ICU readmission prediction using fuzzy support vector machine

In this study, a modified fuzzy support vector machine (FSVM) is proposed for handling noises, outliers and class imbalance problem using a real-world Intensive Care Unit (ICU) dataset. ICU readmission is a critical problem that is associated with physiological variable selection, imbalanced data, noises and outliers. Previously proposed approaches, such as scoring system (APACHE II) or classical data mining techniques did not tackle these issues. In this paper, the ICU readmission information is represented by feature vectors including blood pressure, respiratory rate, arterial base excess, hematocrit, etc. Sequential forward selection (SFS) is applied to select the best predictors for ICU readmission prediction. A fuzzy membership calculation method is proposed, which uses fuzzy cosine algorithm and fuzzy distance function to modify the fuzzy membership values of each sample to deal with the readmission prediction problem with noises and outliers. Fuzzy weights are adopted based on the different distribution of majority class and minority class. The performance of FSVM technique is compared with other existing imbalance-learning methods, such as over sampling, under sampling, and Synthetic Minority Over-Sampling Technique (SMOTE) using 5-fold cross validation evaluation. The experimental results show that the proposed FSVM is a stable and competitive method for ICU readmission prediction on class imbalance dataset with outliers and noises.