Research On Severe Patients' Health Monitoring Using Data-driven Methods

As an important part of the hospital,intensive care unit(ICU)concentrates on the most advanced equipment and excellent medical staff in the hospital,it mainly receive and cure critically ill patients who are life-threatening but may be saved.Mortality prediction and health status monitoring will be carried out after patients enter ICU,so as to achieve optimal allocation of medical resources and good treatment results.However,due to the diversity of patients' state,the general scoring models now widely applied in hospitals discard too much data information,and cannot track the real-time state of patients,so they are difficult to give accurate prediction results.In such a situation,this paper focuses on the establishment of personalized model for each ICU patient,and the main contents are expressed as follow:Aiming at mortality prediction,the research proposes "two-step JITL-ELM" algorithm,which JITL-ELM is used for classification modelling after clustering,JITL collects similar samples from database as the training data,and then ELM builds a local individual model.In this approach,the clustering process reduces the query scope of JITL and improves the searching speed.To solve the problem of missing measurement of some physiological variables,this paper proposes "simplified JITL-ELM",in which only 10 physiological variables are used,and the AUC index is close to the traditional ELM method which uses all variables.Finally,data from 4000 patients in PhysioNet website were selected for experiments.The results showed that the effect of JITL-ELM is better than that of traditional methods,especially the SAPS-I scoring system.In view of health status monitoring,the paper first uses the locally weighted projection regression(LWPR)to approximate the complex nonlinear process with local linear models,in which principal component analysis(PCA)could be further applied to status monitoring,and finally a global weighted statistic will be acquired for detecting the possible abnormalities.Moreover,some improved versions are developed,such as LWPR-MPCA,which also show better performance.Eighteen subjects were selected from Physionet for online monitoring,the experimental results demonstrated that the proposed method can track the real-time status of the patient and is highly sensitive to patients with abnormal state,and it performs best both in learning time and abnormal state detection rate,compared with the traditional PCA and real-time learning PCA algorithm.