Use of new CORE risk score to predict 10 year risk of liver cirrhosis in general population: population based cohort study.

Introduction
The global burden of chronic liver disease is increasing.1 Liver cirrhosis is the 11th most common cause of death worldwide, and liver cancer—predominantly hepatocellular carcinoma—is the sixth most common type of cancer and fourth most common cause of death due to cancer.1 Such end stage liver diseases have poor prognosis, partly because liver disease is often asymptomatic until severe complications occur, such as the classical decompensation signs of oesophageal varices, ascites, or hepatic encephalopathy.2 Metabolic dysfunction associated steatotic liver disease (MASLD) is the most common chronic liver disease, with an estimated global prevalence of 38%.3 The prevalence is particularly high among patients with type 2 diabetes or obesity.4 MASLD and other liver diseases have the potential to progress through different stages of metabolic dysfunction associated steatohepatitis and liver fibrosis to cirrhosis and hepatocellular carcinoma. This makes patients with confirmed or suspected MASLD particularly suitable to be screened for possible advanced fibrosis. MASLD is the most common cause of hepatocellular carcinoma in Sweden and elsewhere.5 By contrast, alcohol related liver disease is less common in the general population but is the leading cause of cirrhosis in many European countries.6 Together, alcohol related liver disease and MASLD are common liver diseases in the general population but are largely under-diagnosed. The primary risk factor for progression to cirrhosis or hepatocellular carcinoma remains the stage of liver fibrosis.7

In recent years, international societies have published recommendations for screening for advanced fibrosis in primary care among patients with or at risk for MASLD or alcohol related liver disease.8
9 Detecting pre-cirrhotic disease early can potentially improve prognosis through interventions including lifestyle changes (alcohol abstinence in alcohol related liver disease and weight loss in MASLD) or by novel treatments such as resmetirom for patients with fibrotic metabolic dysfunction associated steatohepatitis.10 Patients with screening detected compensated cirrhosis may benefit from surveillance for oesophageal varices and hepatocellular carcinoma, for example.8 As a first line test, guidelines recommend the non-invasive, blood based FIB-4 score. FIB-4 was originally developed to detect advanced fibrosis in patients with HIV and hepatitis C viral co-infection,11 but it has since been applied to other groups of patients with liver disease. Some concerns have been raised about its performance in the general population or in the primary care setting, where the prevalence of advanced fibrosis is relatively low.12
13 Although FIB-4 and other novel diagnostic alternatives are relatively successful in detecting fibrosis at the time of testing, they are less effective for predicting future cirrhosis events.13

In this study, we developed and validated a new score—the Cirrhosis Outcome Risk Estimator (CORE)—intended for use in the primary care setting. CORE predicts the 10 year risk of cirrhosis, its related complications, or hepatocellular carcinoma. We directly compared CORE with FIB-4 and validated it in two external cohorts.

