Using machine learning approach to predict short-term mortality risk of acute myocardial infarction after emergency admission

The short-term mortality risk of Acute Myocardial Infarction (AMI) remains high all over the world. Accurately predicting risk of death may assist physicians' decisions on medical treatments and advance care planning for patients. The Taiwan National Health Insurance Research Database (NHIRD) was applied to construct a model to predict whether a diagnosed AMI patient could turn into a severe condition within a short interval. We used IQR and propensity score matching to select control group subjects, and Kaplan Meier method to define a duration boundary for distinguishing two patient groups. Both chi-square test and p-value were applied to evaluate and screen statistical significance of each feature. An XGBoost model was applied to construct a prediction model and compare prediction results at different stages. System performance was optimized by tuning parameters based on increasing recall rate as much as possible. The proposed prediction system can determine mortality risk of patients immediately and prevent delays in treatment. The results showed that although various attempts were tried to optimize the system with different feature combinations, the prediction performance was yet unsatisfied due to NHIRD providing limited longitudinal diagnosed records and medicine information only. No biometric feature was collected in the NHIRD and it only achieved a low recall rate of 56%. It is firmly believed that features obtained from biometric screening test could provide important factors for a better prediction system and precision treatment in practical clinical applications.