Mortality prediction for ICU patients using just-in-time learning and extreme learning machine

Many types of severity or prognostic scoring systems have been developed for patients in intensive care units (ICUs). They provide evaluation of patients' status so that they can get the best distribution of the intensive care. However, the accuracy and reliability of the existing systems is still not ideal. A new combination of just-in-time learning (JITL) and extreme learning machine (ELM) was proposed, aiming at improving mortality prediction accuracy. JITL was utilized to gather the most relevant data samples for patient-specific modeling while ELM was chosen for fast model building. In this study, 4000 records of ICU patients from PhysioNet database were selected, including 554 dead and 3446 survival records in which physiological parameters values were used for mortality prediction. In terms of the area under receiver-operating curve (AUC), JITL-ELM achieved the best performance, compared with ELM, BP neural network, logistic regression model and traditional score models.