Accurately identifying cirrhosis and its complications to create the novel statewide Indiana digital cirrhosis registry

Abstract

Administrative datasets are important for cirrhosis research but limited by suboptimal cirrhosis identification. We developed and validated algorithms to accurately identify cirrhosis and its complications in real-world, statewide dataset. From 2017 to 2020 Indiana Patient Care Network data, 15,636 records were grouped by combinations of code and lab criteria (group A: cirrhosis codes, B: FIB-4/APRI criteria, C: cirrhosis complication codes, D: code/lab for liver disease). Diagnoses were confirmed by chart review in 4.5% of 15,636 records. Positive predictive values (PPV) were calculated for various algorithms which were externally validated in hepatology clinic (n = 1,039) and emergency department-based cirrhosis cohorts (n = 2,124). Charts meeting criteria for group A and at least one other group (“AX”, e.g., ABC) had an overall PPV of 86%. Highest PPVs were seen in ACD and ABCD and confirmed during external validation: 88% and 97% (hepatology cohort), 79% and 93% (ED cohort). Without complication codes, ABD showed strong PPVs: 86%(internal), 92%(hepatology), 72%(ED). ICD-10-based definitions alone were suboptimal for complications: ascites (57%), hepatic encephalopathy (HE:55%). PPV for HE was improved with addition of medications but remained < 80%. Taken together, we provide algorithms to identify both compensated and decompensated cirrhosis in real-world data. Using the “AX” algorithm, we created the statewide Indiana Digital Cirrhosis Cohort to support future research across cirrhosis stages.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-39585-2.

Introduction

The prevalence of cirrhosis has increased to becoming a top-ten cause of death and a major contributor to healthcare utilization^1–3. Chronic liver disease related hospitalization costs were estimated to be around $81 billion from 2012 to 2016 with cirrhosis related hospitalizations accounting for $7.37 billion in expenses in 2014^4,5. In parallel, there are growing number of administrative data sources owing to the advent of electronic health records (EHR) which can facilitate health services research (HSR) to understand and address the epidemic of liver disease^6–8. Within the world of “big data”, administrative databases currently rely on the International Classification of Diseases (ICD) for disease classification to parse patient populations and identify outcomes of interest⁹.

In cirrhosis-based HSR, administrative data, often comprised of large databases spanning multiple hospital systems with long-term longitudinal follow-up have been a rich resource to understand the burden of liver disease define drivers of poor cirrhosis- related outcomes and measure the impact of policy change on these outcomes^5,10–13. The validity of these studies relies on accuracy in identifying individuals with cirrhosis; however, prior studies have raised concern about the performance of ICD-based algorithms in multiple administrative and EHR-based datasets^14–16. A 2022 systemic review assessed 18 publications looking at different ICD code combinations to arrive at a consensus of codes which constituted a combination of nine of the most used ICD-9/10 codes and externally validated the positive predictive value in UK and US cohorts. The range in positive predictive value (PPV) was noted to be from 70 to 100% highlighting that PPV calculation is dependent on the disease prevalence in the population and validation populations need to be carefully selected^17,18. This review also highlighted that prior validation studies of ICD-based cirrhosis algorithms have largely been limited to single-center cohorts, specialty clinics, national datasets with narrow population characteristics or using ICD-9 codes and pre-2015 data, leaving a critical gap in assessing their generalizability and performance within diverse, real-world health systems. To date, no studies have used a contemporary, state-wide administrative dataset to test the generalizability of ICD-10 based cirrhosis identification.

Additionally, accuracy of current ICD-10 code-based definitions is low in identifying complications of cirrhosis such as hepatic encephalopathy and ascites where ascites ICD codes are particularly poor with PPV between 43% and 63% even when combined with other codes^17,19,20.

Given the worsening healthcare burden of cirrhosis, and the need to leverage large administrative databases for HSR, we set out (1) to create an algorithm to identify individuals with cirrhosis and cirrhosis-related complications using administrative data from a unique statewide data network, the Indiana Digital Cirrhosis Cohort, and (2) to test its performance in 2 separate validation cohorts.

