Adaptive State Monitoring Of ICU Patients Based On Just-In-Time Learning And Machine Learning

As the important part in hospital, intensive care units (ICUs) are always equipped with advanced clinical instruments and complete medical service. The ICU patients are treated with the implementation of centralized treatment and careful nursing by experienced medical staff, in order to ensure the survival of patients and reduce mortality as much as possible. After entering the ICUs, patients are served with mortality prediction and health state monitoring for resource allocation and diagnosis. Due to the various states of ICU patients, the offline general-type models widely used for patients' monitoring may fail to adapt to the changing states of ICU patients. According to basic idea "similar input results in similar output", an adaptive model framework combined with just-in-time learning (JITL) and machine learning is proposed for mortality prediction and health state monitoring of ICU patients.A novel combination of just-in-time learning and extreme learning machine (ELM), termed as JITL-ELM, is proposed for mortality prediction. JITL is used to gather the most relevant data samples while ELM is used to build an individualized model.4000 cases with 24 physiological parameters from PhysioNet database are selected for experiments in which 137 features are created. The experiments show the prediction effect of JITL-ELM is improved greatly compared with the traditional evaluation method and single machine learning methods. Similarly to JITL-ELM, just-in-time learning and principal component analysis (PCA), termed as JITL-PCA are combined for ICU patients' health condition monitoring. Twelve subjects were selected from the Physiobank's Multi-parameter Intelligent Monitoring for Intensive Care II (MIMIC II) database, and five vital signs of each subject were chosen. JITL-PCA achieves the best monitoring performance in terms of adaptability to changes in patient status and sensitivity for abnormality detection. The experimental results demonstrate the fault detection rates of JITL-PCA respectively increase by 20% and 47% compared with the traditional PCA and fast moving-window PCA (Fast MWPCA).