Effects of time-series data on machine learning-based survival analysis models for predicting readmission events in intensive care units

Abstract

One of the problems in the intensive care unit (ICU) is the lack of ICU beds. In proper bed management, it is important to predict readmission events, in which a patient who has been discharged is readmitted to the ICU. There have been many studies using survival analysis models for predicting readmission events in an ICU, which consider the last examination results such as vital signs and blood tests. However, there are few reports using survival analysis models for predicting readmission events in an ICU, which consider the past examination results as time-series features. On the other hand, clinicians treat these time-series data as important prognostic factors. Since we think it is meaningful to consider the impact of using time-series data as features in survival analysis models, we used Random Survival Forest (RSF) and DeepSurv as survival analysis models to build a prediction model for ICU readmissions for ICU discharge patients and investigated the impact of time-series data on the model. As a result, we show that time-series data can improve the prediction accuracy of survival time analysis models and survival analysis models considering time-series data may assist clinicians in their predictions and contribute to appropriate bed management.