Multi-Class Classification of ECG Data for Comprehensive Cardiac Abnormality Detection Through Machine Learning

Abstract

Cardiovascular diseases are considered the major cause of death worldwide surpassing cancer. However, despite the broad category of diseases, research has been limited to binary classification i.e. normal and abnormal class leaving behind the accurate classification of specific diseases that affect the ECG waveform. PTB – XL database offers a wide variety of ECG records, but little research is dedicated to extracting morphological features for multi-class classification. Therefore, the paper used the open database to filter the ECG signal records having single unique labels and pre-processed them using the Butterworth bandpass filter and DWT db8. The Bandpass filter corrected baseline wander and reduced noise however, a high signal-to-noise ratio was achieved after applying 8-level DWT. The processed signals were fed into the Pan-Tompkins algorithm to extract R peaks. These peaks served as a baseline to identify other morphological features i.e. P-QRS-T intervals and amplitudes. These extracted features were labelled into 1 normal and 4 abnormal classes. There was a class imbalance in the dataset that could cause bias while training models. Therefore, SMOTE-NC was applied to upsample the dataset. The new dataset was split into the training set and the testing set. These sets were given as inputs to CNN and DNN models for a 5-fold loop. The performance was evaluated for both models using metrics like F1 score, recall, precision and accuracy. The CNN model achieved a mean accuracy of 81% whereas the mean accuracy for DNN was 84%. It was also noted that among the 5 classes, HYP was consistently being classified accurately at 98%.