ICU Mortality Prediction Based on Key Risk Factors Identification

Predicting ICU mortality and finding key risk factors make sense for both doctors and patients. Although there has been a number of research pertaining to ICU mortality prediction systems and algorithms, plenty of room still exists for improvement in practical prediction results and identification of important risk factors. In this study, we use C5 decision tree model to predict mortality of ICU patients and identify key risk factors. Totally 4367 records of ICU patients from a local grade-A tertiary hospital were selected for motality prediction, including 244 dead records with demographic information and physiological parameters. In order to solve the problem of inconsistent data sampling frequency, we extracted 96 statistical indicators based on the original records, such as the kurtosis value of red blood cells (HXB_kurt), the skewness coefficient of red blood cells (HXB_skew). Totally 41 indicators as the final input of the prediction model were extracted through feature extraction method. The experimental results show that C5 decision tree model outperform C&RT, CHDID, KNN, Logistic, SVM and Random Forest in five different performance indicators. Moreover, worst-case status and state of changes in respiratory, body temperature, care level, diastolic blood pressure and age were found to be the key risk factors.