Methods

Cohort creation

This is a retrospective study where we created an internal discovery cohort of all individuals 18 and older who had at least one encounter within the Indiana Network for Patient Care (INPC) from January 1, 2017, to December 31, 2020 and fulfilled inclusion and exclusion criteria described below^21,22. INPC is the largest and most comprehensive regional health information exchange covering approximately two-thirds of Indiana’s population. Research access to this data is managed by the Regenstrief Institute. Established in 1995, it includes data from over 100 separate healthcare entities providing data which includes: major hospitals, health networks, and insurance providers. When combined, the information from these institutions represent data on over 18 million patients in the form of 10 billion clinical observations, 951 million encounter records, and over 147 million mineable text reports. Within the structured data, the INPC strings data across years, institutions, between patients, and other outside data sources. In addition, each patient record is also geocoded, linked to a census track and associated to community-level socioeconomic variables available through American Community Survey as previously described^23,24.

Inclusion and exclusion criteria

Charts were grouped based on ICD-10 criteria (Table 1). Group A contained codes for cirrhosis diagnosis. Group B contained high risk labs for cirrhosis based on FIB-4 and APRI. Group C contained codes for complications of cirrhosis. Group D contained codes for causes of cirrhosis or labs for viral causes of cirrhosis such as hepatitis C. Participants who solely had an “ascites” diagnosis code (K70.11, K71.51, R18.8) without any other cirrhosis-specific complication code (group C) were excluded after our initial review due to a high prevalence of non-cirrhosis related ascites. For Group B, we calculated FIB-4 and APRI using an algorithm that loops through each individual’s available labs (AST, ALT, & platelet count) in chronological order and pairs lab values for if they occur within a proximity of 90 days to calculate the score. Due to the high negative predictive value (NPV) of these scores, we excluded patients with lab values indicating they are low-risk for cirrhosis (FIB-4 < = 2.67 and APRI score < = 1.0) from Group A unless they also met criteria from Group C or D.^25,26 In our initial work, charts without ICD codes for cirrhosis (not in group A) but with other group criteria (e.g., BC, CD, BCD, etc.) had low PPVs (10%-48%, data not shown) and often had low prevalence in the INPC data (0.6%-3.7%) and were therefore not further validated.

Table 1.

Definitions of criteria used to determine entry into cirrhosis cohort.

Group	Group description	ICD codes or criteria
A*	ICD codes for cirrhosis	K70.30, K70.31, K71.7, K74.3, K74.4, K74.5, K74.6, K74.60, K74.69
B	High risk labs for cirrhosis	FIB4 > 2.67 and/or APRI score > 1.0
C	ICD codes for complications of cirrhosis	Ascites: K70.11, K71.51, R18.8; Hepatic encephalopathy†: B19.0, B19.11, B19.21, K70.31, K71.11, K72.01, K72.11, K72.91; Hepatocellular carcinoma: C22.0, C22.8, C22.9; Hepatorenal syndrome: K76.7; Spontaneous bacterial peritonitis: K65.2; Variceal bleeding: I85.01, 185.11
D	ICD codes or lab studies for diseases that cause cirrhosis	Hepatitis B: B16.0, B16.1, B16.2, B16.9, B17.0, B18.0, B18.1, B19.10, B19.11, Z22.51; Hepatitis C: B17.10, B17.11, B18.2, B19.20, B19.21, Z22.52; Nonalcoholic steatohepatitis‡: E08., E09., E10., E11., E12., E13., E66.01, E66.09, K73.0, K73.1, K73.2, K73.8, K73.9, K74.60, K74.69, K75.81, and K76.0; Wilson’s: E83.00, E83.01, E83.09; Hemochromatosis: E83.110, E83.118; Alpha-1-Antitrypsin Deficiency : E88.01; Alcohol: K70.0, K70.10, K70.11, K70.2, K70.30, K70.31, K70.40, K70.41, K70.9; Autoimmune/Biliary: AIH, PBC, PSC: K74.3, K74.4, K74.5, K75.4, K83.01, K83.09, K83.8, and K83.9. Hepatitis B or C related lab study result labeled abnormal or any values indicative of a positive value (e.g., “confirmed”, “positive”, “reactive”, “detected”)

* Group A participants could not have low risk FIB-4 or APRI labs (if labs were present in the record).

