Esophageal varices detection and bleeding risk assessment with artificial intelligence: a systematic review

Abstract

Background and Aims

Esophageal varices (EVs) result from portal hypertension in decompensated liver disease secondary to liver cirrhosis. Diagnosis and grading is done using EGD and mucosal examination. However, interobserver differences may occur when analyzing EGD results. Recent application of artificial intelligence (AI) algorithms yielded mixed results for varices detection and bleeding risk estimation. The capabilities of AI in the detection and grading of EVs need to be evaluated.

Methods

A systematic review was conducted with Preferred Reporting Items for Systematic reviews and Meta-Analyses guidelines. PubMed, EMBASE, and MEDLINE were searched from January 1956 to July 2023. Data were manually identified and extracted by reviewers, assessing outcomes of AI tools used, EV detection accuracies, and bleeding risk prediction accuracies. Average accuracies were derived from result sections or manual calculations.

Results

Sixteen studies with 26,170 patients and 19 AI systems and algorithms were included after a review of 1670 studies. The most common AI systems were artificial neural network and random forest. The categorical boosting machine learning (ML) algorithm was most accurate for prediction of bleeding (100%), whereas the radiomic model ML tool was the least accurate for EV detection (49%). Overall, AI had an average EV detection accuracy of 78.67% and variceal bleed accuracy of 83.2%.

Conclusions

AI could provide an accessible interface for EV prediction and estimation of bleeding risk. Limitations include the dependence on a single dataset for efficacy, assessment with specific AI tools, and potential overinterpretation of broad geographic variances. Newer algorithms should have larger datasets with reproducible validity to strengthen the predictive value in clinical practice.