Utilizing Electronic Medical Records To Predict Major Adverse Cardiovascular Event Of Acute Coronary Syndrome

Major adverse cardiovascular event(MACE)prediction and assessment is a technology to study the quantity-dependent relationship of pathogenic risk factors with morbidity and mortality of acute coronary syndrome and other cardiovascular diseases.It is widely accepted as a core during prevention and treatment of those diseases.Prediction results can provide clinical decision support for clinicians to develop reasonable treatment and care plan so as to reduce the patient's risk and medical expenses.Traditional cohort studies employ inclusion and exclusion criteria to control the quality of patients,and select hand-picked risk factors to develop the risk assessment models.Although the cohort models have been widely applied in the real clinical scenarios and received widespread clinical acceptance,they have some inherent limitations such as the differences between cohort patients and real clinical patients caused by inclusion criteria,the limited performance because of only using a few risk factors and the difficulty of including new risk factors.Recently,with the rapid development of electronic medical records(EMR)system,a large volume of EMR has become available,which offers the great potential to develop brand-new MACE prediction models.Compared with cohort study,EMR-based methods do not have strict inclusion criteria,which allows the data to reflect the real clinical scenarios exactly.Moreover,the machine learning algorithms can include new risk factors easily as more patient features available in EMR.However,EMR-based methods still contain some deficiencies:1)EMR data are underutilized 2)Missing data and imprecision data leads to more uncertainties,which influences the model's prediction accuracy.To alleviate the aforementioned limitations,we propose an EMR-based MACE prediction approach.First,to make full use of collected data,we employed two natural language processing(NLP)technologies to extract patient features from their admission records while processing lab test data.After that,four state-of-the-art machine learning algorithms were utilized to train the standalone MACE prediction models.And then,we use rough set theory to calculate the weight of each standalone model,followed by Dempster-Shafer evidence theory to combine the weighted outputs of standalone models to obtain the proposed ensemble MACE prediction model.In particular,we regards the cohort model,GRACE,as an expert who provides valuable "evidences"and combine these evidences into the final model to improve its performance.We comparatively evaluate the performance of the proposed method on a real clinical dataset consisting of 2,930 ACS patients samples collected from the cardiology department of a Chinese hospital.The experimental results indicate:1)Employing NLP technology to mining the unstructured admission records can improve the model's performance significantly.2)The proposed ensemble model achieves the best overall performance in AUC and Accuracy when compared with standalone models and other ensemble models.