Development and validation of a machine learning-based model for varices screening in compensated cirrhosis (CHESS2001): an international multicenter study

Abstract

Background and aims: The prevalence of high-risk varices (HRV) is low among compensated cirrhotic patients undergoing EGD. Our study aimed to identify a novel machine learning (ML)-based model, named ML EGD, for ruling out HRV and avoiding unnecessary EGDs in patients with compensated cirrhosis.

Methods: An international cohort from 17 institutions from China, Singapore, and India were enrolled (CHESS2001). The variables with the top 3 importance scores (liver stiffness, platelet count, and total bilirubin) were selected by the Shapley additive explanation and input into a light gradient-boosting machine algorithm to develop ML EGD for identification of HRV. Furthermore, we built a web-based calculator for ML EGD, which is free with open access (http://www.pan-chess.cn/calculator/MLEGD_score). Unnecessary EGDs that were not performed and the rates of missed HRV were used to assess the efficacy and safety for varices screening.

Results: Of 2794 enrolled patients, 1283 patients formed a real-world cohort from 1 university hospital in China used to develop and internally validate the performance of ML EGD for varices screening. They were randomly assigned into the training (n = 1154) and validation (n = 129) cohorts with a ratio of 9:1. In the training cohort, ML EGD spared 607 (52.6%) unnecessary EGDs with a missed HRV rate of 3.6%. In the validation cohort, ML EGD spared 75 (58.1%) EGDs with a missed HRV rate of 1.4%. To externally test the performance of ML EGD, 966 patients from 14 university hospitals in China (test cohort 1) and 545 from 2 hospitals in Singapore and India (test cohort 2) comprised the 2 test cohorts. In test cohort 1, ML EGD spared 506 (52.4%) EGDs with a missed HRV rate of 2.8%. In test cohort 2, ML EGD spared 224 (41.1%) EGDs with a missed HRV rate of 3.1%. When compared with the Baveno VI criteria, ML EGD spared more screening EGDs in all cohorts (training cohort, 52.6% vs 29.4%; validation cohort, 58.1% vs 44.2%; test cohort 1, 52.4% vs 26.5%; test cohort 2, 41.1% vs 21.1%, respectively; P < .001).

Conclusions: We identified a novel model based on liver stiffness, platelet count, and total bilirubin, named ML EGD, as a free web-based calculator. ML EGD could efficiently help rule out HRV and avoid unnecessary EGDs in patients with compensated cirrhosis. (Clinical trial registration number: NCT04307264.).