RESEARCHERS have developed a machine learning model that has matched the prognostic accuracy of the invasive gold-standard test for predicting hepatic decompensation in compensated advanced chronic liver disease, using only routine laboratory and demographic data.
Hepatic venous pressure gradient (HVPG) has remained the gold standard for prognostic assessment in compensated advanced chronic liver disease (cACLD), yet its invasiveness has limited broad clinical use. Established non-invasive alternatives such as liver stiffness measurement (LSM) have also required specialised infrastructure not universally available, leaving a gap for scalable, accessible risk stratification tools using machine learning and routine clinical data.
Researchers developed the Cirrhosis Risk Identifier (CIRI), a machine learning model using 11 demographic and routine laboratory parameters, and benchmarked it against HVPG, LSM, FIB-4 and MELD. Patients with cACLD were identified via the US Optum Clinformatics Data Mart for model development and internal validation, with external validation performed in a prospective European tertiary care cohort.
CIRI was trained on 112,618 patients and internally validated on 18,852 patients in the US cohort, while 210 HVPG-characterised patients formed the European external validation cohort. Steatotic liver disease was the leading aetiology in both cohorts (54% and 44% respectively), with median follow-up of 18.5 and 27.5 months.
CIRI achieved 1- and 2-year time-dependent AUROCs of 0.816 and 0.815 in the US cohort, outperforming MELD and FIB-4 (both P<.001). In the European cohort, AUROCs of 0.836 and 0.769 were comparable to HVPG (both P>.900) and superior to LSM (both P<.05). CIRI independently predicted decompensation in both cohorts (US: adjusted subdistribution hazard ratio 1.67, P<.001; Europe: aSHR 1.66, P=.017), and a cut-off of -8.25 or higher identified patients at comparable decompensation risk to HVPG readings of 10mmHg or higher, indicating clinically significant portal hypertension.
This machine learning approach has demonstrated prognostic performance comparable to HVPG and has exceeded that of widely used non-invasive tests, suggesting it could enable scalable, repeatable risk stratification in cACLD. Pending further prospective validation, the CIRI model may help clinicians guide surveillance intensity and target preventive strategies towards patients at highest short-term risk of hepatic decompensation, without requiring invasive pressure measurement.
Kramer G et al. A machine learning approach to non-invasive prediction of hepatic decompensation in compensated advanced chronic liver disease: the CIRI model. Journal of Hepatology. 2026;DOI:10.1016/j.jhep.2026.06.042.
Featured image: Pakorn on Adobe Stock
Each article is made available under the terms of the Creative Commons Attribution-Non Commercial 4.0 License.
