# Blood-based Vienna 3P/5P risk models accurately predict first hepatic decompensation in compensated advanced chronic liver disease

## TL;DR

Blood-based 3P/5P models accurately predict liver decompensation in patients with chronic liver disease as well as invasive tests, offering a non-invasive alternative.

## Contribution

The 3P/5P models are shown to be as effective as invasive HVPG for predicting decompensation in cACLD patients.

## Key findings

- The 5P model achieved an AUROC of 0.800 for predicting severe portal hypertension.
- The 5P model's time-dependent AUROC for decompensation prediction was comparable to HVPG.
- 3P and 5P models outperformed LSM and ANTICIPATE±NASH in decompensation prediction.

## Abstract

Invasive measurement of hepatic venous pressure gradient (HVPG) is the gold standard for diagnosing clinically significant portal hypertension (CSPH, i.e. HVPG ≥10 mmHg), which indicates an increased risk of decompensation. We evaluated the blood-based Vienna 3P/5P models for non-invasive assessment of portal hypertension (PH) severity and their prognostic value. Their performance was compared to HVPG, liver stiffness measurement (LSM) and the ANTICIPATE±NASH model.

Patients with compensated advanced chronic liver disease (cACLD) who underwent HVPG measurement and LSM within the prospective VICIS (Vienna Cirrhosis Study) were included. We assessed the ability of each model to detect CSPH and severe PH (HVPG ≥16 mmHg), predict decompensation, and stratify risk. Outcome prediction was further validated in an external cohort.

Among 266 patients with diverse etiologies of cACLD, median HVPG was 11 (8-16) mmHg with a CSPH and severe PH prevalence of 62.8% and 25.6%, respectively. The 3P/5P models correlated with HVPG (both p <0.001), achieving AUROCs of 0.704 (5P) for CSPH and 0.800 (5P) for severe PH prediction. During a median follow-up of 23.9 (15.3-32.6) months, 48 (18%) patients decompensated. HVPG and the 5P model showed similar time-dependent AUROCs (at 0.5 and 1 year: 0.753-0.822), superior to ANTICIPATE±NASH (AUROCs: 0.689-0.691) and LSM (AUROCs: 0.621-0.636). The 5P (adjusted subdistribution hazard ratio [aSHR]: 1.32, p<0.001) and 3P (aSHR: 1.15, p = 0.010) models predicted decompensation independent from age, sex, LSM, etiological cure and non-selective beta blocker use. Proposed cut-offs for the 3P/5P models distinguished between patients at low and high risk of decompensation (Grays test p <0.001).

The blood-based 3P/5P models demonstrated significant prognostic value for predicting hepatic decompensation and identifying patients with cACLD at high risk. Importantly, the 5P model performed comparably to HVPG.

This study addresses the clinical need for accessible, reliable, and cost-effective non-invasive tools to predict hepatic decompensation in patients with compensated advanced chronic liver disease, given the limited availability of hepatic venous pressure gradient and liver stiffness measurement. By demonstrating that the Vienna 3P/5P models – machine learning tools based solely on routine laboratory parameters – achieve comparable prognostic accuracy to hepatic venous pressure gradient and outperform other non-invasive tools, such as liver stiffness measurement or the ANTICIPATE±NASH model, these findings have significant implications for clinicians providing care for patients with compensated advanced chronic liver disease. The models' simplicity, repeatability and wide availability could facilitate timely risk stratification and improved clinical management across diverse healthcare settings.

NCT03267615.

Image 1

•Blood-based 3P/5P machine learning models predict decompensation in cACLD.•The 5P model has similar prognostic value as invasive HVPG measurement.•Both models perform favourably against LSM and ANTICIPATE±NASH.•3P/5P models might support individualized prognostication and treatment in cACLD.

Blood-based 3P/5P machine learning models predict decompensation in cACLD.

The 5P model has similar prognostic value as invasive HVPG measurement.

Both models perform favourably against LSM and ANTICIPATE±NASH.

3P/5P models might support individualized prognostication and treatment in cACLD.