Combined use of the ELF test and CLivD score improves prediction of liver‐related outcomes in the general population

Effective and feasible population screening strategies are needed for the early detection of individuals at high risk of future severe liver‐related outcomes. We evaluated the predictive performance of the combination of liver fibrosis assessment, phenotype profile, and genetic risk.


| INTRODUC TI ON
Chronic liver disease is one of the leading causes of morbidity and mortality worldwide. 1In Europe, liver disease is estimated to be the second leading cause of years of working life lost. 2,3Chronic liver disease often develops silently and diagnosis is frequently delayed until the patient presents with complications of end-stage liver disease. 4,5This has attracted increasing interest in developing noninvasive strategies that are both effective and practical for the early detection of individuals at high risk of future advanced liver-related clinical outcomes.Of the common non-invasive tests to detect advanced liver fibrosis, elastography is the most accurate 6 but impractical as a first-line tool for large-scale screening in the population because it requires costly local equipment and operator expertise.
Blood-based fibrosis markers have emerged as more feasible firstline tests for this purpose. 6The enhanced liver fibrosis (ELF) test, a direct blood-based marker of fibrosis, has shown superiority over indirect markers, such as FIB-4 or APRI, in detecting advanced liver fibrosis in several liver diseases, including alcohol-related liver disease and non-alcoholic fatty liver disease (NAFLD). 7,8[11] The predictive performance of the ELF test and other fibrosis markers decreases for longer term (>5 year) predictions, 12,13 which is logical because fibrosis stage only provides static information on the current disease status, not the drivers of disease progression.
Combining fibrosis markers with risk factors of disease progression could help improve predictions in the general population.The chronic liver disease (CLivD) score could be used for this purpose. 14sed on age, sex, alcohol use, waist-hip ratio, diabetes, smoking, and, if available, gamma-glutamyltransferase (GGT), the CLivD provides an estimate of the 15-year risk of incident severe liver disease.
The CLivD score together with fibrosis markers may enable better long-term predictions, but this has not been studied.Moreover, genetic factors in the form of polygenic risk scores (PRSs) may provide additional predictive information, 15 but the added gain from PRSs above and beyond fibrosis markers and clinical risk factors remains unclear and scarcely studied.
In this study, we evaluated the performance of a combination of the CLivD score, ELF test, and a PRS for predicting liver-related outcomes in the general population.The findings can aid in implementing future screening strategies to identify individuals at high risk for liver-related outcomes.Due to the regional two-stage stratified cluster sampling procedure, the sample is considered representative of the entire Finnish population.The Health 2000 Survey protocol has been described in detail previously. 16Briefly, baseline data were gathered via structured Average weekly alcohol use in grams of ethanol was assessed by a quantity-frequency questionnaire as described elsewhere. 16oking was categorized as current, former, or never smokers.Waist and hip circumference, weight, and height were measured.Diabetes was defined as fasting glucose ≥7.0 mmol/L or HbA1C ≥48 mmol/ mol or a previous diagnosis of diabetes or prescription of medication for diabetes.Metabolic syndrome was defined as the presence of at least three of the following: waist circumference ≥94 cm for men or ≥80 cm for women; serum triglycerides ≥1.7 mmol/L; serum HDL-C < 1.0 mmol/L for males or <1.3 mmol/L for females; blood pressure ≥130/85 mmHg or anti-hypertensive medication; and fasting serum glucose ≥5.6 mmol/L or diabetes medication. 17

| CLivD score
The non-laboratory CLivD non-lab score was calculated based on age, sex, smoking status (current vs. previous/never), alcohol use (number of weekly drinks), waist-to-hip ratio, and diabetes (yes vs. no) as

Key points
In this study, the combination of the ELF test, which assesses liver fibrosis, and the CLivD score, a risk-based scoring system for severe chronic liver disease, was investigated to improve predictions of hospitalization, cancer or death related to liver disease in the general population.The findings revealed that using both the ELF test and CLivD score provided better predictions compared to using the ELF test alone.This means that this combined approach can help identify individuals at a higher risk of liver problems, enabling early preventive measures.
| 2109 described in the Supplementary methods.The laboratory version, CLivD lab , also included GGT (U/L).Both CLivD scores were categorized into subgroups (minimal, low, intermediate, and high risk) based on previously reported thresholds (Supplementary methods).