† Additional nonspecific hepatic encephalopathy codes were used after cohort identification and include G92, G93.40, K72.90, R41.0, R41.82. K76.82 was developed in October 2022 and therefore was not applied to this cohort.

^‡ Metabolic dysfunction associated steatohepatitis was defined as having at least one of the following codes: K75.81or K76.0 or having a combination of at least one of the following E08., E09., E10., E11., E12., E13., E66.01, E66.09 with at least one of K73.0, K73.1, K73.2, K73.8, K73.9, K74.60, K74.69, K75.81, and K76.0.

Hence, the ICD-10 code criteria for inclusion into the internal discovery cohort based on our four groups (Table 1) is as follows:

• Diagnosis code for cirrhosis (Group A) without lab data within 2017–2020.

OR.

• Diagnosis code for cirrhosis (Group A) AND at least one of the following within 2017–2020:

  • Either lab values indicating high-risk FIB-4 or APRI score for cirrhosis (Group B) or no lab data (excluding those with lab data indicating they are low-risk for cirrhosis).
  • Diagnosis code for cirrhosis complication (Group C).
  • Diagnosis code or test result for specific liver disease etiology (Group D).

Participants were excluded if their age was less than 18 years old on the index date or they had liver transplant coded prior to their index date, indicated by ICD-10 codes T86.4 and Z94.4 and ICD-9 codes 996.82 and V42.7. An index date was defined as the first date where inclusion criteria codes appeared in a patient’s chart during the study period (2017–2020)^25,26. Based on these inclusion and exclusion criteria our cirrhosis registry cohort included patients that fall within the “AX” group. This is illustrated in Fig. 1 with the groups included in “AX” is highlighted with a black outline.

Fig. 1.

Subsets and subgroups of patients included in the internal discovery cohort. Subgroups that fall within the black circle belong to the AX algorithm that was used for PPV calculations during validation in each cohort. Subgroups B, BD, and D were excluded in the initial criteria, while CX subgroups (C, BC, BCD, CD) were not found to have high enough PPV.

Internal validation of codes to identify cirrhosis

To better understand which EHR features improved cirrhosis identification, we created subgroups of this “AX” cohort based various combinations of groups (Fig. 1). As noted above, subgroups within A (“AX”) are: A with no labs, AB (but not C and D), ABC (but not D), ABCD, ABD (but not C), AC (but not B and D), ACD (but not B), and AD (but not B and C). Subgroups within Group C (“CX”) were created excluding A in a similar fashion (C, BC, BCD, and CD). To test the accuracy of different subgroups at identifying patients with cirrhosis, we manually reviewed each individual’s EHR record to serve as the gold standard for the presence of cirrhosis and its complications. We created a standardized guide with a standardized definition of cirrhosis to minimize discrepancies between our eleven independent reviewers (5 transplant hepatologists, 4 hepatology trained physicians assistants and 2 hepatology fellows) (Supplemental Table 1). For the first 111 charts reviewed, 2 independent reviewers reviewed each record and achieved an overall agreement rate of 85% for presence of cirrhosis with a Cohen’s kappa of 0.68 (Supplemental Table 2). In the 15% where there was disagreement, a third reviewer reviewed the charts to break the tie. The disagreements were largely due to missing data or overlap with alcohol related hepatitis leaving the diagnosis of cirrhosis less certain. Then another 594 records were adjudicated by single blinded reviewers (total internal validation cohort n = 705).

Validation of AX algorithms identify cirrhosis in two separate cohorts

We also compared the accuracy of different subgroups at identifying patients with cirrhosis in 2 independent cohorts (Table 2). The first cohort is that of outpatient hepatology clinic patients at Indiana University Health which serves as a cohort enriched with cirrhosis²⁷. This cohort included 1039 patients seen between 2019 and 2021. Cirrhosis was confirmed in this cohort by a detailed chart review that assessed histology or characteristic clinical, lab, imaging, and endoscopic findings. The second cohort is that of all-comers to any one of 18 emergency department at Indiana University Health during 2021 which serves as a real-world cohort²⁴. This cohort included 2142 patients whose cirrhosis diagnosis was confirmed through chart review by hepatology providers based on a combination of clinical evaluation by gastroenterology/hepatology providers, imaging modalities such as ultrasonography, transient elastography or CT scan revealing characteristic features of nodular liver with ascites or splenomegaly, and liver biopsy. The details of these cohorts are described in their respective publications and Table 2^24,27.

