Machine Learning Classifications of Multiple Organ Failures in a Malaysian Intensive Care Unit

Abstract

Multiple organ failures are the main cause of mortality and morbidity in the intensive care unit (ICU). The progression of organ failures in the ICU is usually monitored using the Sequential Organ Failure Assessment (SOFA) score. This study aims to perform the classification of multiple organ failures using machine learning algorithms based on SOFA score. Ninety-eight ICU patients’ data were obtained retrospectively from Universiti Malaya Medical Centre for analysis. Several machine learning algorithms which are decision tree, linear discriminant, naïve Bayes, support vector machines, k-nearest neighbor, AdaBoost, and random forest were used for the classification. The classifiers were trained on 80% of the patients with 10-fold cross-validations and assessed on 20% of patients using 34 variables in the ICU. The random forest algorithm was able to achieve 99.8% accuracy and 99.9% sensitivity in the training dataset. Meanwhile, the AdaBoost algorithm achieved 99.1% sensitivity in the testing dataset. This study demonstrates the performances of different machine learning algorithms in the classification of multiple organ failures. The feature selection shows respiratory rate and mean arterial pressure (MAP) as the most important variables using chi-square test while insulin and fraction of oxygenated hemoglobin are the most important predictors by the mutual information test.