Identifying patients with hepatic encephalopathy using administrative data in the ICD-10 era

Introduction

Hepatic Encephalopathy is a common complication of cirrhosis resulting in relapsing-remitting mental status changes ranging from deficits in executive function to coma. Incident HE is associated with an abrupt increase in mortality¹ and frequent hospitalization.² To further our understanding of the burden and impact of HE at the population level, we require valid algorithms to identify patients in administrative data. An ICD-9 code is specific for HE (572.2), offering a 0.92 positive predictive value (PPV) and 0.36 negative predictive value (NPV).³ When applied in an algorithm to patients with ICD-9 codes for cirrhosis (e.g. 571.5), Kanwal found that a PPV and NPV of 0.86 and 0.87.⁴ Unfortunately, the switch to ICD-10 in 2015 rendered algorithms validated using ICD-9 invalid. Kaplan previously showed that lactulose and rifaximin use correlated with grade of HE for Child classification.⁵ Herein, we validate a diagnostic coding algorithm for HE using ICD-10 and medication-records.

Methods

We prospectively enrolled 300 persons with Child A-B cirrhosis and portal hypertension but no current or prior history of HE from 2016–2017 in a study designed to yield predictors of HE approved by the University of Michigan Institutional Review Board. As published elsewhere,⁶ all patients had definitive diagnoses of (as well as diagnostic codes – K70.3,K74.6,K74.69 – for) cirrhosis and were followed for up to 3 years. Overt HE was defined according to the West-Haven Grade (grades≥2) by the treating transplant hepatologist. Patients were called or evaluated every 3 months for HE and medication changes. When overt HE was reported at an outside facility, records were retrieved, reviewed by two independent investigators, and defined by gross disorientation or coma with response to medical therapy. We surveyed all ICD-10 billing-codes generated for each patient during follow up. We sought to evaluate the test characteristics of diagnostic codes and medication records (any prescription for lactulose or rifaximin) for the presence of HE. Candidate codes were selected according to default codes for HE and a manual review of the codes registered in 100 consecutive records of discharged patients. These included K72.90 (hepatic failure), K72.91 (hepatic failure with coma), G93.40 (encephalopathy, NOS), G93.49 (Other encephalopathy). We then validated our findings in an external cohort of 300 persons with cirrhosis chosen randomly from a 42,575 veterans with hepatitis C identified from 2000–2016 and followed until 2019.(see Supplement for additional details).

Results

Overall our cohort was aged 60(52–66) years, 56.3% male, and 70% Child class A. All patients had portal hypertension, 76% had varices, and 41% had a history of ascites (predominantly well controlled). The median MELD-Na score was 9 (7–13). During follow-up of 538±263 days, 68 (22.7%) developed HE within 228±208 days. The ICD-10 code which best identified patients who developed HE was K72.90 while the optimal strategy was to use recorded lactulose or rifaximin prescriptions.(Table 1) In our validation cohort, 37(12.3%) developed HE. Diagnostic algorithms performed with equivalent characteristics. However, positive predictive values for lactulose and lactulose-or-rifaximin use (0.73 and 0.71) were lower. Lactulose was used in 5(1.7%) and 13(4.3%) of patients without HE in the prospective and retrospective cohorts.

Table 1:

Performance of ICD-10 Codes and Medications for the Identification of Persons with Hepatic Encephalopathy

Codes/Medications	Cohort	Proportion of exposed with HE	Sensitivity (95% CI)	Specificity (95% CI)	PPV (95% CI)	NPV (95% CI)	AUROC
K72.90	UM	28/33	0.41 (0.30–0.53)	0.98 (0.95–0.99)	0.85 (0.69–0.93)	0.85 (0.80–0.89)	0.70
	VA	17/19	0.46 (0.31 – 0.62)	0.99 (0.97 – 1.0)	0.90 (0.69 – 0.97)	0.93 (0.89 – 0.95)	0.73
K72.91	UM	0/1	0	0.99 (0.98–1.00)	0	1.00 (0.98–1.00)	0.50
	VA	1/1	0.03 (0.01 – 0.14)	1.0 (0.99 – 1.0)	1.0 (0.21 – 1.0)	0.88 (0.84 – 0.91)	0.51
G93.40	UM	0/0	N/a	N/a	N/a	N/a	N/a
	VA	6/8	0.16 (0.08 – 0.31)	0.99 (0.97 – 1.0)	0.75 (0.41 – 0.93)	0.89 (0.85 – 0.92)	0.58
G93.49	UM	0/0	N/a	N/a	N/a	N/a	N/a
	VA	0/0	N/a	N/a	N/a	N/a	N/a
Lactulose	UM	52/57	0.77 (0.65 – 0.85)	0.98 (0.95 – 0.99)	0.91 (0.81 – 0.96)	0.93 (0.90 – 0.96)	0.87
	VA	35/48	0.95 (0.82 – 0.99)	0.95 (0.92 – 0.97)	0.73 (0.59 – 0.83)	0.99 (0.97 – 1.0)	0.94
Rifaximin	UM	25/26	0.37 (0.26 – 0.49)	0.99 (0.98 – 1.00)	0.96 (0.81 – 0.99)	0.84 (0.80 – 0.88)	0.68
	VA	20/23	0.54 (0.38 – 0.69)	0.99 (0.97–1.0)	0.87 (0.68 – 0.96)	0.94 (0.90 – 0.96)	0.76
Lactulose or rfiaximin	UM	64/66	0.94 (0.85 – 0.98)	0.96 (0.93 – 0.98)	0.88 (0.78 – 0.93)	0.98 (0.96 – 0.99)	0.95
	VA	35/49	0.95 (0.82 – 0.99)	0.95 (0.92 – 0.97)	0.71 (0.58 – 0.82)	0.99 (0.97 – 1.0)	0.77
Lactulose and rifaxmin	UM	20/20	0.29 (0.20 – 0.40)	1.0 (0.98 – 1.00)	1.0 (0.84 – 1.00)	0.83 (0.78 – 0.87)	0.65
	VA	29/22	0.54 (0.38 – 0.69)	0.99 (0.97 – 1.0)	0.91 (0.72 – 0.98)	0.94 (0.90 – 0.96)	0.95

UM = University of Michigan prospective cohort. VA = Veterans Affairs, retrospective random sample from national cohort. NPV = negative predictive value, PPV = positive predictive value

Discussion

We find that lactulose-or-rifaximin use in patients with codes for cirrhosis is the optimal strategy for the identification of HE in administrative data. Previously the Kanwal group attempted to validate ICD-10 codes for HE prospectively but could not for lack of incident HE.⁷ Combining a prospective study designed to predict HE and a large, national sample from two different health-systems, we have overcome this gap. Although K72.90 (hepatic failure) identifies patients with HE with the same PPV and NPV offered by ICD-9 572.2, it also reproduces the same suboptimal insensitivity (41%). We find that, in persons with cirrhosis, recorded use of lactulose or rifaximin in the medical record is an accurate method to identify HE. This strategy can reliably identify patients with HE in any database that includes pharmacy linkage.

Contextual factors

These data must be interpreted in the context of the study design. First, the performance of specific codes may vary between centers given differing software platforms or local coding decisions. Second, K72.90 was the best performing code even though it is agnostic to the underlying etiology, lacks any specific mention of HE in its name, and is likely susceptible to variable interpretations across centers. Third, some patients, roughly 2–4% both samples, received lactulose for reasons other than HE. This accounts for the lower performance of lactulose-based identification for Veterans.

Conclusion

These data empower the use of administrative data with medication records for the identification and study of contemporary patients with HE.