Table 2.

Characteristics of discovery and validation cohorts.

Variables	Internal discovery cohort# N = 14254)	Hepatology clinic cohort (N = 1039)	Emergency department cohort (N = 2124)
Age (years, median, IQR)	(60 (52–68)	58.8 ± 11.8	61.3 ± 12.9
Gender - Female (N,%)	6559 (46.0%)	488 (47%)	994 (47%)
Race
White (N,%)	12,421 (87.1%)	891 (85.8%)	1900 (91%)
Black or African American (N, %)	1264 (8.9%)	66 (6.4%)	195 (9%)*
Asian	174 (1.2%)	–	–
Hispanic Ethnicity	601 (4.2%)	–	–
Insurance/Payer
Commercial	3832 (26.9%)	346 (33.3%)	597 (28%)
Medicaid	6305 (44.2%)	197 (19.0%)	420 (20%)
Medicare	3298 (23.1%)	432 (41.6%)	1107 (52%)
Etiology of cirrhosis
Hep-C/Hep-B	3816 (26.7%)	224 (21.9%)	448 (21%)
AIH/PBC-PSC	1851 (13%)	–	–
Alcohol	5278 (37%)	277 (27.0%)	803 (38%)
MASLD	9661 (67.8%)	378 (36.9%)	563 (27%)
Hemochromatosis/Wilsons/A1AT	342 (2.4%)	–	–
Comorbid Conditions
Alcohol Use Disorder (AUD)	4176 (29.3%)	134 (12.9%)	–
Coronary Artery Disease (CAD)	4580 (32.1%)	106 (10.2%)	–
Type 2 diabetes	6532 (45.8%)	252 (24.2%)	–
Obesity	2330 (16.3%)	95 (9.1%)	–
Substance Use Disorder (SUD)†	2330 (16.3%)	38 (3.7%)	–
MELD 3.0 score	14 (10–19)	13.3 (6.2)‡	17.0 (11.0–23.6)
Cirrhosis Complications
Ascites	5798 (40.7%)	565 (54.4%)	1029 (49%)
Hepatic Encephalopathy	5963 (41.8%)	387 (37.2%)	675 (32%)
Variceal Bleeding	1070 (7.5%)	–	720 (34%)§
Spontaneous Bacterial Peritonitis	754 (5.3%)	–	–
Hepatorenal Syndrome	598 (4.2%)	–	–
Hepatocellular Carcinoma	1592 (11.2%)	85 (8.2%)	–

# Individual records meeting “AX” criteria (membership to group A and at least one other group, i.e., AB or ACD).

* Non-white race.

† Includes AUD and SUD.

‡ Mean (SD).

§ All patients with varices regardless of bleeding.

Validation of codes to identify hepatic encephalopathy and Ascites related to cirrhosis

We next determined the accuracy of ICD-10 codes in identifying HE and ascites. EHR-based HE was identified using ICD-10 codes listed in Table 1. In addition, due to known poor performance of these codes¹⁹, we also developed a secondary definition of HE based on medication orders for lactulose or rifaximin or both. These combinations were then compared to the results of manual chart review to calculate the PPV for identifying HE. Ascites was identified using codes K70.11 Alcoholic hepatitis with ascites, K71.51 Toxic liver disease with chronic active hepatitis with ascites, and/or R18.8 Other ascites. These combinations were then compared to the results of manual chart review to calculate the PPV for identifying ascites.

Statistical analysis

Descriptive statistics including means with standard deviations, medians with interquartile range and portions were calculated for the cohort’s characteristics as appropriate. Interobserver reliability of cirrhosis by chart review was estimated by the overall proportion of agreement between any two chart reviewers. Positive predictive values with standard errors were calculated as the proportion of true positives (according to chart review) among all identified as cirrhosis cases for each subgroup of interest.

Ethics declarations

All research was conducted in accordance with both the Declarations of Helsinki and Istanbul and was approved by Indiana University Institutional Review Board (IRB). The Indiana University IRB also waived the need for informed consent due to the retrospective nature of this study.

