Abstract:
Sepsis is blood poisoning disease that occurs when body shows dysregulated host response to an infection and cause organ failure or tissue damage which may increase the mortality rate in ICU patients. As it becomes major health problem, the hospital cost for treatment of sepsis is increasing every year. Different methods have been developed to monitor sepsis electronically, but it is necessary to predict sepsis as soon as possible before clinical reports or traditional methods, because delayed in treatment can increase the risk of mortality with every single hour. For the early detection of sepsis, specifically in ICU patients, different machine learning models i.e., Linear learner, Multilayer perceptron neural networks, Random Forest, Lightgbm and Xgboost has trained on the data set proposed by Physio Net/ Computing in Cardiology Challenge in 2019. This study shows that Machine learning algorithms can accurately predict sepsis at the admission time of patient in ICU by using six vital signs extracted from patient records over the age of 18 years. After comparative analysis of machine learning models, Xgboost, Randomforest and Lightgbm model achieved a highest accuracy of under the range of 0.89-0.96, precision of 0.90-0.96, and recall 0.78-0.96 under the precision-recall curve on the publicly available data. Early prediction of sepsis can help clinicians to implement supportive treatments and reduce the mortality rate as well as healthcare expenses.
