A machine learning approach to non-invasive prediction of hepatic decompensation in compensated advanced chronic
Georg Kramer1, Anders Hildeman2, Caterina Tozzi3
1Division of Gastroenterology and Hepatology, Department of Medicine III, Medical University of Vienna, Vienna, Austria; Vienna Hepatic Hemodynamic Lab, Division of Gastroenterology and Hepatology, Department of Medicine III, Medical University of Vienna, Vienna, Austria; Clinical Research Group MOTION, Medical University of Vienna, Vienna, Austria.
Insights
A new machine-learning model, the Cirrhosis Risk Identifier (CIRI), accurately predicts hepatic decompensation in patients with compensated advanced chronic liver disease (cACLD). CIRI offers a non-invasive alternative to the gold-standard HVPG measurement for risk stratification.
Area of Science:
- Hepatology
- Machine Learning in Medicine
- Clinical Prognostics
Background:
- Compensated advanced chronic liver disease (cACLD) requires accurate prognosis for patient management.
- Hepatic venous pressure gradient (HVPG) is the gold standard but is invasive.
- Existing non-invasive methods have limitations in risk stratification for cACLD.
Purpose of the Study:
- To develop and validate the Cirrhosis Risk Identifier (CIRI), a machine-learning model for predicting first decompensation in cACLD.
- To benchmark CIRI's performance against established prognostic tools like HVPG, LSM, FIB-4, and MELD.
- To assess CIRI's utility in non-invasively identifying high-risk cACLD patients.
Main Methods:
- Development and validation of the CIRI model using demographic and routine laboratory data.
- Training and internal validation in a large US cohort (Optum CDM).
- External validation in a prospective European cohort with HVPG measurements.
Main Results:
- CIRI demonstrated robust performance, with AUROCs comparable to HVPG and superior to LSM, MELD, and FIB-4 in predicting decompensation.
- The model independently predicted decompensation in both US and European cohorts.
- A specific CIRI cutoff identified patients at a decompensation risk similar to HVPG ≥10mmHg.
Conclusions:
- The machine-learning-based CIRI model provides accurate risk stratification for cACLD patients.
- CIRI's non-invasive nature and performance comparable to HVPG suggest its potential for widespread clinical use.
- CIRI may help identify at-risk cACLD patients, guiding surveillance and preventive strategies.
Background & Aims:
Hepatic venous pressure gradient (HVPG) is the gold standard for assessment of prognosis in compensated advanced chronic liver disease (cACLD), yet its invasiveness limits broad clinical use. We developed and validated the Cirrhosis Risk Identifier (CIRI) model, a machine-learning tool using 11 demographic and routine laboratory parameters trained to prognosticate first decompensation; and benchmarked its performance against HVPG, liver stiffness measurement (LSM), FIB-4 and MELD.
Methods:
Patients with cACLD were identified in the US Optum Clinformatics Data Mart (Optum CDM) for model development and internal validation and externally validated in a prospective European tertiary care cohort. The primary endpoint was first decompensation (ascites, encephalopathy, variceal bleeding) with hepatocellular carcinoma and death as competing events.
Results:
CIRI was trained on 112,618 and internally validated on 18,852 patients with cACLD in Optum CDM, with 210 HVPG-characterized patients with cACLD (European cohort) included for external validation. In both Optum CDM and Europe, steatotic liver disease was the leading etiology (54%; 44%), with median follow-up of 18.5 months (IQR 7.2-40.4) and 27.5 months (21.2-34.8), respectively. In Optum CDM, CIRI showed higher 1- and 2-year time-dependent AUROCs (0.816, 0.815) than MELD and FIB-4 (both p <0.001). In Europe, AUROCs (0.836, 0.769) were comparable to HVPG (both p >0.900) and superior to LSM (both p <0.05). CIRI predicted decompensation independently (Optum CDM: adjusted subdistribution hazard ratio [aSHR]: 1.67, p <0.001; Europe: aSHR: 1.66, p = 0.017) with a cut-off of ≥-8.25 identifying patients at comparable decompensation risk as HVPG ≥10 mmHg (CSPH).
Conclusion:
The machine-learning-based CIRI model yielded robust prognostic performance and accurate risk stratification in patients with cACLD. CIRI's discrimination of decompensation risk was comparable to that of HVPG, highlighting its potential future utility to non-invasively identify "at-risk" patients with cACLD.
Impact And Implications:
Hepatic decompensation marks the key clinical transition from compensated to decompensated advanced chronic liver disease and is associated with substantially worse prognosis. However, current risk assessment remains limited by the invasiveness of the gold-standard hepatic venous pressure gradient (HVPG) measurement and the infrastructure required for established non-invasive tests such as liver stiffness measurement (LSM). In this study, we developed and validated "Cirrhosis Risk Identifier" (CIRI), a machine-learning model based solely on routine laboratory and demographic data, which predicted first hepatic decompensation with performance comparable to HVPG and superior to widely used non-invasive tests including LSM, MELD and FIB-4. These findings are important for clinicians and researchers aiming to identify patients with cACLD who are at increased short-term risk of hepatic decompensation. Pending further prospective validation, CIRI could support scalable and repeatable risk stratification and help guide surveillance intensity and targeted strategies intended to prevent hepatic decompensation for high-risk patients.
Clinical Trial Number:
NCT03267615.
