Prediction of sepsis using artificial neural network and optimal brain surgeon

Abstract

Sepsis is a severe threat to global health. Approximately, the mortality rate of sepsis in the Intensive Care Unit (ICU) is 42%. In 2017, 11 million sepsis-related deaths were reported among 49 million cases, 20% of all-cause of deaths worldwide. Detection and prediction of sepsis in earlier stage allow patients to get earlier care and better results, but sepsis is often unknown until the late stages. Substantial bodies of research studies on sepsis prediction have mainly focused on rules-based severity scores, which are transparent and straightforward; unfortunately, they have imperfect sensitivity and specificity in identifying and predicting sepsis. Typically, various sepsis predictions approach that would allow for predicting in an earlier stage, which can reduce the mortality rate and treatment cost. So, machine learning algorithms can be a choice for predicting sepsis. Therefore, this current thesis identified the features influencing early sepsis prediction and examining the features impelling the clinical severity scores used for the prediction of sepsis. The thesis also developed a hybrid optimal brain surgeon algorithm for sepsis prediction and tested the proposed algorithm's accuracy. The research methodology adopted for this thesis is an experimental simulation. The datasets used in this research were adopted from MIMIC-III, which comes with vast electronic health records. A systematic literature review was performed, and significant features of the MIMIC-III dataset for sepsis prediction were obtained by applying Automatic Backward Elimination (ABE) algorithm, Generalized Linear Model (GLM), and Correlation Matrix (CM). After that, the research built a hybrid-sepsis prediction model using machine learning techniques to train and test with selected features for model selection. Then Optimal Brain Surgeon (OBS) algorithm was used to simplify the architecture of the neural network for making an explainable deep learning-based sepsis prediction model. This is where hybridization has taken effect. The pruning algorithm OBS uses Hessian information and considers the time delay for measuring the saliency. Second-derivative information is used to compromise between the difficulty of the network and the training set error. The thesis's finding revealed that the AUROC of the predictive model was 0.882. The hybrid OBS algorithm pruned network is 80.0% with the same accuracy of the prediction model. This result indicates that the proposed hybrid model is efficient with high prediction accuracy and slight complexity compared with some previous prediction techniques. Early prediction of sepsis can reduce mortality rates and save treatment costs among ICU patients.