Machine Learning Helps Predict Portal Hypertension in Chronic Liver Disease

Machine learning models may have clinical utility for predicting portal hypertension in patients with compensated advanced chronic liver disease.

Machine learning models (MLMs) may identify clinically significant portal hypertension (CSPH) or severe portal hypertension (PH) in patients with compensated advanced chronic liver disease (cACLD), investigators reported in the Journal of Hepatology.

The multicenter study evaluated the ability of different MLMs using only noninvasive readouts to predict the risk for severe PH and CSPH in patients with cACLD.

The internal VIENNA cohort included 163 patients with cACLD from the prospective Vienna Cirrhosis Study from 2017 to 2021. External validation was conducted with 1069 patients with cACLD from multiple international collaborators.

Participants in the VIENNA cohort were assessed for hepatic venous pressure gradient (HVPG) measurement, patient demographics, disease activity, and clinical and biochemical parameters. Severe PH and CSPH were defined by HVPG at least 16 mm Hg and at least 10 mm Hg, respectively. The researchers applied 3-parameter (3P) and 5-parameter (5P) classification MLMs to the combined external patient datasets for validation.

The study population included 1232 patients with cACLD (VIENNA; n=163; external cohort from 7 sites, n=1069). The proportion of patients with cACLD who had CSPH and severe PH in VIENNA (67.4%/35.0%) and the external cohort (70.3%/34.7%) was comparable.

The presented 5P and 3P MLMs have promising clinical utility for the noninvasive prediction of HVPG ≥10mmHg and HVPG ≥16mmHg in cACLD patients.

The most suitable 3- and 5-parameter sets were identified for MLMs for PH risk prediction. The 3P model included platelet count (PLT), total serum bilirubin (BILI), and international normalized ratio (INR), and the 5P model also included cholinesterase and gamma-glutamyl transferase and activated partial thromboplastin time in place of INR.

The best performing models were logistic regression (LR), random forest (RF), XGBoost, support vector machine, and multilayer perceptron. The 5P and 3P models outperformed liver stiffness measurement alone to predict severe PH, and the 5P and 3P MLMs for predicting severe PH resulted in area under the curve (AUC) values above 0.739 in all models, reaching 0.887 and 0.813, respectively, with LR.

The 5P and 3P models outperformed the models trained on all parameters. For predicting CSPH, logistic regression had the best performance, with AUC scores of 0.813 (5P) and 0.784 (3P).

The 5P and 3P models were then applied to the external cohorts for validation. For predicting CSPH, the internally trained 5P model had reliable performance in 1 dataset (LR: AUC=0.691). Validation of the internally trained 3P MLMs in predicting CSPH demonstrated more robust performance, with AUCs greater than 0.8 in 3 datasets.

When applying the HVPG of at least 16 mm Hg threshold, the internally trained 5P model attained an AUC of 0.694 with random forest in 1 dataset and 0.521 in another dataset with LR. The internally trained 3P MLMs reached an AUC greater than 0.85, but the performance was heterogeneous.

The study authors also assessed the prediction performance of the 5P and 3P models in the merged cohort of all study datasets with use of repeated cross-validation. For CSPH prediction, LR had better performance vs other models (AUC=0.773 with 3P; AUC=0.754 with 5P). For predicting severe PH, the LR models also had the best performance (5P AUC=0.812; 3P AUC=0.735).

For the single-dataset resolution, the validation results improved slightly and were less heterogenous vs the internally trained setting. The AUC scores were greater than 0.625 in all cohorts for predicting CSPH and severe PH.

Among several study limitations, the patient datasets in the validation cohort were heterogeneous, and the findings may not reflect under-represented disease etiologies such as cholestatic and rare liver diseases.

“The presented 5P and 3P MLMs have promising clinical utility for the noninvasive prediction of HVPG ≥10mmHg and HVPG ≥16mmHg in cACLD patients,” the study authors noted. “This approach can be clinically used for prioritization for treatment to prevent decompensation and for selection of patients for clinical trials.”

Disclosure: Some study authors declared affiliations with biotech, pharmaceutical, and/or device companies. Please see the original reference for a full list of authors’ disclosures.

References:

Reiniš J, Petrenko O, Simbrunner B, et al. Assessment of portal hypertension severity using machine learning models in patients with compensated cirrhosis. J Hepatol. Published online September 21, 2022. doi:10.1016/j.jhep.2022.09.012