Methods

Data
We used the AMORIS (Apolipoprotein-related Mortality Risk) cohort to develop our risk prediction model. The cohort has been described thoroughly elsewhere.14 Briefly, the AMORIS cohort is based on people undergoing health examination with laboratory data collected in the Stockholm region during 1985-96, mostly from primary care or occupational health screenings. AMORIS encompasses approximately 35% of the Stockholm population at the time, with a sociodemographic composition representative of the population. The individuals were matched to several Swedish regional and national registers to ascertain outcomes, of which the most relevant for this study are the inpatient and outpatient registers, the cancer register, and the cause of death register.
What makes AMORIS unique is that all laboratory analysis was done by a single laboratory with standardised and documented methods, where they offered measurements of many biomarkers at once in a standard package. AMORIS therefore contains liver related biomarkers for most people, even without suspicion of liver disease.
The outcome the model predicts is the 10 year risk of major adverse liver outcomes (MALO), which includes diagnosis of cirrhosis (including alcohol related cirrhosis), oesophageal varices (bleeding or non-bleeding), hepatic ascites, hepatorenal syndrome, portal hypertension, hepatic encephalopathy, chronic liver failure, hepatocellular carcinoma, and liver transplantation. The ICD (international classification of diseases) codes used to define this outcome are shown in supplementary table A.
We chose the candidate predictors on the basis of subject matter knowledge of risk factors for liver related disease, together with availability in both the training data and the primary care setting. Potential predictors included age, sex, body mass index, and the laboratory variables aspartate aminotransferase, alanine aminotransferase, γ-glutamyl transferase, platelet count, albumin, bilirubin, cholesterol, glucose, and triglycerides. These parameters are readily available in primary care across most countries and are relatively cheap.
Participants started follow-up at their first recorded laboratory value among the candidate predictors (baseline). We excluded participants with none of the laboratory values considered for the prediction model and those with a previous diagnosis of MALO. We further excluded people with known liver disease, such as viral or autoimmune hepatitis, or other alcohol and drug related disorders at or before baseline (supplementary table B). This is because such patients would either have established contact with specialised care or be considered at substantial risk for MALO outside of the target population (primary care). We did not censor individuals who developed specific liver diseases after baseline, as such information would not be available in a prediction setting. We also excluded individuals outside the age span 18-80 years at baseline and those with extreme values for some laboratory variables (supplementary table C). See supplementary figure A for a flowchart of the derivation of the final dataset.
We considered deaths from non-hepatic causes to be competing events. We censored individuals at the end of the study period (31 December 2020) or at emigration.
We imputed missing predictor values with multiple imputation using chained equations.15 We did 100 imputations as some predictors had a very high degree of missingness. The percentage of missing values for each predictor is shown in table 1, and a dendrogram of the relations between the missing values can be found in supplementary figure B. Additionally, a total of 73 laboratory values were manually truncated for the sake of the statistical modelling (supplementary table C).

Statistical analysis
We used cause specific flexible parametric survival models for both the outcome of MALO and for the competing outcome of non-MALO death.16 We modelled the continuous predictors by using restricted cubic splines. We included interactions between aspartate aminotransferase, alanine aminotransferase, and γ-glutamyl transferase owing to their correlation and how they have appeared in other prediction models; between age and γ-glutamyl transferase, albumin, and cholesterol; and between sex and γ-glutamyl transferase, aspartate aminotransferase, alanine aminotransferase, and platelets. Inclusion of these interactions was motivated by the age and sex dependent clinical guidelines at Karolinska University Hospital for normal and abnormal concentrations of these predictors. We included linear interaction terms with log(time) for all the candidate predictors to account for some potential non-proportionality in the hazards.
The model development was based on the methods described in Harrell 2015.17 Firstly, we fitted a full risk model with all candidate predictors. To improve the usability of the model, we applied a suitable transformation to the predicted risks of this model and used them as the outcome in a normal linear regression model with the same regression formula as in the full model. Then we did backwards elimination to successively remove each predictor by smallest reduction in R2. From this, we chose an appropriately simplified model on the basis of a remaining R2 of ≥95%. For more details on the model development, please see the supplementary materials. Whenever references to “cumulative incidence” or “incidence” are made, we estimated by using the Aalen-Johansen estimator.18

Model performance
We assessed the performance of the developed prediction model through three different aspects. First discrimination, which evaluates how well a risk model separates people who will experience the outcome from those who will not. We assessed discrimination by using receiver-operator characteristic curves and the time dependent area under the curve.19 Second calibration, which assesses how well predicted risks correspond to the observed risks in the data. We used calibration curves in which the individuals in each dataset were grouped by their predicted risks and the average predicted risk in each group was compared with its estimated cumulative incidence.20 Third clinical utility, which relates the risk prediction back to the clinical decision problem. We assessed this by using net benefit and decision curves to compare the relative benefits and harms of true positives and false positives.21 For more details on how these performance measures were estimated, please see the supplementary materials.

