Evaluating machine learning models for cardiac irregularities (early detection, prediction for heart disease)

Abstract

The following research’s purpose is to tackle one of the most significant global health challenges, responsible for millions of deaths annually. Heart diseases are a serious and common threat to nowadays world and, despite constant medical advances, predicting cardiovascular diseases remains a problem. To address this issue, after a personal experience regarding heart disease, the researcher focused on understanding and creating four machine learning models with limited domain knowledge. Random Forest, Support Vector Machines, Deep Learning and XGBoost are the models leveraged in this study, which, once applied to public available datasets, provided unexpected results. The datasets used differ in size, quality and feature composition which mirror real-world conditions of clinical practice. Methods such as normalization, encoding, and Synthetic Minority oversampling technique (SMOTE) were implemented directly on the datasets to enhance the model performance when it come to accuracy, recall and precision. The final results showed how XGBoost is the most consistent and reliable model in between datasets. However, Deep learning, while being the second best, it provided unexpected results when working with small datasets. The findings underline how, to improve timing and precision in early detection of cardiovascular diseases, it is important that proper preprocessing and handling of data is performed, together with choosing the right model. This study’s purpose is to show the base of how important the integration of machine learning models with domain experts’ judgement is. Future research is certainly needed, and the thesis suggests practical directions to further improve clinical applicability in cardiac care.