Study On Data-driven Major Adverse Cardiovascular Event Prediction For Acute Coronary Syndrome

Major adverse cardiovascular events(MACE)prediction after acute coronary syndrome(ACS),as a pivotal role in preventive nursing and precise treatments,can assist physicians to deliver appropriate treatment interventions,and decrease the mortality and morbidity after ACS.Although traditional cohort-based models,e.g.,GRACE,CRUSADE,etc.,are valuable for MACE prediction,they are developed to predict only one specific type of MACE,cannot incorporate newly-discovered potential risk factors,and lack the ability to tackle with missing data.Recently,with the rapid development of electronic health records(EHR),more and more data-driven models have been proposed to explore the huge potential of EHR.Compared with the traditional cohort studies,the data-driven prediction models are relaxed from the strict inclusion and exclusion criteria,easier to include new potential risk factors into model,and capable of dealing with missing data.However,there still exist several challenges for the EHR data-driven models,such as:1)The EHR data is typically imbalanced as some types of MACE are rarely occurred in clinical practice;and 2)The type-type,feature-feature and sample-sample relations hidden in EHR data are seldom exploited by the existing models.To address the above challenges,this thesis proposes a relational multi-type MACE prediction model based on a novel boosted-resampling framework.Specifically,to tackle with the class-imbalance problem,the boosted-resampling framework is proposed by applying both over-sampling and under-sampling on minority-class and majority-class samples,respectively,and iteratively learning a new and stronger MACE prediction model from the selected small subset data to correct the previously wrongly predicted patient samples.To utilize the potential correlational information among types,features and samples,this thesis proposes to formulate the prediction problem of multi-type MACE as a multi-label classification problem,and to incorporate the relational information into the model by using regularization terms to encode the corresponding correlations.A case study was conducted on an actual ACS clinical dataset which consists of 2,930 ACS patient samples and collected from a Chinese hospital to evaluate the performance of the proposed model.The experimental results show:1)In dealing with the class-imbalance problem,the proposed model can effectively identify the minority samples,which is more robust and suitable for predicting MACE than state-of-the-art machine learning models;2)Integrating type-type,sample-sample,and feature-feature relations into the model can effectively boost the performance of MACE prediction.The EHR data-driven multi-type MACE prediction method proposed in this thesis is significantly superior to both the traditional machine learning methods and cohort study models in predicting multi-type MACE,inlcuding both ischemic and hemorrhage events.The proposed methodology can be smoothly shifted to tackle with other diseases by utilizing a larger scale of EHR data,as a crucial advantage over traditional techniques for clinical risk prediction and prevention.