Prediction of Icu Readmission Using Lightgbm Classifier

Decision to discharge a patient from an intensive care unit (ICU) is difficult, with a risk of prolonging unnecessarily the stay and a risk of ICU readmission, which is associated with adverse outcomes. We propose to use light gradient boosting (lightGBM) classifier to build decision-making systems to predict which patients are most likely to experience ICU readmission within the first three days. The classifier was developed and tested using MIMICIII database. We extracted many clinical data from the electronic health records (EHR) stored in MIMICIII database. Then, several statistical, temporal and spectral features were extracted. In addition, some feature selection methods were used to select the most informative data. The performance of the classifier was good, with an area under the curve of 78.57%, suggesting the superiority of using LightGBM classifier over previously tested other machine learning (ML) classifiers. An external validation of the method is warranted. This work is financially supported by OPSIDIAN.