Comparison of Predictive Models for Transferring Stroke In-Patients to Intensive Care Unit

Abstract

The cost for Intensive Care Unit (ICU) resources is extremely high and it affects healthcare budget that provides quality healthcare service for patients. Thus, the need for a predictive model for the decision to transfer stroke in-patients to the ICU is very important to utilize the resources effectively. Also, it will help to lower morbidity and mortality rates through earlier detection and intervention. In this research, initially, a Decision Tree (DT) model, an Artificial Neural Network (ANN) model, a Support Vector Machine (SVM) model, and a Logistic Regression (LR) model are evaluated for predicting the need to transfer the stroke in-patients to the ICU or not. The study is conducted on a clinical dataset consisting of 1,415 observations using the vital signs with six variables. This original dataset was having data imbalance and hence the result was misleading. In order to overcome this situation the Synthetic Minority Oversampling Technique was applied on the dataset. A DT model, an ANN model, a SVM model, and a LR model are evaluated again on the balanced dataset for prediction. Tree-based ensemble approaches such as Generalized Boosted Model, Adaptive Boosting (AdaBoost.M1), Random Forest and Bagged AdaBoost (Adabag) are used to improve the accuracy and performance of models. These methods were trained and tested on the balanced stroke in-patients dataset. The boosting model, AdaBoost.M1, and bagging model, Random Forest, achieved better accuracy compared to the other models. Therefore, these two models could be used for helping healthcare professionals in decision-making.