Predictive model of in-hospital mortality in liver cirrhosis patients with hyponatremia: an artificial neural network approach

Abstract

Hyponatremia can worsen the outcomes of patients with liver cirrhosis. However, it remains unclear about how to predict the risk of death in cirrhotic patients with hyponatremia. Patients with liver cirrhosis and hyponatremia were screened. Eligible patients were randomly divided into the training (n = 472) and validation (n = 471) cohorts. In the training cohort, the independent predictors for in-hospital death were identified by logistic regression analyses. Odds ratios (ORs) were calculated. An artificial neural network (ANN) model was established in the training cohort. Areas under curve (AUCs) of ANN model, Child-Pugh, model for end-stage liver disease (MELD), and MELD-Na scores were calculated by receiver operating characteristic curve analyses. In multivariate logistic regression analyses, ascites (OR = 2.705, P = 0.042), total bilirubin (OR = 1.004, P = 0.003), serum creatinine (OR = 1.004, P = 0.017), and international normalized ratio (OR = 1.457, P = 0.005) were independently associated with in-hospital death. Based on the four variables, an ANN model was established. Its AUC was 0.865 and 0.810 in the training and validation cohorts, respectively, which was significantly larger than those of Child-Pugh (AUC = 0.757), MELD (AUC = 0.765), and MELD-Na (AUC = 0.769) scores. An ANN model has been developed and validated for the prediction of in-hospital death in patients with liver cirrhosis and hyponatremia.