Validation of a pan-ELastography Machine-learning (ELM) score to predict clinically significant portal hypertension in compensated advanced chronic liver disease

BACKGROUND: Clinically significant portal hypertension (CSPH) drives decompensation and mortality in advanced chronic liver disease (ACLD). Although non-selective β-blockers (NSBB) reduce risk, accurate identification of patients with CSPH requires invasive hepatic venous pressure gradient (HVPG) measurement. The non-invasive Baveno-VII CSPH criteria based on liver stiffness measurement (LSM) and platelet count (PLT)-yield 40-50% indeterminate ("gray-zone") results and vary across etiologies and elastography techniques. Spleen stiffness measurement (SSM) has been proposed to improve the accuracy of the Baveno-VII CSPH criteria. We developed and validated a machine-learning (ML) model integrating pan-elastographic LSM and SSM results with clinical variables to improve CSPH rule-out and rule-in accuracy while minimizing indeterminate cases. METHODS: We analyzed 1,435 compensated ACLD patients with paired HVPG, LSM, SSM, and clinical parameters. LSM and SSM were obtained by vibration-controlled transient elastography (VCTE), two-dimensional shear-wave elastography (2D-SWE), or point-SWE (p-SWE). Models were trained (n=943) and internally validated (n=150) using harmonized LSM/SSM from different technologies and clinical variables (PLT, Child-Pugh, age, gender, etiology). Cut-offs were selected for 100% negative predictive value (NPV) to rule-out and 100% positive predictive value (PPV) to rule-in CSPH. External validation was conducted in 342 patients across seven centers, comparing ML performance against Baveno VII, Baveno-SSM single- and dual-cut-off criteria, and, in the VCTE subgroup, ANTICIPATE and NICER scores. RESULTS: A Random Forest-based model based achieved the highest performance (external validation: AUC=0.91, Brier Score=0.13), with cut-offs≤0.45 (rule-out) and ≥0.60 (rule-in) yielding NPV=0.90 (95%C.I.0.84-0.94) and PPV=0.96 (95%C.I.0.92-0.98). The ML gray-zone was 12.3%, versus 47.9% (Baveno VII;p