The early prediction of Sepsis is a highly important task regarding its impact on the mortality ratio for ICU patients. The earlier the Sepsis occurrence is foreseen, the sooner the patient can have the chance to be intervened and the patient can be prevented from getting major organ damage. However, it is a challenging task for healthcare professionals even though they have expert knowledge and standardized prediction scoring systems. Therefore the machine learning models that evaluate the patients’ regularly measured covariates as features, and make predictions by utilizing them have been acknowledged and have gathered attention to automatize and improve the early prediction process. However, the nature of the problem of early prediction of Sepsis requires patients to make an estimation based on past measurements, and reliable estimation cannot be made with the use of instant measurement results alone. To this end, additional features obtained by recent measurements have been proposed to be involved in the machine learning model’s estimation process by basing a specific dataset with a specific definition of the illness. This study broadly evaluates the previously proposed features to measure the effectiveness of their single and joint use. Unlike the previous research, this evaluation was done by multiple datasets, different types of additional features, and various machine learning methods for validation purposes. Experimental results have proved that the proposed machine learning model has a significant effect on early prediction, and the best improvement is obtained with the XGBoost classifier. However, it should be noted that quantity, especially the quality of the features used to feed the model has a critical effect on the model’s performance. Consequently, using additional features enhances the model’s performance dramatically.
Read full abstract