Establishment and verification of a diagnostic model of liver cirrhosis with spontaneous bacterial peritonitis

Abstract

Objective To screen the influencing factors of liver cirrhosis with spontaneous bacterial peritonitis (SBP) so as to establish and verify a diagnostic model of cirrhosis with SBP. Methods A total of 7 461 patients with liver cirrhosis were collected from 2 Grade-A hospitals in Chongqing (Hospital A and B) between July 2015 and December 2019. According to the occurrence of SBP during hospitalization, they were divided into SBP (n=1 173) and non-SBP (n=6 288) groups. A total of 3 776 patients (70%) from Hospital A were randomly selected as the training set, while the remaining 1 619 patients (30%) and the 2 066 patients from hospital B were subjected as the internal and external validation sets respectively. Univariate and logistic analyses were used to screen variables, and logistic regression, random forest (RF), decision tree (DT) and XGBoost models were subsequently established using the training set. Then, the optimized logistic regression model was built on the basis of the above 4 models. Finally, the 5 models were applied in the internal and external verification sets for evaluating and comparing the diagnostic value of different models for cirrhosis with SBP. Results The machine learning algorithm suggested that the 7 common influencing factors with significance in all models were as follows: decompensation stage (OR=5.354, 95%CI: 3.770-7.803), lymphocyte percentage (OR=0.951, 95%CI: 0.939-0.962), total bilirubin (OR=1.003, 95%CI: 1.002-1.004), abnormal C-reactive protein (OR=1.626, 95%CI: 1.310-2.017), international normalized ratio (OR=1.346, 95%CI: 1.091-1.681), prealbumin (OR=0.990, 95%CI: 0.987-0.993) and model for end stage liver disease (MELD) score (OR=1.038, 95%CI: 1.015-1.063). The AUC of the internal verification of the optimized logistic model was 0.860, with a sensitivity of 0.872 and a specificity of 0.719; while the corresponding values of the external verification were 0.818, 0.662 and 0.812, respectively. Delong test showed that there was no statistical difference in AUC between the optimized logistic regression model and the other models with favorable performance. Conclusion The risk prediction models established by machine learning algorithm for liver cirrhosis complicated with SBP have high diagnostic value. Among them, the optimized logistic regression model has performed well in both internal and external verification, and is able to provide reference for clinical diagnosis of SBP.

Keywords

• liver cirrhosis
• spontaneous bacterial peritonitis
• machine learning
• diagnostic model