| Summary: | Abstract. Early repolarization (ER) pattern was considered a benign finding until 2008, when it was associated with sudden cardiac arrest (SCA). Since then, the interest of the medical community on the topic has grown, stating the need to develop methods to detect the pattern and analyze the risk of SCA. This thesis presents an automatic detection method of ER using supervised classification. The novelty of the method lies in the features used to construct the classification models. The features consist of prototypes that are composed by fragments of the ECG signal where the ER pattern is located. Three different classifier models were included and compared: linear discriminant analysis (LDA), k-nearest neighbor (KNN) algorithm and support vector machine (SVM). The method was tested in a dataset of 5676 subjects, manually labeled by an experienced analyst who followed the medical guidelines.
The algorithm for the detection of ER is composed of different stages. First, the ECG signals are processed to locate characteristic points and remove unwanted noise. Then, the features are extracted from the signals and the classifiers are trained. Finally, the results are fused and the detection of ER is evaluated.
Accuracies of the different classifiers showed results over 90%, demonstrating the discrimitative power of the features between ECG signals with and without the ER pattern. Additionally, dimensionality reduction of the features was implemented with Isomap and generalized regression neural networks (GRNN) without affecting the performance of the method. Moreover, analysis of critical cases that are difficult to label was performed based on the distances to the classifier decision boundary, improving the sensitivity of the detection. Hence, the method presented here could be used to discriminate between ECG signals with and without the ER pattern.
|
|---|
