Risk Stratification of Acute Coronary Syndrome using Machine Learning : An analysis of CLEOS-CPDS data

Abstract: Chest pain is one of the most common complaints amongst patients seeking urgent medical care at hospitals. Chest pain can be a symptom of serious cardiovascular disease such as acute coronary syndrome (ACS), however, most underlying causes are benign. Risk stratification in early stages of medical evaluation is difficult. As a consequence, many patients with chest pains are unnecessarily admitted to hospitals. There is precedent for using machine learning (ML) to aid in predicting cardiovascular disease. In this thesis, our goal is to investigate the feasibility of using ML as a complement to safely discharge patients. We use data collected by the ‘Clinical Expert Operating System - Chest Pain Danderyd Study’ (CLEOS-CPDS). Several models are developed to predict the risk of ACS following chest pains and for identifying important factors. Our best performing model on the highest risk class uses the random forest algorithm. The model has a recall score of 0.58 on the highest risk class using a subset of the medical history background. It admits 10 out of 12 patients who ultimately suffers from ACS, however, only 7 out of 12 are classified as high risk. Identified important features are mostly known risk factors, some of which are used in current risk calculations. However, less known factors such as chest pain radiation and associated symptoms, are also identified as important. The conclusion is that it is feasible to use a machine learning model to aid in risk stratification of ACS in early stages of evaluation, but that the current model needs improvement. In future work, a larger and more complete dataset with a longer follow-up period of patients may be highly beneficial to improve the model performance and verify the conclusions of this thesis.