-squared test and the Student’s t test, respectively. We performed a post hoc analysis using only the K72.0 code in the algorithm. Significance was determined using p values < 0.05. Statistical analysis was performed using R statistical software.

Results

Electronic Medical Record Search

Between October 1st, 2015, and September 30th, 2018, 1,212,402 unique patients had either an inpatient or outpatient encounter in the EMR (Fig. 1). A total of 29,457 had at least one T-code recorded, and 3,660 had a K-72 code recorded, and there were 661 total encounters that had both T-codes and K-72. The K72.9 code for Hepatic failure unspecified was more common than the K72.1 (chronic hepatic failure) or the K72.0 (acute and subacute hepatic failure) codes (Supplemental Table 1). Among these patients, 172 did not meet the predefined laboratory criteria leading to a cohort of 489 potential patients. After excluding duplicate patients and those who had undergone a prior bone marrow or liver transplant as well as those with acetaminophen overdose, 345 patients were eligible for this analysis.

Fig. 1.

Patient population. The EMR at a large referral center was searched using a combination of K72 and T-codes. After filtering by laboratory criteria and excluding patients with a prior liver or bone marrow transplant, a total of 345 unique patients were identified

Clinical Features

The median age of the cohort was 57 years (IQR 43–66), 184 (53%) male, and 307 (89%) Caucasian. After manual chart review and application of the expert opinion causality method, 37 (10.7%) cases were determined to be probable-DILI cases with a DILIN causality score of 1, 2, or 3, and 308 (89.3%) cases were categorized as non-DILI cases with an alternative diagnosis for their liver injury encounter (Table 1). The probable-DILI cases were significantly more likely to have a hepatocellular injury pattern (57% vs 32.5%, p = 0.01) at the onset with a higher mean ALT (1507 vs 531 IU/L, p = 0.04) and a trend toward a higher mean AST as well (1526 vs 716 IU/l, p = 0.09). In addition, total bilirubin levels and the severity scores were significantly higher in the probable-DILI patients compared to the non-DILI patients. However, patient age, sex, and ethnicity were similar in the two groups.

Table 1. Clinical characteristics of the probable-DILI and non-DILI cases.

From: Combining K-72 Hepatic Failure with 15 Individual T-Codes to Identify Patients with Idiosyncratic Drug-Induced Liver Injury in the Electronic Medical Record

	Probable-DILI N = 37 (%)	Non-DILI N = 308 (%)	P value
Age (years)	50.4 ± 19.4	53.9 ± 17.4	0.29
Female	18 (48.6)	143 (46.4)	0.94
Race
Caucasian	31 (83.8)	276 (89.6)	0.04
African American	1 (2.7)	20 (6.5)
Asian	2 (5.4)	2 (0.6)
American Indian or Alaska Native	0	1 (0.3)
Other/unknown	3 (8.1)	9 (2.9)
Initial laboratory values
AST (IU/1)	1526 ± 2774	716 ± 1726	0.09
ALT (IU/1)	1507 ± 2717	531 ± 1188	0.04
ALP (IU/1)	364 ± 437	242 ± 309	0.11
Bilirubin (mg/dl)	7.7 ± 8.1	4.6 ± 6.7	0.03
INR	1.6 ± 1.0	1.8 ± 1.3	0.29
R-value at onset
< 2 = Cholestatic	13 (35.1)	166 (53.8)	0.01
2–5 = Mixed	3 (8.1)	42 (13.6)
> 5 = Hepatocellular	21 (56.8)	100 (32.5)
DILIN Severity Index
Mild (1)	0	8 (2.6)	< 0.01
Moderate (2)	2 (5.4)	12 (3.9)
Moderate-severe (3)	19 (51.3)	276 (89.6)
Severe (4)	10 (27.0)	11 (3.6)
Fatal (5)	6 (16.2)	1 (0.3)

Values expressed as mean ± standard deviation or number (Percent)

