Sudden cardiac death Prediction Using Time-Frequency Analysis of Electrocardiogram Signal

Abstract:

Sudden cardiac death (SCD) is a prevalent cardiovascular disease that affects approximately 3 million individuals globally each year, often without displaying any noticeable symptoms prior to the event. While the underlying causes of SCD remain unclear, ventricular fibrillation is believed to play a role in its pathophysiology. Given that symptoms typically appear only an hour before death, opportune prediction is crucial for an effective cardiac resuscitation. In recent years, there has been an extensive number of researches focused on developing methods of diagnosis and prediction of SCD using electrocardiogram (ECG) signals and heart rate variability. The goal of this thesis is to detect SCD using processing methods of ECG signal through time-frequency conversion technique. In this study two datasets namely the sudden cardiac death Holter dataset and the MIT-BIH normal sinus rhythm dataset were employed, which are available in Physionet database. In this study two approaches for predicting of SCD have been proposed. In the first approach, the 25-minute interval prior to ventricular fibrillation has been segregated into one-minute segments. Subsequently, each one-minute segment has been decomposed into time-frequency sub-bands employing experimental wavelet analysis. Nonlinear features are then extracted from the decomposed signal. Additionally, this approach has been applied onto 60 minutes prior to SCD. In the second approach, initially the 60-minute interval prior to ventricular fibrillation has been divided into one-minute segments. Each one-minute segment is subsequently decomposed into time-frequency sub-bands using empirical mode decomposition (EMD). Then nonlinear features are extracted from the decomposed signal. Ultimately in both approaches, support vector machines and k-nearest neighbors has been implemented for classification between healthy individuals and those at risk of SCD. Our findings in the first approach indicate that in both cases of 25 and 60 minutes before SCD, the total average accuracy of 94.20% and 93.29% has been yielded respectively, while employing support vector machine for classification. The highest results of the second approach have been achieved in the case of 60 minutes with an average total accuracy of 94.03% when using k-nearest neighbor for classification.