Non-Invasive Machine Learning Approach Identifies Portal Hypertension

abdomen tomography
A machine learning algorithm was developed to detect clinically significant portal hypertension CT and MRI scans.

A machine learning algorithm developed to detect clinically significant portal hypertension (CSPH) from computer tomography (CT) scans or magnetic resonance images (MRI) could identify patients with CSPH more successfully than invasive clinical approaches, according to results published in Clinical Gastroenterology and Hepatology.

Liver and spleen images were obtained from patients with cirrhosis within 14 days of recording their hepatic venous pressure gradient through transjugular catheterization. The CT scan cohort (n=679 patients providing 10,014 liver and 899 spleen images) were scanned between 2016 and 2017 and the MRI scan cohort (n=271 patients providing 45,554 liver and spleen images) were collected between 2018 and 2019 for the CHESS1701 and CHESS1802 studies, respectively.

All data were collected from 9 centers, 8 located in China and 1 in Turkey. CSPH was defined by a venous pressure gradient ≥10 mmHg. All patients were sampled 6 times, and the samples were randomly assigned in a 3:1:1 ratio for training, validation, or test datasets.

Related Articles

The machine learning model was trained using deep convolutional neural networks.

For the CT cohort, the synthesized model was able to identify CSPH with an area under the receiver operating characteristic curve (AUC) of 0.912 (95% CI, 0.854-0.971) in the validation set and 0.933 (95% CI, 0.883-0.984) in the test set. The investigators observed a similarly accurate performance of their model for the MRI cohort, with an AUC of 0.924 (95% CI, 0.833-1.000) in the validation set and 0.940 (95% CI, 0.880-0.99) in the test set. The investigators reported that they tested their model 6 times, finding AUC values >0.888 with no significant difference (P >.05) among their results.

The investigators compared the results of their non-invasive diagnostic tool to commonly implemented invasive clinical practices and reported lower AUC values than their machine-learning approach. Specifically, fibrosis index was reported as AUC 0.805 (95% CI, 0.770-0.841), CSPH risk score as 0.792 (95% CI, 0.752-0.833), Lok score as 0.791 (95% CI, 0.49-0.832), FIB-4 as 0.762 (95% CI, 0.723-0.801), and APRI as 0.742 (95% CI, 0.701-0.784).

The limitations of this study include the fact that they base their model on hepatic venous pressure gradient measurements in patients with cirrhosis, which is not a typical clinical measurement collected in this patient population. Furthermore, this algorithm was designed as a classifier, and it is unclear if their model could be used to monitor changes to liver and spleen health over time.

The investigators concluded that their deep learning model was able to analyze CT and MRI data in patients with cirrhosis and identify CSPH through a non-invasive and rapid approach.

Follow @Gastro_Advisor


Liu Y, Ning Z, Ōrmeci N, et al. Deep convolutional neural network-aided detection of portal hypertension in patients with cirrhosis [published online March 20, 2020] Clin Gastroenterol Hepatol doi:10.1016/j.cgh.2020.03.034