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

Mortality prediction for patients in intensive care unit (ICU) is necessary to prioritize resources as well as to help the medical staff to make decisions, and hence more accurate methods for identifying high risk patients are very important for improving clinical care. However, many existing approaches including some scoring systems now being used in the hospital are not good enough since they try to establish a global/average offline model, which may be unsuitable for a specific patient. Thus, a more robust and effective monitoring model adaptable to individual patients is needed. To establish a more personalized model, this study proposes a two-step framework, in which the first step is for clustering and while the second one is for mortality predication. A novel method combining just-in-time learning (JITL) and extreme learning machine (ELM), referred to JITL-ELM, is proposed for mortality prediction, which applies global optimization of variables and neighborhood of appropriate samples to build an accurate patient-specific model. In addition, a simplified JITL-ELM with less key physiological variables is developed. In the experiment, 4000 real clinical records of ICU patients are collected to validate the proposed algorithm, of which the AUC index is 0.8568, which is much better than the existing traditional global/average models, and furthermore the simplified JITL-ELM still performs well.