fiLiver and spleen elastography-based prediction models for clinically significant portal hypertension and high-risk esophageal varices: a systematic review of conventional and machine-learning approaches.

The presence of clinically significant portal hypertension (CSPH) and high-risk esophageal varices (HREV) are major predictors of decompensation and mortality in chronic liver disease. Liver and spleen elastography are validated noninvasive markers of portal pressure, and increasing numbers of studies have integrated elastography measurements with conventional statistical and machine-learning (ML) models to predict these outcomes. This systematic review identified, critically appraised and synthesized studies developing or validating prediction models incorporating liver and/or spleen elastography for predicting CSPH and/or HREV. A structured literature search identified studies developing or validating prediction models using transient elastography, shear-wave elastography or magnetic resonance elastography in adults with chronic liver disease. Risk of bias and applicability were assessed using PROBAST+AI. Owing to substantial heterogeneity in elastography modalities, prediction algorithms and outcome definitions, findings were synthesized narratively. Eleven studies published between 2015 and 2025 met the inclusion criteria. Three evaluated machine-learning algorithms using elastography as an input feature (LightGBM, logistic regression/XGBoost hybrid and artificial neural network), whereas eight used conventional cutoff or regression approaches. The best-performing model combined liver stiffness, spleen stiffness and fibrosis-4 within logistic regression and XGBoost, achieving AUROCs of 0.902-0.927 across three independent datasets for HREV and outperforming earlier liver-stiffness-only models. The two newest multicenter externally validated ML studies showed low risk of bias, whereas older, smaller single-center studies frequently demonstrated high or unclear risk, particularly within the analysis domain because external validation was lacking. ML-assisted models integrating liver and spleen elastography with laboratory parameters show promising performance, warranting further external validation before widespread clinical implementation.