Developing an Effective Machine Learning System for the Diagnosis of Cardiovascular Diseases using Electrocardiogram Signals

Abstract

Cardiovascular diseases, identified by the World Health Organization as the predominant cause of global mortality, present a formidable health challenge. This issue is further exacerbated in middle- and low-income countries due to the lack of resources for effective disease management and prevention. The primary diagnostic tool for these conditions is the electrocardiogram (ECG), which provides a graphical representation of the heart's electrical activity over time. Given the high mortality rates and severe implications of cardiovascular diseases, early detection becomes critical. In this context, machine learning techniques are emerging as key contributors to disease prediction and management. Most of the available studies uses the MIT-BIH database (only 48 patients) and/or focus mainly at the detection of arrhythmias. The proposed system employs machine learning techniques on PTB-XL, the to-date largest freely accessible dataset (18885 patients) to augment the analysis of cardiac events using ECG signal data. Initially, the system processes the ECG signals to extract unique attributes, each indicative of different signal characteristics generated by the 12 leads. These attributes are then used to identify valuable features for training various machine learning models. These models are designed to test for 57 diagnostic values of the ECG, encompassing a broad range including heart blocks, arrhythmias, fibrillations, and heart enlargements. Upon completion of training, each model is evaluated based on statistical performance parameters and data interpretative capacity. The model that demonstrates the most effective balance between statistical accuracy and interpretative prowess is selected as the final model for heart event prediction. At present, the model can accurately distinguish a normal class from any abnormal condition with an accuracy of 73.19%. However, accuracy fluctuates between different conditions, with a high of 99.8% for First Degree heart block, and a low of 69.3% for Left ventricular hypertrophy-thickened walls. As this system undergoes further refinement, it promises to enhance its diagnostic capabilities. By augmenting the capacity of healthcare practitioners to predict abnormal heart activity, this system could contribute significantly to distinguishing between healthy individuals and those with cardiovascular diseases, thereby facilitating early intervention and efficient disease management. The integration of technology and healthcare, forecasts a promise in the transformation in the landscape of disease detection, prevention, and management.