Code availability

The methods described above and Table 1 delineate the combination of ICD10 codes used to the various subgroups evaluated as definitions of cirrhosis. The Regenstrief Institute may not share any of the code used by Regenstrief’s Data Core for data processing and analysis. While the code may not be shared, fully de-identified or aggregate datasets may be shared on request providing the following conditions are met: (1) only the minimum necessary data points are shared; (2) the data is strictly de-identified, meaning it includes only aggregate counts and does not include any date fields; (3) the data set to be shared has gone through an Honest Data Broker review by the Data Core and (4) an agreement is put in place with Regenstrief specific to this sharing. There may be charges associated with putting this agreement in place.

Results

Cohort description

Overall, 14,254 individuals’ record fulfilled the inclusion criteria using the “AX” algorithm (Table 2). Median age was 60 years (IQR 52–68) with females making up 46% of the cohort. The majority were white (87%) and 9% were black. 44% had Medicaid insurance while commercial insurance and Medicare were 27% and 23% respectively. MASH was the most common etiology of cirrhosis with 30% followed by alcohol at 26% and viral hepatitis at 24%. Median MELD 3.0 was 14. Hepatic encephalopathy was the most common complication of cirrhosis with a 15.3% prevalence followed by ascites at 14.9%. Figure 1 shows prevalence of group membership in the cohort with Group ABD and ABCD being the most common (37.2% and 36.5%, respectively) followed by AD (12.2%). The remaining groups were prevalent in < 10% of the cohort. These group prevalences were also observed in the 2 external validation cohorts (data not shown).

Positive predictive value for cirrhosis identification

The PPVs across different subgroups are illustrated in Fig. 2 and detailed in Supplemental Table 3. The overall positive predictive value (PPV) in our cohort was 86%. The three subgroups with the highest PPV during internal validation were AC (cirrhosis specific code with a complication code), ACD (adds liver disease code to the AC group) and ABCD (adds high risk FIB-4 or APRI to the ACD group) achieved PPV of 100%, 100%, and 92%, respectively. For the AC group, the PPV remained 100% in the hepatology clinic cohort but decreased to 50% in the real-world ED cohort (SE 0.35, the AC group had low prevalence of 0.1% in the ED cohort). For the ACD and ABCD groups, PPV for the hepatology clinic cohort was 88% and 97% but declined to 79% and 93% in the ED cohort. While the ABC group did not perform well on internal validation with a PPV of 75% (SE 0.13), it had excellent PPVs in the hepatology clinic cohort (97%) and the ED cohort (82%). For groups that may identify compensated cirrhosis such as AB, ABD and AD, the AD group did not perform well in any well in the cohorts while ABD had PPV of 86% in the internal cohort, 92% in the hepatology cohort and 72% in the ED cohort.

Fig. 2.

Positive predictive value for detection of cirrhosis in different cohorts among the different AX algorithms.

Positive predictive value for cirrhosis complications identification

Table 3 describes the PPV of ICD-10 based ascites definition and possible definitions of hepatic encephalopathy using ICD-10 and medication data. Of the 705 manually reviewed charts, ascites was present in 153 records by chart review (39%) and PPV for the ICD-10 based definition was 57%. HE was present based on chart review in 98 records (14%). The overall best performing was definitions that required the prescription of lactulose or rifaximin paired with an ICD-10 code or those where the record had dual prescription for lactulose and rifaximin with PPVs ranging from 68 to 76%. For example, the PPVs of 73% (0.06) for ICD-10 encephalopathy codes and lactulose and rifaximin and 76% for lactulose and rifaximin prescription while the PPV for an encephalopathy ICD-10 code alone was 55% and for an ICD-10 code or a prescription for lactulose or rifaximin was 53%.

Table 3.

Positive predictive value for definitions of Ascites and hepatic encephalopathy.

Criteria	PPV	SE	N*
Ascites code	0.56	0.03	104
Any encephalopathy code†	0.55	0.05	120
Lactulose only	0.64	0.05	98
Rifaximin only	0.70	0.05	74
Lactulose or Rifaximin	0.61	0.05	106
Lactulose and Rifaximin	0.76	0.05	66
Any encephalopathy code and Lactulose	0.69	0.05	71
Any encephalopathy code or Lactulose	0.54	0.04	147
Any encephalopathy code and Rifaximin	0.71	0.06	58
Any encephalopathy code or Rifaximin	0.57	0.04	136
Any encephalopathy code and (Lactulose or Rifaximin)	0.68	0.05	74
Any encephalopathy code and (Lactulose and Rifaximin)	0.73	0.06	55
Any encephalopathy code or (Lactulose or Rifaximin)	0.53	0.04	152

* Number meeting criteria (“exposed”).

