State of the Art Methodology for Prediction of Patient’s Mortality within the Early Hours of ICU
Patients who are brought to Intensive Care Units (ICUs) are chronically ill or disabled and need intensive treatment. These patients are at much greater risk of dying than the average patients in hospitals. The primary goal of this paper is to cluster and segregate high-risk patients so that they can seek more effective and aggressive treatment and decrease their chance of fatality. An early estimate of the survival of the patient is based on initial (first 24hrs) test tests, chart events, and patient demographic information were established to accomplish this completely. The secondary purpose of this initiative is for the classifiers to identify low-risk patients who do not need to be handled as aggressively in order to provide appropriate treatment, thereby reducing medical costs. This paper has employed eight classifiers over feature selection and extraction methods involving PCA, LDA, LLE, t-SNE, and Chi2 Test were used for novel preprocessed dataset files: Linear Support vector (SVM), K Nearest Neighborhood (KNN), Linear Logistic Regression, Naive Bayes, Decision Tree, Random Forest, Boosted Trees, Regular SVM, and Multilayer Perceptron (MLP). Several studies were performed using the well available and labeled medical dataset called Medical Information Mart for Intensive Care III (MIMIC-III) to gain an understanding of quality of the proposed method.
Keywords–Mortality Prediction, ICU, Patients, T-distributed Stochastic Neighbor Embedding, Principal Component Analysis, Linear Discriminant Analysis, MIMIC-III, Random Forest, Boosted Tree, Decision Trees