| ELF test
Assays of tissue inhibitor of matrix metalloproteinase-1 (TIMP-1), hyaluronic acid (HA), and amino terminal peptide of pro-collagen III (PIIINP) were performed using frozen serum samples in accordance with the manufacturer's instructions (Siemens Healthineers) using the ADVIA Centaur XPT analyser (Siemens Healthineers) at the Biomarkers Team, Finnish Institute for Health and Welfare (Helsinki, Finland).We used previously recommended ELF cut-offs: ≥9.8 to indicate high risk of advanced fibrosis or cirrhosis, and ≥11.3 to indicate high risk of cirrhosis.

| Genetic data
To assess genetic risk, we used PRS-5 based on variations in PNPLA3, TM6SF2, GCKR, MBOAT7, and HSD17B13 -the strongest and best validated genetic risk variants in fatty liver disease. 18,19PRS-5 correlates with liver fat content and predicts hepatocellular carcinoma and liver-related outcomes in external validation datasets. 15,18,19netic data to calculate the PRS-5 were available in a subset of 5209 individuals.

| Outcome data
Follow-up data were obtained from national electronic healthcare registers through linkage using the unique personal identity code assigned to all Finnish residents.Data for hospitalizations were obtained from the Care Register for Health Care (HILMO), which covers all hospitalizations in Finland since 1969.One or several ICD diagnoses are assigned to each hospitalization at discharge; these diagnostic codes are systematically recorded in the HILMO register.Data for malignancies were obtained from the Finnish Cancer Registry, which has nationwide cancer records since 1953.
Vital status and cause-of-death data were obtained from Statistics Finland.According to Finnish law, each person who dies is assigned a cause of death on the official death certificate issued by the treating physician based on medical or autopsy evidence, or based on forensic evidence when necessary; the death codes are then verified by medical experts at the register and recorded according to systematic coding principles.Data reporting to all these registries is mandatory by law, and the general quality is consistent and virtually 100% complete. 20,21llow-up was conducted until December 2015.The primary outcome was fatal and non-fatal advanced liver disease requiring hospital admission or causing liver cancer, or liver-related death defined in line with a recent consensus paper 22 ; the specific ICD-10 codes are provided in Table S1.

| Statistical methods
For comparing groups, we used chi-squared or Mann-Whitney tests as appropriate.The primary analyses were performed using the non-laboratory CLivD non-lab score because this version of the CLivD model is more practical in real-world settings. 14We then repeated the analyses using the CLivD model containing GGT (CLivD lab ) to confirm the consistency of the findings (see Supplementary appendix).We calculated the probability of having elevated ELF values over the spectrum of CLivD scores using logistic regression.We performed univariate and multivariable Cox regression analyses to analyse associations of the ELF test, CLivD score, and PRS-5 with liver-related outcomes.No transformation of these predictors was performed.We used Harrell's Adequacy Index to assess the fraction of new predictive information provided by a combination of the ELF test and CLivD score and/or PRS-5 compared to the ELF test alone. 23Harrell's Adequacy Index is calculated as 1-(a/b), where a is the likelihood ratio statistic for a Cox model with ELF as the only independent variable, and b is the likelihood ratio statistic for a Cox model with ELF and CLivD score and/or PRS-5 as independent variables.Time-dependent area under the curve (AUC) values were calculated based on competing-risk Fine-Grey regression analysis.
The absolute risk of liver-related outcomes at 10 years was predicted based on Fine-Grey regression.Subgroup analyses were by alcohol risk use (≥30 grams of ethanol per day for men and ≥20 g/day for women), obesity (BMI ≥30 kg/m 2 ), and diabetes.Data were analysed in R software version 3.6.1 using the packages survival, rms, car, tableone, pec, cmprsk, and riskRegression.

