Establishment and validation of a nomogram model for riskprediction of hepatic encephalopathy: a retrospective analysis.

To establish a high-quality, easy-to-use, and effective risk prediction model for hepatic encephalopathy, to help healthcare professionals with identifying people who are at high risk of getting hepatic encephalopathy, and to guide them to take early interventions to reduce the occurrence of hepatic encephalopathy. Patients (n = 1178) with decompensated cirrhosis who attended the First Affiliated Hospital of Guangxi University of Chinese Medicine between January 2016 and June 2022 were selected for the establishment and validation of a nomogram model for risk prediction of hepatic encephalopathy. In this study, we screened the risk factors for the development of hepatic encephalopathy in patients with decompensated cirrhosis by univariate analysis, LASSO regression and multifactor analysis, then established a nomogram model for predicting the risk of getting hepatic encephalopathy for patients with decompensated cirrhosis, and finally performed differentiation analysis, calibration analysis, clinical decision curve analysis and validation of the established model. A total of 1178 patients with decompensated cirrhosis who were hospitalized and treated at the First Affiliated Hospital of Guangxi University of Chinese Medicine between January 2016 and June 2022 were included for modeling and validation. Based on the results of univariate analysis, LASSO regression analysis and multifactor analysis, a final nomogram model with age, diabetes, ascites, spontaneous peritonitis, alanine transaminase, and blood potassium as predictors of hepatic encephalopathy risk prediction was created. The results of model differentiation analysis showed that the AUC of the model of the training set was 0.738 (95% CI 0.63-0.746), while the AUC of the model of the validation set was 0.667 (95% CI 0.541-0.706), and the two AUCs indicated a good discrimination of this nomogram model. According to the Cut-Off value determined by the Jorden index, when the Cut-Off value of the training set was set at 0.150, the sensitivity of the model was 72.8%, the specificity was 64.8%, the positive predictive value was 30.4%, and the negative predictive value was 91.9%; when the Cut-Off value of the validation set was set at 0.141, the sensitivity of the model was 69.7%, the specificity was 57.3%, the positive predictive value was 34.5%, and the negative predictive value was 84.7%. The calibration curve and the actual events curve largely overlap at the diagonal, indicating that the prediction with this model has less error. The Hosmer-Lemeshow test for goodness of fit was also applied, and the results showed that for the training set, χ