Population-level screening for liver disease: Opportunities and challenges

Cirrhosis accounts for nearly 1.3 million deaths worldwide in 20231 and ranks as the ninth leading cause of death in the United States.2 The epidemiology of cirrhosis is undergoing a marked shift, with steatotic liver disease (SLD) surpassing chronic hepatitis B and C as the dominant underlying etiology.3 Most concerning, liver disease is often clinically silent, with many patients remaining undiagnosed until they present with advanced disease or experience an adverse liver-related outcome. Despite this substantial burden, no dominant risk-based strategy is currently implemented to identify individuals at risk for advanced liver disease in the general population. In response to this major gap in care, the LiverRisk score was developed using data from 7 prospective studies conducted in Europe and Hong Kong, including participants from the community, primary care, hepatology clinics, and alcohol rehabilitation programs.4 The model includes age, sex, fasting glucose, cholesterol, aspartate aminotransferase (AST), alanine aminotransferase (ALT), platelet count, and gamma-glutamyl transferase (GGT), and is set to predict liver stiffness as measured by transient elastography. The external validation of the LiverRisk score in 2 European cohorts showed acceptable-to-excellent discrimination with good calibration. The prognostic performance of the score was subsequently evaluated in the UK Biobank, where it demonstrated excellent discrimination for predicting the 10-year risk of liver-related mortality. In this issue of Hepatology, Hernaez et al.5 aimed to externally validate the prognostic performance of the LiverRisk score in the United States. To achieve this, they analyzed the performance of the score among 170,998 patients without known chronic liver disease receiving care within the Veterans Health Administration (VHA), the largest integrated US healthcare system, which serves a sociodemographically heterogeneous population. Model discrimination was assessed using the C-statistic for liver-related events, hepatocellular carcinoma (HCC), and liver-related mortality in the overall cohort and across racial and ethnic subgroups. Hazard ratios were estimated across LiverRisk score categories for each outcome. The authors found that the discriminative performance of the LiverRisk score varied according to the outcome being predicted and the subpopulation in which it is used. The score showed modest discrimination for liver-related events in the overall cohort (C-statistic 0.61) and among non-Hispanic White (C-statistic 0.62) and Hispanic (C-statistic 0.60) Veterans. Its performance was better for liver-related mortality, with acceptable discrimination in the overall cohort (C-statistic 0.70) and among non-Hispanic White (C-statistic 0.70) and Hispanic (C-statistic 0.77) patients. Findings were similar for HCC, with notably excellent discrimination among Hispanic patients (C-statistic 0.82). However, the LiverRisk score had limited performance in non-Hispanic Black Veterans, with modest discrimination for liver-related mortality (C-statistic 0.63) and low discrimination for liver-related events (C-statistic 0.59), as well as HCC (C-statistic 0.53). Expectedly, higher LiverRisk scores were consistently associated with greater risk of each outcome across populations, with the strongest associations observed for HCC. Quantitative tools to identify individuals at increased risk of future liver events and related complications may guide population health and linkage to appropriate care. The LiverRisk score represents an advance over commonly used tools, demonstrating superior discriminative performance compared with the fibrosis-4 (FIB-4) score for all 3 outcomes of interest and with the AST-to-Platelet Ratio Index (APRI) for liver-related mortality. However, some limitations should be considered. Although the LiverRisk score relies primarily on routinely available laboratory variables, GGT, one of its key components, is not commonly obtained in clinical practice. This limitation is highlighted by the fact that only slightly than 1% of the Veteran cohort had complete laboratory parameters to calculate the score. GGT, a key enzyme involved in glutathione metabolism, is a sensitive but nonspecific marker of liver injury. While abundant in hepatocytes, it is also present in multiple extrahepatic tissues, including the kidneys and pancreas.6 GGT also serves as an indirect marker of sustained heavy alcohol use and higher body mass index,7 and epidemiological studies have suggested that GGT is independently and positively associated with all-cause, cardiovascular, and liver-related mortality.8 The incorporation of GGT in the LiverRisk score may partially explain its superior predictive performance compared with APRI and FIB-4 for future liver disease and related complications. However, further research is needed to define the optimal clinical context and thresholds for incorporating GGT into liver disease screening and risk stratification strategies in US cohorts. Additional variables in the LiverRisk score, such as fasting glucose and total cholesterol, also warrant consideration. Despite that these are commonly available in clinical settings, the reliability of obtaining a truly fasting glucose and its changes over time in response to diet and anti-diabetic medications is unknown and could impact score performance. Similar questions arise for total cholesterol, which is a marker of liver synthetic function but is majorly altered by lipid-lowering therapy. Equally important is the observed heterogeneity in performance across racial and ethnic groups. While the LiverRisk score showed acceptable-to-excellent discrimination for liver-related mortality and HCC on the whole, its performance was limited among non-Hispanic Black Veterans. This discrepancy may reflect differences in the demographic composition of the derivation and validation cohorts. Black patients represented 12% of the validation cohort, whereas race and ethnicity were not reported in the derivation cohort, presumably due to limited representation. This finding warrants careful consideration, as the use of risk prediction models with differential performance across populations could inadvertently exacerbate existing healthcare disparities. Future work needs to be done to develop robust prediction models across population strata. When evaluating a risk prediction model, both discrimination and calibration should be assessed, as they represent distinct and complementary aspects of predictive accuracy. Discrimination reflects a model’s ability to distinguish between individuals at higher versus lower risk, whereas calibration assesses the agreement between the predicted risk and the observed risk (ie, the “truth”). For example, the LiverRisk score may correctly identify a 2-fold difference in risk between 2 patients, accurately classifying them as higher versus lower risk. However, the score may predict absolute risks of 5% and 10% when the observed risks are actually 20% and 40%. In other words, the LiverRisk score may correctly rank patients by risk, indicating excellent discrimination, but systematically underestimates (or overestimates) the true risk, reflecting suboptimal calibration. The original LiverRisk formula had excellent calibration; however, calibration was not reported in this paper in the Veteran population. It will be important to assess calibration in the future. Additional considerations include generalizability and transparency. The VHA cohort used for external validation had substantially higher event rates than the derivation cohort, including a 15-fold higher rate of liver-related events and a 9-fold higher rate of HCC. This difference is also reflected in the greater proportion of patients classified as medium risk (2.2% vs. 0.8%) and high risk (0.5% vs. 0.1%) by the LiverRisk score in this Veteran cohort compared with the derivation cohort. A combination of factors is the most likely explanation, including the possible presence of unrecognized liver disease at baseline, the high burden of comorbidities in the Veteran population, and laboratory testing, especially GGT, being obtained for cause such as alcohol consumption or concern about liver disease. Another important consideration is that the mathematical formula of the LiverRisk score has not been published. The absence of the full model specification limits further independent validation and restricts opportunities for model updating and recalibration, which are particularly important for generalizability. We urge transparency and eventual publication of the formula to maximize its impact on liver health worldwide. In summary, this work by Hernaez et al.5 represents a meaningful step forward in the field’s efforts in population-based liver health approaches. Future studies should focus on evaluating model calibration in additional populations and across specific subgroups, exploring strategies for model recalibration, and assessing the impact of risk-based tools on clinical decision-making and patient outcomes.