| RE SULTS
Of the 5795 individuals in this study, 46% were men, 80% were current alcohol drinkers, and 27% current smokers.The mean age was 52 years and mean BMI 26.9 kg/m 2 (Table 1).The probability that a person had an ELF value ≥9.8 (indicative of advanced liver fibrosis) was <4% at the lower end of the CLivD score, and this probability rose to >68% at the higher end of the CLivD score (Figures 1 and   S1).Similarly, the probability of having an ELF ≥11.3 (indicative of cirrhosis) rose from <0.2% at the lower end of the CLivD score to >8% at the higher end.
In the univariate Cox regression, ELF, CLivD, and PRS-5 were all significantly associated with liver-related outcomes (Tables 2 and   S2).These associations remained significant in the multivariable models (Table 2).The coefficients of the various Cox models are provided in Table S3.Time-dependent AUC values tended to decrease for longer term predictions with ELF and CLivD, and this decrease was steeper for ELF than for CLivD.For PRS-5, AUC values remained poor at all follow-up times (0.43-0.55).
The highest AUC values at 1, 5, and 10 years of follow-up were observed in the model including ELF and CLivD (Table 2).Figure 2 shows how the predictive performance in terms of AUC values significantly improved for the model containing both ELF and CLivD compared to ELF alone; the delta-AUC at 10 years was 0.097 (p < .0001).Adding PRS-5 to this model did not improve the predictive performance.These findings were consistent for the CLivD lab score, which also incorporates GGT in risk predictions (Table S2).
The findings were also consistent in subgroups of individuals with obesity, diabetes, or alcohol risk use; 5-year AUCs for the model with ELF and CLivD were 0.94 in individuals with diabetes, 0.93 in obese individuals, and 0.84 in those with alcohol risk use (Tables 3   and S4).

Based on Harrell
(Table S5).However, additional inclusion of the PRS-5 added only 5.7% of new predictive information to the combination of the ELF test and CLivD score.Findings were consistent for the CLivD lab score (Table S6).
Figure 3 shows the 10-year absolute risk of liver-related outcomes over the spectrum of different ELF and CLivD non-lab scores.For example, at an ELF value of 11.3, the 10-year risk ranged from 0.3% to 33% depending on the CLivD score.

| DISCUSS ION
We found that combining the CLivD score, as a measure of risk of progression, with the ELF test, as an estimate of fibrosis stage,  We found that the predictive value, in terms of discrimination of the ELF test, tended to decrease over longer follow-up times.
Addition of the CLivD score improved these long-term predictions.
Fibrosis markers, including the ELF test, tend to only reflect the current liver status, whereas the CLivD score reflects the risk of liver disease progression based on the individual risk profile.Therefore, it makes sense that combining a marker of current disease status (ELF test) with a marker of progression risk (CLivD score) improves longterm outcome predictions in particular.
Previous studies have shown that non-invasive liver fibrosis tests poorly predict liver-related outcomes in an unselected general population 25,26 ; therefore, guidelines advocate that fibrosis markers only be used in at-risk populations. 6We recently showed that the CLivD score performed better at selecting individuals for liver fibrosis testing than traditional risk factors, such as harmful alcohol use, diabetes, or obesity considered in isolation. 27The CLivD score combines the effects of several risk factors and also considers low levels of alcohol intake and adiposity, providing a more sensitive estimate of overall risk compared to an impressive increase in a single risk factor. 28e ELF test has been widely validated 29 and shown to outperform FIB-4 and APRI for detection of advanced fibrosis or cirrhosis in both NAFLD and alcohol-related liver disease patients. 7,8 addition, in alcohol-related liver disease and mixed-aetiology liver diseases, the ELF test predicts liver-related outcomes as least as well as liver histology. 9,30,311]32 The ELF test is reportedly a cost-effective first-line test for screening advanced liver fibrosis in NAFLD 33 and alcohol-related liver disease. 34fortunately, due to a lack of platelet data, we were unable to    Note: The outcome is incident liver-related events over a 10-year follow-up.The estimates were computed using inverse probability of censoring weighted with competing risks.