Validation
The prediction model was internally validated using bootstrap validation with a total of 500 bootstrap samples.22 The bootstrapping was combined with the multiple imputation by first imputing 100 datasets, and then bootstrap sampling five datasets from each imputation.23 We repeated all modelling steps until we found an approximated model with three laboratory values. We always considered age and sex as preselected in the variable selection. We used these bootstrap samples both to optimism correct the model performance and to construct 95% confidence intervals.
For external validation, we had access to data from the UK Biobank and FINRISK 2002-12 and Health2000 (“FINRISK”) cohorts.24
25 Both validation datasets used the same outcome of 10 year risk of MALO according to supplementary table A and the same additional exclusion criteria at baseline according to supplementary table B. No missing values were imputed in the external validation cohorts (that is, complete case analysis).

Comparison with FIB-4
We compared the developed prediction model (CORE) with the FIB-4 score. We made special references to the established FIB-4 cut-off values 1.30 and 2.67, which are used to rule out and rule in advanced fibrosis. As FIB-4 is not expressed on the probability scale (0-100%), it could be directly compared with CORE only in terms of discrimination (receiver-operator characteristic curves and area under the curve), not calibration and net benefit. Therefore, we “re-calibrated” the FIB-4 score to put it on the probability scale, as shown in supplementary figure C. We reported the statistical analysis according to the TRIPOD guidelines.26

Patient and public involvement
This study was based on de-identified data from national health registers and the AMORIS laboratory database. We cannot determine exactly which members of the AMORIS cohort were included in this study, and contacting the entire cohort personally was not possible for logistic and financial reasons. The researchers did not have the necessary infrastructure or funding available to further pursue additional patient or public partnership for this specific project.

Results

Data

Table 1 summarises the characteristics of the training data. Of the 480 651 individuals included, 7168 (1.5%) had MALO diagnosed during follow-up. The median follow-up was 27.7 (interquartile range 22.6-32.2) years, corresponding to 12.3 million person years of follow-up. Figure 1 describes the observed MALO events in the training data. A total of 7168 MALO events occurred during follow-up, of which 1331 (19%) occurred during the first 10 years, giving a cumulative incidence at 10 years of 0.27% (95% confidence interval 0.26% to 0.29%).

Table 2 shows a comparison between the three datasets used in this study. Notably, the 10 year incidence of MALO was higher in the validation datasets: 0.40% in the UK Biobank and 0.42% in FINRISK compared with 0.27% in the training data. The participants in the UK Biobank were notably older than those in the other two cohorts, and concentrations of γ-glutamyl transferase and aspartate aminotransferase were higher in the validation data than in the training data.

Model development
We developed one flexible parametric survival model for MALO and one for the competing event of non-MALO death. We used these two models to calculate (using numerical integration) the predicted 10 year risk of MALO, accounting for the competing event. These predictions represent the “full model.”
We then logit-transformed the predicted risks of the full model. We chose this transformation instead of log-log and probit because it produced the highest R2 between the predicted values of the full model and the final simplified model. When we did the backwards elimination on the logit-transformed predicted risks of the full model, γ-glutamyl transferase and aspartate aminotransferase were consistently the two strongest predictors. The third most important predictor was alanine aminotransferase, although third place ranking varied slightly during the bootstrap validation among alanine aminotransferase, albumin, cholesterol, and platelets, with alanine aminotransferase generally performing better. Sex and age—always known and “free” to collect—were automatically included in this simplified model. We observed strong interaction effects between these and a few of the laboratory based predictors, and they are known to be strongly associated with the competing risk. Age, sex, γ-glutamyl transferase, aspartate aminotransferase, and alanine aminotransferase could together approximate the predictions of the full model with an R2 of 96%. As this approximation was very good, we did not need to consider any additional biomarkers.
On the basis of these decisions, we finalised the simplified model with the predictors γ-glutamyl transferase, aspartate aminotransferase, alanine aminotransferase, age, and sex, which became the Cirrhosis Outcome Risk Estimator (CORE) model. A web calculator is available at www.core-model.com, and R code for calculating CORE is available at https://github.com/rickstra/CORE. The mathematical formula for CORE is provided in the supplementary materials.

