Feature Selection using Generalized Linear Model for Machine Learning-based Sepsis Prediction

Abstract

Sepsis is a life-threatening condition of patients in an intensive care unit. Early sepsis detection can reduce the mortality rate and cost of treatment among the patients of the Intensive care unit (ICU). Machine Learning-based model can be used to predict sepsis early using Electronic Health Record (EHR) which consists of big data. Features selection plays a vital role for reducing overfitting and the accuracy of the ML-based prediction model. In this paper, Generalized Linear Model (GLM) was used to select the significant features related to sepsis using MIMIC-III dataset which is a rational database that contains ICU patient’s data at Beth Israel Deaconess Medical center. In addition, developed a sepsis prediction model using Artificial Neural Network (ANN) and Random Forest (RF) and validated those models using confusion matrix. After that, clinical severity scores were also calculated with the same dataset. Finally, compared the Area Under the Receiver Operating Characteristic (AUROC) between ML-based model and clinical severity score. The accuracy of ML-based prediction model with GLM is better than clinical severity scores like SOFA, qSOFA and SIRS.