Mortality prediction in critically ill patients using machine learning score

Scoring tools are often used to predict patient severity of illness and mortality in intensive care units (ICU). Accurate prediction is important in the clinical setting to ensure efficient management of limited resources. However, studies have shown that the scoring tools currently in use are limited in predictive value. The aim of this study is to develop a machine learning (ML) based algorithm to improve the prediction of patient mortality for Malaysian ICU and evaluate the algorithm to determine whether it improves mortality prediction relative to the Simplified Acute Physiology Score (SAPS II) and Sequential Organ Failure Assessment Score (SOFA) scores. Various types of classification algorithms in machine learning were investigated using common clinical variables extracted from patient records obtained from four major ICUs in Malaysia to predict mortality and assign patient mortality risk scores. The algorithm was validated with data obtained from a retrospective study on ICU patients in Malaysia. The performance was then assessed relative to prediction based on the SAPS II and SOFA scores by comparing the prediction accuracy, area under the curve (AUC) and sensitivity. It was found that the Decision Tree with SMOTE 500% with the inclusion of both SAPS II and SOFA score in the dataset could provide the highest confidence in categorizing patients into two outcomes: death and survival with a mean AUC of 0.9534 and a mean sensitivity 88.91%. The proposed ML score were found to have higher predictive power compared with ICU severity scores; SOFA and SAPS II.