DILI drug-induced liver injury, AST aspartate aminotransferase, ALT alanine aminotransferase, ALP alkaline phosphatase, INR international normalizad ratio, DILIN drug-induced liver injury network

Probable-DILI Cases

The most common drug classes identified in the probable-DILI cases were systemic antibiotics (T-36, 29.7%), other systemic anti-infectives and anti-parasitics (T-37, 16.2%), hormones (T-38, 13.5%), and diuretics and other drugs (T-50, 13.5%) (Table 2). The most commonly identified individual agents were all antibiotics and included trimethoprim-sulfamethoxazole (8%), ceftriaxone (8%), amoxicillin-clavulanate (5%), isoniazid (5%), and azithromycin (5%). Only five of the probable-DILI cases had evidence of pre-existing liver disease prior to their DILI diagnosis (three with nonalcoholic fatty liver disease-associated cirrhosis and two with hepatitis C virus-related cirrhosis). Of the 37 probable-DILI cases, 30 (81.1%) were previously identified using the algorithm that combined K-71 and T-codes. [8] Twenty-four of the probable-DILI cases had the K72.0 code (acute and subacute hepatic failure) alone, while eight had the K72.9 code (hepatic failure unspecified) and five had both K72.0 and K72.9 codes.

Non-DILI Cases

Among the 308 non-DILI cases, the most frequent alternative etiologies of liver injury were ischemic hepatitis (37.7%), cancerous replacement of the liver (14.6%), and alcohol-related liver disease (10.1%) (Fig. 2).

Fig. 2.

Adjudicated causes of liver injury in the 308 non-DILI cases. NAFLD Nonalcoholic fatty liver disease

K72.0 Post hoc Analysis

Given that 29 of the probable-DILI cases had an associated K72.0 code, we performed a post hoc analysis where only the K72.0 code was used in the searching algorithm (rather than all 3 K72 codes) combined with T-codes and laboratory criteria. The positive predictive value using the K72.0 only algorithm improved to 15% (29 of 193).

Discussion

Idiosyncratic DILI remains largely a clinical diagnosis of exclusion due to the lack of a confirmatory, objective laboratory test. Although recent studies involving in vitro test systems derived from peripheral lymphocytes of DILI patients and various serum biomarkers show promise, further validation is required [13, 14]. The diagnosis of DILI is therefore based on a robust clinical assessment that includes medication review, radiographic and laboratory evaluation, and the elimination of more common causes of liver injury [15, 16]. However, to improve our understanding of DILI pathogenesis and the role of genetic, environmental, and immunological factors, it is important to be able to reliably identify clinical cases from administrative databases [17].

