Predictive classification of ICU readmission using weight decay random forest

Abstract

Intensive care unit (ICU) readmissions of critically ill patients result in significant increases in mortality rates and costs, but most readmissions could be avoided. Therefore, the medical management community has devoted considerable effort to developing predictive classifications for ICU readmissions. However, the existing classification methods lack effective feature engineering and are dependent on large quantity of imbalanced and sparse data. In this paper, we use an objective quantitative data set to estimate the probability of ICU readmission for patients who have been transferred from the ICU to the general ward at various risk levels. To implement valuable feature selection for imbalanced time series data, we integrate the missing value analysis and the likelihood ratio test for the distribution characteristics of time series indicators and introduce a weight decay random forest model to achieve ICU readmission classification based on sparse data. Using these approaches, we can rank the most relevant factors that affect the probability of ICU readmission and identify the missing indicators that have the greatest impact on ICU readmission classification. Comprehensive experimental results show that our proposed method can outperform other traditional methods according to seven different performance indicators.