Feature Selection and Non-Euclidean Dimensionality Reduction: Application to Electrocardiology

Heart disease has been the leading cause of human death for decades. To improve treatment of heart disease, algorithms to perform reliable computer diagnosis using electrocardiogram (ECG) data has become an area of active research. This thesis utilizes well-established methods from cluster analysis, classification, and localization to cluster and classify ECG data, and aims to help clinicians diagnose and treat heart diseases. The power of these methods is enhanced by state-of-the-art feature selection and dimensionality reduction. The specific contributions of this thesis are as follows. First, a unique combination of ECG feature selection and mixture model clustering is introduced to classify the sites of origin of ventricular tachycardias. Second, we apply a restricted Boltzmann machine (RBM) to learn sparse representations of ECG signals and to build an enriched classifier from patient data. Third, a novel manifold learning algorithm is introduced, called Quaternion Laplacian Information Maps (QLIM), and is applied to visualize high-dimensional ECG signals. These methods are applied to design of an automated supervised classification algorithm to help a physician identify the origin of ventricular arrhythmias (VA) directed from a patient's ECG data. The algorithm is trained on a large database of ECGs and catheter positions collected during the electrophysiology (EP) pace-mapping procedures. The proposed algorithm is demonstrated to have a correct classification rate of over 80% for the difficult task of classifying VAs having epicardial or endocardial origins.