Study of ICU Mortality Prediction and Analysis based on Random Forest

Accurately predicting the mortality of ICU patients is a tricky problem in modern clinical medicine. Promoting the mortality prediction can effectively improve the ICU utilization and allow patients in need to enter the ICU as soon as possible. Due to the incomplete collection of patient vital signs, the prediction of patient death usually involves a large component of manual intervention. For example, doctors need to pre-classify patient background information and manually judge whether the patient will die In the light of their experience, etc. There is no complete set of vector features that can be used. ICU mortality prediction in ICU still lacks a unified vector for feature selection. This paper used random forest to predict and analysed ICU patient death according to the data set downloaded from Kaggle website which emphasis on the chronic condition of diabetes, through data from MIT's GOSSIS (Global Open-Source Severity of Illness Score) initiative. Our model approaches encouraging performance (Accuracy=0.9241, F1-score=0.96, Recall=0.99, Precision=0.93), and the most important features are selected, the feasibility of unified vector modelling is proved.