2. 1 |
Health 2000 Data were sourced from the Health 2000 Survey, a multidisciplinary epidemiological study performed in Finland in 2000-2001, which was coordinated by the Finnish Institute for Health and Welfare (previously known as National Public Health Institute of Finland).
interviews and questionnaires, clinical measurements, blood tests, and clinical examinations by a physician.The Epidemiology Ethics Committee of the Helsinki and Uusimaa Hospital Region approved the Health 2000 Survey protocol, and all participants provided signed informed consent.The Health 2000 Survey sample collection was transferred to THL Biobank in 2015 after approval from the Coordinating Ethics Committee of the Helsinki and Uusimaa Hospital District and Ministry of Social Affairs and Health.Of the original sample of 8028 adults aged ≥30 years, 5837 had participated in a health examination, had complete data to calculate the CLivD score, and had blood samples available for ELF analyses.After the additional exclusion of 42 subjects with a diagnosis of liver disease at or before baseline (International Classification of Disease [ICD-10]: K70-K77 or C22.0), the final study cohort comprised 5795 individuals.
provides better prediction of liver-related outcomes than the ELF test alone.The CLivD score added 50.7% of new predictive information to the predictions made by the ELF test.The model combining the ELF test with the CLivD score provided AUCs of 0.85 at 5 years and 0.80 at 10 years, which are considered excellent given that these predictions concern the rare outcome of clinical liver-related events in the general population.The strengths of our study include the representative general population sample and linkage to high-quality national healthcare registries with high coverage.20,21,24Representability of the Heath 2000 sample was ensured and participation rate (80%) exceptionally high for a population-based survey.16Moreover, socioeconomic inequality in Finland is relatively low.Regarding clinical outcomes, we included all types of liver-related outcomes considered clinically relevant and representative of severe liver disease (hospitalization, death, liver cancer).The sample includes both extensive phenotype data and genetic data and represents the largest sample of ELF measurements to date.
compare the ELF test to other blood-based fibrosis markers, such as FIB-4 and APRI.We found that adding genetic data to the ELF test and CLivD score failed to noticeably improve risk predictions regarding liverrelated outcomes.In a recent UK Biobank study, Innes et al. 35 came to the same conclusion, although they evaluated other fibrosis F I G U R E 2 Time-dependent area under the curve (AUC) values at 1, 5, and 10 years of follow-up for liver-related outcomes by a competing-risk model including only the Enhanced Liver Fibrosis (ELF) test, or ELF and CLivD non-lab score.*p = .004,**p = .00001.

F I G U R E 3
Heatmap showing the 10-year absolute risks (%, z-axis) for liver-related outcomes according to the Enhanced Liver Fibrosis (ELF) value and CLivD non-lab score.Cut-offs for previouslydefined CLivD non-lab risk groups (minimal, low, intermediate, and high) are shown on the x-axis.Analysis is by Fine-Grey competing-risk regression.TA B L E 4 Time-dependent predictive performance of ELF alone and a sequential strategy with CLivD non-lab as the initial test and the ELF test targeted to those with a CLivD score above a specific threshold.

ab
Test negative = CLivD non-lab in the minimal risk category or ELF <9.8.Test positive = CLivD non-lab in the low/intermediate/high-risk category and ELF ≥9.8.Test negative = CLivD non-lab in the minimal or low-risk category or ELF <9.8.Test positive = CLivD non-lab in the intermediate/high-risk category and ELF ≥9.8.Limitations of our study include the low number of liverrelated outcomes.Nonetheless, all of the predictors considered (ELF, CLivD, and PRS-5) are continuous variables developed in previous studies.Therefore, in the current validation setting, there was a minimum of 19 outcome events per predictor in the Cox models, which should ensure sufficient sample size.As the study was based on the Finnish population, our findings should be validated in other populations.In conclusion, combined use of the ELF test and CLivD score in the general population provides excellent prediction of liver-related outcomes.Further studies are warranted to evaluate a screening strategy based on the CLivD score followed by the ELF test.
's Adequacy Index, inclusion of the CLivD score added 50.7% of new predictive information to the ELF test TA B L E 1 Baseline characteristics.
TA B L E 2