Deep Learning-Based Prediction of High-Risk Gastroesophageal Varices Using Two-Dimensional Shear Wave Elastography

Abstract

Aims: Bleeding from gastroesophageal varices (GEV) is a medical emergency associated with high mortality. We aim to construct an artificial intelligence-based model of two-dimensional shear wave elastography (2D-SWE) of the liver and spleen to precisely assess the risk of GEV and high-risk GEV (HRV). Methods: This was a multicenter, prospective study conducted from October 2020 to September 2022 across 12 hospitals in China. Patients with compensated advanced chronic liver disease (cACLD) were enrolled, with informed consent obtained. A total of 1136 liver stiffness measurement (LSM) images and 1042 spleen stiffness measurement (SSM) images generated by 2D-SWE. We leveraged deep learning methods to uncover associations between image features and patient risk; in this manner, we constructed models to predict GEV and HRV. Results: A multimodality deep learning risk prediction (DLRP) model was constructed to assess GEV and HRV based on LSM and SSM images and clinical information. Validation analysis revealed that the area under the curve (AUC) values of DLRP were 0.91 for GEV (95% confidence interval [CI], 0.90–0.93, p < 0.05) and 0.88 for HRV (95% CI, 0.86–0.89, p < 0.01), which were significantly and robustly better than those of canonical risk indicators, including the values of LSM (0.63 and 0.68 for GEV and HRV) and SSM (0.75 for both GEV and HRV). Moreover, the DLRP model outperformed the model using individual parameters. In HRV prediction, the 2D-SWE SSM images (0.75) were more informative than LSM (0.68, p < 0.01). Conclusion: Our DLRP model shows excellent performance in predicting GEV and HRV, outperforming the canonical risk indicators LSM and SSM. Additionally, the 2D-SWE SSM images provided more information and thus better accuracy in HRV prediction than the LSM images. Trial Registration: ClinicalTrials.gov; NCT04546360.