† Encephalopathy codes are all codes for hepatic encephalopathy listed in Table 1(group C) and non-specific encephalopathy codes (G92, G93.40, K72.90, R41.0, R41.82). K76.82 was developed in October 2022 and therefore was not applied to this cohort to identify encephalopathy.

Discussion

We used 3 large, contemporary datasets to develop and validate algorithms to identify patients with cirrhosis. This includes a nearly 15,000-person statewide large administrative database, a hepatology clinic-based cohort of 1,039 patients and 2,124 patients with cirrhosis visiting the emergency department. Building on prior work in this field, our study provides guidance on which algorithms are best fit for identifying a cirrhosis cohort based on the pre-test likelihood of cirrhosis and availability of laboratory and/or pharmacy data. We found that groups with ICD-10 codes for the cirrhosis diagnosis with an additional feature of liver disease achieved adequate PPVs even in those without decompensations related to cirrhosis which have previously been a key feature of ICD code-based cirrhosis identification. ICD-10 based definitions of ascites and hepatic encephalopathy have marginal PPVs which improved with medication order data.

Our study is the first to use a state-wide population to test the performance of algorithms using multiple data elements to identify patients with cirrhosis. In line with prior literature, we show that combinations of groups of ICD codes outperformed single groups of codes at identifying those with cirrhosis in both our internal validation cohort and the two external validation cohorts^{14–17,28–30}. For example, a study using random forest and variable of importance analysis showed that code combinations outperform single codes to identify cirrhosis¹⁵. Notably, we had over 90% PPV when combinations of criteria were used, including AC, ACD and ABCD in the internal validation group. This result is similar but much more robust and comprehensive compared to a recent systematic review that combined ICD codes from different studies to create a consensus combination of codes and validated them with two cohorts: a UK cohort of 300 patients and a US cohort of 113 patients with PPV of 83% and 89% respectively¹⁸. The consensus combination of codes included two cirrhosis diagnosis codes (alcoholic cirrhosis and other unspecified cirrhosis of the liver) combined with complication codes for esophageal varices, portal hypertension, hepatic failure, and hepatorenal syndrome. This methodology is limited to finding those with cirrhosis and a complication, completely missing those with compensated cirrhosis. In our study, we extend prior findings to show adequate PPVs in those without complications codes, such as the AB and ABD group, PPVs remained adequate at 82% and 86%, respectively.

Similarly, while algorithms with cirrhosis and cirrhosis complications codes had excellent performance in our hepatology clinic cirrhosis cohort, the AC algorithm performed poorly in our ED cohort. In our ED cohort, the ABC, ABD, and ABCD algorithms performed well overall with PPVs ranging from 72 to 93% indicating that the addition of high-risk lab data (FIB-4 > 2.67 or APRI > 1.0) to ICD codes significantly improves the PPV in a real-world cohort that is not enriched with liver disease. Lab data has been shown in prior studies to aid in detecting decompensations^31,32. Taken together, using FIB-4 or APRI when available offers a tangible improvement in cirrhosis detection and are more widely validated then the Bonacini discriminant score which is only used in hepatitis C and has largely been replaced by the aforementioned scores³³. Furthermore, our results emphasize that the pre-test probability of cirrhosis diagnosis is important to consider. Here, our observed PPVs in the ED validation cohort are an important guide to selecting an algorithm for the general population (i.e., ABD in those without complications and ACD or ABCD for those with complications).

