Machine learning methods for prediction in intensive care

Abstract

Background: An average ICU patient is estimated to be described by more than 200 dieren t parameters, making it likely that there is more information in this data than what is currently being extracted from it by humans. Machine learning methods could assist clinicians by analysing this large amount of ICU data to build models that predict the occurrence of specic clinical problems earlier than an experienced intesivist would. Purpose: To evaluate the applicability of machine learning methods for predicting kidney dysfunction and predened hyper-inammatory states. Materials and Methods: A database of 1548 patients from a randomised controlled trial studying intensive insulin therapy in a surgical intensive care unit was used for this study 3 . A total of 12 prediction tasks were considered, consisting of development and recovery from hyper-inammatory states and kidney dysfunction, with predictions ranging from 1 to 4 days in advance. Four dieren t machine learning algorithms were used: Decision trees (DT), First Order Random Forests (FORF), Naive Bayes (NB) and Tree Augmented Naive Bayes(TAN). Results: Criteria for discrimination and calibration were Area Under the Receiver Operator Characteristic Curve (aROC) of at least 80% and a HosmerLemeshow H-statistic p-value greater than 0:05. Except for the prediction of development of inammation, all prediction tasks regarding development satised the required criteria. While recovery from kidney dysfunction was predicted up to 4 days in advance, none of the predictions of recovery from hyper-inammatory states satised the criteria completely. Table 1 shows results of a subset of prediction tasks. Conclusions: For the ICU database studied and the predictive tasks considered, standard machine learning techniques, result in predictive models with good performances according to the selected criteria.