Figure 2 summarises the model predicted risks in the training data. On average, the predicted 10 year risk of MALO was 0.26%; 4%, 0.5%, 0.2%, and 0.05% of participants had a predicted risk of at least 1%, 5%, 10%, and 20%, respectively.

Model performance

Figure 3 shows receiver-operator characteristic curves comparing CORE and FIB-4 in the training data. At every level of sensitivity, CORE has a better specificity than FIB-4; for every level of specificity, CORE has a higher sensitivity than FIB-4. For example, FIB-4 at the cut-off value of 1.30 had a sensitivity of 0.53 and a specificity of 0.86. The equivalent cut-off value for CORE, which gives a fixed sensitivity of 0.53, instead has a specificity of 0.96; and the cut-off value that gives a fixed specificity of 0.86 has a sensitivity of 0.73. Similarly, the FIB-4 cut-off value of 2.67 had a sensitivity of 0.18 and a specificity of 0.992, whereas for CORE this was 0.34 and 0.998, respectively, when the other was fixed to match FIB-4. The area under each receiver-operator characteristic curve was 0.88 (95% confidence interval 0.87 to 0.89) for CORE and 0.79 (0.78 to 0.80) for FIB-4.

Figure 4 presents time dependent areas under the curves in the training and validation cohorts. The area under the curve of CORE (optimism corrected using bootstrap validation) ranges from 0.94 (95% confidence interval 0.91 to 0.98) at one year to 0.88 (0.87 to 0.89) at 10 years. This is consistently much higher than the area under the curve for FIB-4, which ranges from 0.86 (0.84 to 0.89) at one year to 0.79 (0.78 to 0.80) at 10 years. In the validation cohorts, the 10 year area under the curve for CORE was lower than in the training data: 0.81 (0.77 to 0.87) in FINRISK and 0.79 (0.78 to 0.80) in the UK Biobank. However, in the UK Biobank (for which FIB-4 was available), the 10 year area under the curve for FIB-4 was 0.73 (0.72 to 0.74), similarly lower than CORE as in the training data.

Figure 5 shows calibration plots for CORE in the training and validation cohorts. Overall, the calibration was good in all three datasets, showing good agreement between predicted and observed risks. Some indication of overprediction exists in AMORIS and FINRISK for the especially high risks (>30%), but very few individuals make up those levels of risk (as indicated by the very wide error bars). The UK Biobank data instead show very good calibration at those high risks but indicate slight overprediction at the intermediate risks around 3-7%.
The top panel of figure 6 shows decision curves in the training data for four different theoretical treatment strategies for people at risk of MALO within 10 years: treating everyone as if they will develop MALO, treating none, or using either FIB-4 or CORE for risk stratification followed by treatment of those above the risk threshold indicated on the horizontal axis. “Treatment” in this regard can be seen as the need for further evaluation—for example, by referral to hepatology or for vibration controlled transient elastography. We can see that, for all potential risk thresholds below 20%—meaning the scenarios in which the benefit of correctly treating one person who will have MALO outweighs the harm of unnecessarily treating four or more people who will not—using CORE is the optimal strategy of the four for balancing the benefits and harms of true positives and false positives.
In the bottom panel of figure 6, we have added the decision curves for using CORE in the two validation datasets. We see that the net benefit is consistent overall across all three datasets. As the validation datasets had a higher incidence of MALO than the training data, the net benefit starts higher when the risk threshold is low but—as the discrimination performance was lower—the net benefit drops off faster and coincidentally converges with the net benefit of the training data. This indicates that CORE would bring a similar net benefit in all three populations.
In supplementary figure D, we break down the types of MALO events (between compensated cirrhosis, decompensated cirrhosis, and hepatocellular carcinoma) detected by CORE and FIB-4 in the training data. We here use hypothetical cut-off values of 0.4% and 5% for CORE, which would refer the same number of individuals as the established FIB-4 cut-off values of 1.30 and 2.67, respectively. We see that CORE finds more cases of all three types and especially more cases of compensated cirrhosis when using the lower cut-off values (0.4% and 1.30).
As the current screening recommendations are aimed at certain risk groups, we repeated the internal validation of CORE in a subpopulation of 52 202 people who would qualify for screening according to European Association for the Study of the Liver guidelines.8 How this subpopulation was derived is explained in the supplementary materials. This subpopulation had a 10 year risk of MALO of 0.59% (95% confidence interval 0.53% to 0.66%) and an estimated 10 year area under the curve of 0.894 (95% confidence interval 0.855 to 0.929.)

