Predicting intracerebral hemorrhage patients' length-of-stay probability distribution based on demographic, clinical, admission diagnosis, and surgery information

Abstract

The vast majority of patients with intracerebral hemorrhage (ICH) suffer from long and uncertain length of stay (LOS). The aim of our study was to provide decision support for discharge and admission plans by predicting ICH patients' LOS probability distribution. The demographics, clinical predictors, admission diagnosis, and surgery information from 3,600 ICH patients were used in this study. We used univariable Cox analysis, multivariable Cox analysis, Cox-variable of importance (Cox-VIMP) analysis, and an intersection analysis to select predictors and used random survival forests (RSF) - a method in survival analysis - to predict LOS probability distribution. The Cox-VIMP method constructed by us effectively selected significant correlation predictors. The Cox-VIMP RSF model can improve prediction performance and is significantly different from the other models. The Cox-VIMP can contribute to the screening of predictors, and the RSF model can be established through those predictors to predict the probability distribution of LOS in each patient.