Construction and validation of a screening model for minimal hepatic encephalopathy in patients with cirrhosis: A...

Clinical practice currently lacks objective and accurate screening tools for minimal hepatic encephalopathy (MHE). Therefore, we aimed to develop an MHE prediction model based on common risk factors.

A total of 514 and 191 cirrhotic patients were included in the training and external validation cohorts, respectively. Best subset selection was applied to screen for predictors. Logistic regression was selected for model development because it outperformed four machine learning algorithms (Random Forest, Adaptive Boosting, Support Vector Machines, and Naive Bayes) in this study. Discrimination, calibration, and clinical decision-making utility of the model were evaluated. Furthermore, the model was compared with the Stroop test for MHE assessment.

From 44 potential predictors, 6 variables were identified as significant and included in the prediction model: upper gastrointestinal bleeding (odds ratio, 4.17; 95% confidence interval, 2.53-6.88), ascites (2.86; 1.59-5.16), albumin (0.76; 0.70-0.82), ammonia-ULN-ratio (5.89; 3.43-10.13), model for end-stage liver disease (1.16; 1.08-1.26), and long-term oral lactulose (0.04; 0.01-0.11). The model exhibited robustness and outperformed the Stroop test for MHE identification, with areas under the receiver operating characteristic curves of 0.882 and 0.867 for the training and validation datasets, respectively. An interactive web-based nomogram is accessible at https://xc-web.shinyapps.io/dynnomapp/.

This model enables rapid MHE screening.