We previously illustrated the effectiveness of searching the EMR for DILI cases using the combination of toxic liver disease codes (K71) and drug poisoning T-codes [8]. In the current study, an algorithm using three hepatic failure codes (K72.0, K72.1, K72.9) combined with T-codes over the same time period identified 345 cases of which 37 were probable DILI compared to the prior algorithm that identified 182 cases of which 121 were probable DILI. Therefore, the algorithm used in the current study had a PPV of only 10.7% for identifying probable-DILI cases compared to 66.5% using the prior method. Importantly, the vast majority of DILI cases found using the K72 algorithm were previously identified with the K71 algorithm. These findings support the robustness of the K71 algorithm in identifying DILI in the EMR. One explanation for this large difference in the PPV of the algorithms was the use in the current study of ICD-10 K72 codes that have proven useful in identifying acute liver injury from a multitude of etiologies [18]. We speculate that the more severe DILI severity scores observed in the probable-DILI cases (Table 1) is due to the fact that these patients had a bonafide acute DILI episode which tends to be characterized by high serum ALT and bilirubin elevations at onset. In contrast, the non-DILI cases frequently did not actually have an acute liver injury episode but rather chronic underlying liver disease like alcoholic cirrhosis, congestive hepatopathy from longstanding CHF, or progressive metastatic cancer which are known to be associated with lower ALT and bilirubin elevations. However, to our knowledge, the K72 codes have not been used to identify cases of idiosyncratic drug-induced liver injury and our study adds to the literature by illustrating that these three codes when combined together have a low yield for identifying DILI in an EMR. Indeed, the positive predictive value of our searching algorithm (10.5%) was similar to the pooled positive predictive value of 14.6% seen in prior algorithms using the less specific ICD-9 codes [19]. Among the K72 codes deployed, the K72.0 (acute/subacute hepatic failure) and K72.9 (unspecified hepatic failure) yielded the most cases (29 and 13, respectively). However, when we repeated our analyses using only the K72.0 code, the positive predictive value remained low at 15%. However, this represents a significant improvement in efficiency since only 193 cases needed to be manually reviewed compared to 345 from Fig. 1. Since the metabolic dysfunction associated with severe hepatic disease (hepatic failure) increases the risk of drug side effects, it is likely our algorithm was identifying mostly non-hepatic adverse drug reactions rather than specific instances of drug hepatotoxicity [20]. In support of this, the majority of the non-DILI cases did have severe liver dysfunction, but this was largely attributed to hepatic ischemia, malignant infiltration of the liver, or sepsis (Fig. 2). It is possible that modifying the search to exclude encounters that also had an ICD-10 code for alcoholic liver disease, congestive hepatopathy, or metastatic cancer could have reduced the total number of non-DILI cases but at the expense of eliminating some bonafide DILI cases. Going forward, it may be worthwhile for other investigators to consider the impact of eliminating cases by using ICD-10 codes for various chronic liver diseases or other causes of liver injury.

There are several limitations regarding the current study. First, cases were retrospectively adjudicated by the study authors and the probability of DILI was determined using the semi-quantitative expert opinion DILIN method. To maximize the reliability, we had three individuals review and score the cases. While these methods may limit the generalizability of our findings, causality assessment by expert opinion is considered a more accurate means of identifying DILI cases compared to standardized instruments like the RUCAM [21, 22]. Secondly, the success of any algorithm using administrative billing codes is dependent on the accuracy and reliability of the initial coding assigned to the encounter. In a study assessing inter-observer reliability of ICD-10 codes for circulatory causes of death, the kappa coefficient was 0.60 among two individual coding offices reviewing the same death records [23]. Therefore, not only is DILI frequently unrecognized by practicing physicians, it also relies on consistent and accurate coding by medical billers to be recorded in the EMR [24].

In conclusion, a novel application of an algorithm using K72 codes for hepatic failure combined with drug toxicity/poisoning T-codes had a low yield for bonafide DILI cases. A large number of non-DILI cases were identified presumably due to the fact that K72 tracks well for liver injury broadly including chronic liver disease with various non-hepatic adverse drug events but less so for DILI. A small improvement in the PPV was found when we restricted the searching to subjects with a K72.0 code alone rather than all three K 72 codes (15% vs 10.5%). This refined search strategy significantly reduces the need for manual chart review (from 345 to 193 cases) although a small number of bonafide DILI cases (eight) would not be identified. We recommend that other investigators also consider conducting a sensitivity analysis involving an exclusionary step for ICD-10 codes of various competing etiologies of liver disease (e.g., alcoholic cirrhosis, metastatic cancer) to determine whether the search efficiency may further improve but at the potential loss of bonafide DILI cases which are notoriously difficult to identify in the EMR.

Abbreviations

ALF

Acute liver failure

ALP

Alkaline phosphatase

ALT

Alanine aminotransferase

AST

Aspartate aminotransferase

DILI

Drug-induced liver injury

DILIN

Drug-induced liver injury network

EMR

Electronic medical record

ICD

International classification of diseases

INR

International normalized ratio

IQR

Interquartile range

SD

Standard deviation

ULN

Upper limit of normal