Others have tried more intricate algorithms to identify patients with cirrhosis. For instance, augmenting ICD codes and lab work with natural language processing of imaging result reports achieved an overall PPV of 92%³². More recently training artificial algorithms including Naïve Bayes, Random Forest, and convolutional neural network with word2vec embeddings to detect patients with cirrhosis based on discharge summaries also led to high accuracy, notably 97.3% for the convolutional neural network model³⁴. These studies are limited by the use of single center electronic health record data without external validation. Artificial intelligence algorithms are prone to overfitting in internally validated analysis where the almost perfect results seen in internal validation do not translate to external validation¹⁸. Furthermore, these algorithms are not readily available and their future use will be limited by being more resource intensive when it comes to implementation¹⁹. Given these limitations, our study provides algorithms that rely on discrete data fields and are ready-to-use for cirrhosis identification.

Our study also assessed PPVs for algorithms for identifying complications of cirrhosis including hepatic encephalopathy and ascites which carry high morbidity and have prognostic implications. For hepatic encephalopathy the performance of the ICD codes alone was very poor and algorithms relying on HE-specific medications such as lactulose and/or rifaximin proved to be a better strategy. This result is similar prior studies in a veteran population and multiple hospital systems in Michigan^19,35. However, compared to both studies, the overall performance of HE medications at detecting HE was inferior in our state-wide cohort; 61% for lactulose and 76% rifaximin compared to 71–89% for Lactulose and 88–100% for rifaximin. It is possible that the variable performance is due to a larger number of records used for validation in our study as well as the difference in our study population (i.e., state-wide population). For ascites, our results confirmed with the poor performance for cirrhosis-related ascites ICD codes seen in prior studies^17,20,29.

Though there have been many studies addressing the issue of identifying patients with cirrhosis in administrative data, our study is unique in its rigorous and comprehensive methodology as we combined multiple different strategies that were used in prior studies. Furthermore, our data is statewide and spans multiple hospital systems and electronic health records. Another important contribution of our study is our final guide for chart reviewers which led to a high interobserver agreement rate. A limitation of our study is that it is not nationwide, which might have an impact on generalizability. We did not include procedural data to identify those with complications such as esophageal varices as procedural reports were not always available. While these types of data are worthy of future exploration, our definitions mirror the data availability of other large administrative datasets that span multiple health systems and are likely more generalizable. We also excluded patients who only had an ascites code without any other liver specific codes due to prior literature showing poor performance^17,19,20.

This may have negatively affected the sensitivity of our final cohort, however, we also noted the poor performance of the ascites codes. In addition, new liver disease specific ICD10 codes are being generated, such as K76.82³⁵, and will need to be added to future definitions. Finally, we used a retrospective design which has its inherit limitations with regard to cirrhosis identification as our chart reviewers had knowledge of data after the index date of cirrhosis identification.

To conclude, in our large statewide dataset, we show that combining ICD-10 codes for cirrhosis, its etiologies and complications with serum-based disease and fibrosis makers can adequately identify individuals with both compensated and decompensated cirrhosis in large administrative datasets. For those with cirrhosis complications, we validate the need for combinations of ICD-10 codes for cirrhosis diagnosis and complications (algorithm AC) when prevalence of cirrhosis is high. For those without cirrhosis complications, the ABD algorithm achieves adequate PPVs in both the specialty and general populations. In the general population, addition of other features such as high-risk labs or etiology codes/labs for liver disease such as algorithms ABC or ABCD) are ideally suited to identify those with cirrhosis. Finally, we also show that identifying complications such as HE and ascites remains poor even when using prescribed medication data. Future studies should focus on developing and implementing a readily available artificial intelligence tool based on natural language processing as this is likely to overcome the limitations of ICD-code and discrete data in identifying cirrhosis related complications. Looking forward, our study also supported the creation of the Indiana Digital Cirrhosis Cohort which has a state-wide membership and is fertile ground for cirrhosis research^36–38. This dataset is poised to provide important data that will inform contemporary cirrhosis-related natural history and clinical care.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Study Concept and Design: Eric Orman, Naga Chalasani, Archita DesaiData Analysis: Hani Shamseddeen, Lauren Lembcke, Sui Lui Hui, Eric Orman, Archita DesaiManuscript Preparation: Hani Shamseddeen, Archita DesaiCritical Manuscript Review: All authors.

Funding

APD is funded by the National Institute of Diabetes and Digestive and Kidney Diseases of the National Institutes of Health under award number R03DK139283 and, previously, by award number K23DK123408.

Data availability

The analytic methods used in this study are detailed in the methods. The datasets generated during and/or analyzed during the current study are available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.