A machine-learning (Random Forest) model — the ELM Score — was developed and validated in 1,435 compensated advanced chronic liver disease (ACLD) patients to non-invasively predict clinically significant portal hypertension (CSPH, defined by HVPG), integrating liver stiffness (LSM), spleen stiffness (SSM), and clinical variables (platelets, Child-Pugh, age, sex, etiology) across three elastography modalities (VCTE, 2D-SWE, p-SWE).
In external validation (n=342), the ELM Score achieved AUC=0.91 and Brier Score=0.13; rule-out cut-off (≤0.45) yielded NPV=0.90 (95% CI 0.84–0.94) and rule-in cut-off (≥0.60) yielded PPV=0.96 (95% CI 0.92–0.98), with a gray zone of only 12.3% vs. 47.9% for Baveno VII (p<0.001), 38.6% for Baveno-SSM dual (p<0.001), and 19.6% for Baveno-SSM single (p<0.05).
- NPV of 0.90 falls short of the 100% target set during training, meaning some CSPH cases may be missed at the rule-out threshold. - External validation was conducted across seven centers but the study design is retrospective and cohort-based, limiting causal inference. - The model has not yet been validated for outcomes (decompensation, mortality) or prospective NSBB treatment decisions.
The ELM Score can substantially reduce indeterminate (gray-zone) results compared to Baveno VII criteria, potentially allowing more compensated ACLD patients to be confidently classified for or against NSBB prophylaxis without invasive HVPG measurement. Clinicians should note that the rule-out NPV of 0.90 — not 100% — means a small residual miss rate persists before withholding NSBB therapy.
Explore related topics
