Study On Risk Assessment Of Acute Coronary Syndrome Based On EHR Data

Cardiovascular disease is one of the leading causes of death worldwide,and Acute Coronary Syndrome(ACS)is one of them.It is important to assess the risk of adverse events in the early treatment stage for ACS patients,such as myocardial infartion,stroke and death.Individualized risk assessment on ACS patients can help physicians choose the most appropriate treatment strategy to reduce ACS mortality.Many traditional scoring models such as GRACE,TIMI,etc.are mainly established based on cohort studies.Although valuable,the samples usually deviate from clinical practice due to the strict enrollment criteria of these cohort studies.In addition,the use of a few risk factors limits the performance of the cohort-based models and makes them difficult to incorporate new risk factors to improve the prediction performance.With the development of hospital informatization in recent years,many machine learning based models have been developed for risk assessment.These models utilize electronic health record(EHR)data faithfully documents various clinical information and thus can be utilized to achieve promising prediction performanceHowever,there are still some problems in EHR-based models.Specifically,existing models usually ignore association information,such as the association between different samples,between different features,and between different diagnostic subtypes(STEMI,NSTEMI,UA)in ACSTo address this problem,this thesis proposes a relational regularized ACS patient risk assessment model,which uses laplace operator to encode the association information between samples and features in the form of regular terms to be introduced in the model learning process.In addition,noted that the existing risk assessment models are mostly shallow models,this thesis proposes a regularized deep learning ACS risk assessment model,which introduces the correlation information between samples into a stacked autoencoder in the form of regular terms.In this manner,the merits of deep neural network architecture can be utilized to imporve the performance of risk prediction.Moreover,taking into account the different subtypes of ACS,this thesis proposes a multi-task adversarial learning model for predicting major adverse coronary events(MACE)in patients with subtypes of ACS based on adversarial networks.The proposed model combines multi-task learning and adversarial training in a joint manner,and can extract specific and shared latent features of each subtypes of ACS to predict major adverse cardiac events in a refined manner.This thesis used EHR data collected from a Chinese hospital to evaluate the proposed model.The experimental results show that our proposed regularization can effectively use the association information between samples and features to improve the performance of risk assessment.As well,the multi-task adversarial learning model can take advantage of the specificities and commonalities between subtypes of ACS to achieve promissing performance in predicting adverse cardiac events.The methodology proposed in this thesis can achieve comparative prediction performance with the state of art machine learning models in ACS risk prediction,make full use of the potential of electronic medical record data,and pave the way for the risk assessment research of other diseases.