Because the estimated predictive performance in the training data was based on partially imputed data of aspartate aminotransferase, alanine aminotransferase, and γ-glutamyl transferase, we also calculated the area under the curve of a complete case dataset with respect to these three variables as a sensitivity analysis. The complete case sample size was 398 149 (83% of total) and the estimated 10 year area under the curve was 0.882 (0.869 to 0.891).

Discussion
In this study, we developed a new risk model—the CORE model—for predicting the 10 year risk of MALO in a primary care setting. It was trained on a very large dataset and validated in two external general population datasets. CORE handily outperformed FIB-4 in both the training data and the validation datasets, positioning it as a first line tool to predict incident MALO in the general population, complementing existing risk stratification strategies.

Strengths and limitations of study
Major strengths of this study include the use of a very large training cohort with near complete follow-up in Swedish health and population registers. We used a careful and statistically sound method while always keeping the end goal of clinical feasibility in mind when developing the CORE model. The performance assessment of CORE and FIB-4 was thorough, and the external validation was done completely independently by different analysts.
Another advantage of CORE compared with FIB-4 is that CORE is potentially cheaper. For example, at the Karolinska University Laboratory in 2024, the combined cost of ordering the constituent laboratory values for CORE was 54 SEK (£4.23; $5.69; €4.88) whereas for FIB-4 it was 117 SEK.27 In practice, all three laboratory values for CORE can be measured in a single lithium-heparin tube, whereas FIB-4 needs a separate EDTA tube for measuring the platelet count.
Several recent risk scores—such as the LiverRisk,28 SAFE,29 and MAF-530 scores—have been developed to detect different stages of fibrosis in a general population/primary care setting but may also be used for prediction purposes. We could not compare CORE directly with these scores owing to data availability (globulin for SAFE and waist circumference for MAF-5 not being available in AMORIS) and the formula for LiverRisk not being publicly available. However, CORE may have benefits over these models as the included parameters are few and inexpensive, are readily available, and do not require other time consuming tasks such as measuring waist circumference. Future studies are needed to directly compare the predictive capacity of CORE with these scores. A previous study using the AMORIS cohort evaluated the discrimination of several scores (FIB-4, NFS, APRI, BARD, Forns) when predicting incident MALO and found FIB-4 to be the best in that aspect.13

Interestingly, the individual parameters with the highest value to include in CORE were γ-glutamyl transferase, aspartate aminotransferase, and alanine aminotransferase. Adding further parameters did not improve the model enough to motivate construction of a more complex model (the “full model” had a 10 year area under the curve of 0.90, compared with 0.88 for CORE, which could approximate the full model to 96%). From a prediction perspective, why a parameter is linked to an outcome is of less interest, with focus instead on the predictive performance and how well it validates. In a general population, these are unspecific markers that signal ongoing liver injury and possibly alcohol consumption, which is known to be a strong risk factor for incident cirrhosis, also in patients with presumed MASLD.31

