Research On ICU Mortality Risk Prediction Based On Electronic Health Records

Electronic health records cover the entire process from the patient’s admission to discharge,making clinical data-driven prediction possible.Patient outcome prediction in the intensive care unit is closely related to intervention choices,care plans and resource allocation.Accurately assessing patient mortality risk,early identification of high-risk groups with poor prognosis and timely intervention are the key to improve patient survival.Research on patient mortality risk prediction in ICU can be divided into two parts: the feature representation of patients and the construction of mortality risk prediction models.This article starts from these two aspects.In patient representation,this paper explores the important statistics under different mortality risk prediction tasks based on the genetic algorithm.An dynamic integration algorithm based on K-means is proposed to construct a mortality risk prediction model.The details are as follows:(1)On patient representation,statistics-based patient time series representation is a widely used patient representation method.However,in the existing research,only simple statistics or combinations of them are used.Why these statistics are chosen is not justified and there is no related research to evaluate the validity of different statistics in different mortality risk prediction tasks.This article explores the effective statistics under the short-term,in-hospital,and long-term mortality risk prediction tasks in order to provide theoretical and experimental support for evaluating the important statistics of time series under different prediction tasks,and also to provide an effective patient representation for the construction of prediction models.(2)In terms of constructing the mortality risk prediction model,many studies have proposed ensemble learning algorithms based on different ideas.However,few studies have applied clustering algorithms to improve the diversity of the base classifier.And the fusion strategy is based on a unified weight vector for different samples,which is difficult to reflect the personalization of the samples.This paper proposes a dynamic ensemble learning algorithm based on K-means(DELAK)and apply it to predict the mortality risk of ICU patients.To verify the superiority of the proposed algorithm,the DELAK ensemble algorithm is compared with single classifiers,different fusion strategies,classic ensemble algorithms and clinical scoring systems.Through the analysis of different combinations of statistics,this paper finds effective statistics under different mortality risk prediction tasks,which can provide some reference for the subsequent research on constructing effective patient representation.Compared with other ensemble algorithms and fusion strategies,the proposed dynamic ensemble learning algorithm achieves better prediction performance,and it significantly improves the prediction effect of the widely used scoring system in clinical practice,which reflects the effectiveness of the proposed algorithm.