Wavelet Based Machine Learning Techniques for Electrocardiogram Signal Analysis

Heart signals, taken from an Electrocardiogram (ECG) machine, consist of P wave, QRS complex and T wave. These signals contain hidden, but vital information, which enable clinicians to pre-diagnose a disease before any symptoms can be observed. This hidden information is better recognized in the Discrete Wavelet Transform (DWT) domain than in time and frequency domains. The reason for this superiority comes from the fact that the DWT domain provides an accurate time and frequency representation of a signal and this enables both detection and extraction of concealed information. For this project, two different classes of ECG signals were studied: Left Bundle Branch Block (LBBB) and normal heart rhythm. We devised a sophisticated processing structure to extract relevant information from these signals. The first processing step was DWT based noise reduction followed by dimension reduction with Principal Component Analysis (PCA). The PCA output was fed to a Support Vector Machine (SVM), which was used to classify the signals into either normal or diseased. The 10-fold cross validated classification results show 99.93% sensitivity, 100% specificity, 100% Positive Predictive Value (PPV) and 99.96% accuracy. The strong classification results indicate that computerized processing of ECG signals can reveal relevant information for disease diagnosis.