A limitation to CORE is that its training data are relatively old, with laboratory values available until 1996. However, CORE performed well and consistently better than FIB-4 in the two more modern validation cohorts, and prediction models are almost always built on historical data. The long follow-up period also allowed for enough events to occur for a robust analysis, something that is especially important in liver diseases that progress over decades.32 Another consequence of the cohort’s age is that defining appropriate metabolic risk groups was difficult owing to reporting at the time. For example, the ICD-9 system in Sweden did not differentiate between type 1 and type 2 diabetes. Other variables used to define metabolic disorder according to the European Association for the Study of the Liver and American Gastroenterological Association guidelines were also scarce. This also applied more generally to the candidate predictors of the model, with high percentages of missing values especially for platelets and bilirubin. However, the large sample size still meant that we had many measurements of these predictors, and we used many imputations to compensate. We also note that multiple imputation is the preferred method even for very high degrees of missing data.33 Although the predictors with particularly high missingness were not included in the final CORE model, we cannot rule out the possibility that these predictors could have had a stronger predictive performance had the measurements been more complete.

Future perspectives
All cohorts used to develop and validate CORE so far stem from western European countries. CORE would thus need further validation using data from other parts of the world where the prevalence of, for example, MASLD, alcoholic liver disease, type 2 diabetes, and viral hepatitis are different. Although we did not know precisely which underlying liver disease led to the MALO diagnosis, this is not really of concern when predicting incident MALO, and we excluded all prevalent cases of known liver disease at baseline. We therefore at least expect CORE to perform consistently in populations with similar prevalences of liver disease as seen here. CORE also needs to be validated in larger patient populations with a higher pre-test probability of having liver fibrosis, such as type 2 diabetes or confirmed MASLD. We have planned future studies in both such populations. External researchers are also invited to validate CORE in other populations and against other risk scores (using code available at https://github.com/rickstra/CORE).
A major obstacle for implementation of any risk score into primary care is accessibility. An online web calculator might not be enough to facilitate wide scale adoption in primary care. A direct integration into healthcare systems, so that CORE can be directly ordered from the laboratory and be automatically calculated, might be necessary. Some healthcare systems, including the largest electronic health record system in Stockholm, Sweden, have recently enabled this for FIB-4. Also, aspartate aminotransferase is not routinely available in some countries, but the strength of aspartate aminotransferase for predicting incident MALO as shown in this study might motivate a change in that practice.
Current risk stratification algorithms in clinical guidelines focus on diagnosing advanced fibrosis.8
9 A limitation with these algorithms is that no data are available on the absolute risk of developing MALO, which can be argued is what really matters to patients. CORE may be used to inform clinicians and patients about the long term risk of incident MALO, with results given as percentages. This may serve as a valuable tool to decide on the need for further evaluations such as vibration controlled transient elastography, referral to hepatology services, or initiation of treatment.
In terms of predicted risk, a CORE risk estimation of 0.4% largely corresponds to a FIB-4 value of 1.3, and a CORE risk estimation of 5% corresponds to a FIB-4 value of 2.67. However, health economic evaluations are needed to inform what risk threshold should motivate referral for hepatology evaluation or second line tests such as vibration controlled transient elastography. Although CORE had a higher net benefit (that is, identified more true positives per false positive or vice versa) than FIB-4 at all risk thresholds, which threshold is appropriate to choose for referral in clinical practice is not apparent. This would depend on differing willingness to pay between countries, as well as on the development of novel therapies. One also needs to separately consider organised screening (for example, by invitation every one to three years for patients with a diagnosis of type 2 diabetes) and opportunistic screening (for example, by testing CORE when patients with obesity visit primary care for other reasons). To our knowledge, this has not been earnestly discussed in the presentation of these types of risk scores or in the field as a whole.

What is already known on this topic
Many people in the general population, especially those with type 2 diabetes or obesity, are at elevated risk of developing major adverse liver outcomes (MALO)

Non-invasive tests that are cheap and easily available in primary care are needed to find patients before major complications such as cirrhosis and hepatocellular carcinoma occur

What this study adds
A large Swedish population based cohort (n=480 000) was used to develop a new risk score—the Cirrhosis Outcome Risk Estimator (CORE)—to predict the 10 year risk of MALO

The CORE model outperformed the currently recommended first line test FIB-4 when predicting future MALO, both in the training data and in external validation

The CORE model is a promising new first line test for finding clinically relevant patients at risk of MALO in primary care