Abstract
Electrocardiograms (ECGs) play a crucial role in diagnosing heart-related conditions, and ensuring the reliability of ECG collections is essential for accurate diagnoses. While previous research has focused on processing ECG signals, many machine learning models developed earlier had limited datasets, making them less suitable for real-world applications. In clinical settings, the inadvertent assignment of ECG recordings to incorrect patients poses a significant challenge. In our work, we address this issue through a comprehensive approach that spans multiple phases, starting with meticulous preprocessing of the dataset. We focus on the highly imbalanced PTB-XL electrocardiography dataset that contains records of 18885 patients. Prior to the preprocessing phase, we perform channel selection to choose more meaningful features for accurate prediction. For feature selection, a crucial step in enhancing classification accuracy, we employ a novel fusion of the Mel Frequency Cepstrum Coefficient (MFCC) and statistical features from the Discrete wavelet transform of input. We achieve state-of-the-art performance in binary classification, i.e., 97.6% using Artificial Neural Networks (ANN), effectively distinguishing between healthy individuals and those with health conditions.
