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

BACKGROUND & AIMS: The non-invasive Baveno VII criteria for clinically significant portal hypertension (CSPH), based on liver stiffness measurement (LSM) and platelet count (PLT), yield indeterminate ("gray-zone") results in approximately 40-50% of patients, and their performance varies 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 the ELM score, a machine learning-based model integrating pan-elastographic LSM and SSM with clinical variables to improve accuracy for CSPH diagnosis, while minimizing indeterminate classifications.

METHODS

We analyzed 1,435 patients with compensated advanced chronic liver disease who underwent paired hepatic venous pressure gradient measurements, LSM, SSM, and clinical assessment. LSM and SSM were obtained using vibration-controlled transient elastography, two-dimensional shear-wave elastography, or point shear-wave elastography. Models were trained (n = 943) and internally validated (n = 150) using harmonized elastography measurements and clinical variables. External validation was performed in 342 patients from seven centers. Cut-offs were selected to maximize rule-out and rule-in performance and compared with the Baveno VII criteria, Baveno-SSM criteria, and, in the vibration-controlled transient elastography subgroup, the ANTICIPATE and NICER scores.

RESULTS

A random forest-based model achieved the highest performance in the external validation cohort (AUC = 0.91; Brier score = 0.13). Cut-offs of ≤0.45 (rule-out) and ≥0.60 (rule-in) yielded a negative predictive value of 0.91 (95% CI 0.84-0.95) and a positive predictive value of 0.96 (95% CI 0.92-0.98). The ELM score reduced the gray zone to 12.3%, compared with 47.9% for the Baveno VII criteria (p <0.001), 38.6% for the Baveno-SSM dual cut-off criteria (p <0.001), and 19.6% for the Baveno-SSM single cut-off criteria (p <0.05). In the vibration-controlled transient elastography external validation subgroup (n = 275), the ELM score achieved rule-in performance comparable to that of the ANTICIPATE and NICER scores while reducing the gray zone to 12.0%, compared with 38.2% and 26.5%, respectively.

CONCLUSIONS

The ELM score outperformed the current Baveno VII criteria and substantially reduced the gray zone across elastography modalities, supporting broader and safer non-invasive identification of patients with CSPH who may benefit from non-selective beta-blocker therapy.

IMPACT AND IMPLICATIONS: Clinically significant portal hypertension is central to prognosis in compensated advanced chronic liver disease, yet current non-invasive strategies classify 40-50% of patients as indeterminate. This study introduces a pan-elastography machine-learning model that integrates liver and spleen stiffness with clinical variables to address this limitation. The results are important for hepatologists and healthcare systems because the ELM score reduced the diagnostic gray zone to ∼12% in external validation, compared with 47.9% for Baveno VII, while maintaining high rule-out and rule-in accuracy (NPV up to 0.91 and PPV up to 0.96). In practice, these findings could support earlier non-invasive identification of patients eligible for non-selective β-blocker therapy, more targeted use of HVPG in indeterminate cases, and harmonized risk stratification across elastography modalities, while recognizing that prospective validation and regulatory approval are required before routine clinical implementation.