Abstract

Ventricular tachycardia (VT) is a type of abnormally fast heart rate that arises from abnormal electrical conductivity in the ventricles of the heart. Most VTs involve an abnormal origin of electrical activation inside the ventricles. An effective way to treat VT is catheter ablation that destroys the origin of VT by radiofrequency energy. To accurately localize the origin of VT therefore is an important factor for the success of ablation therapy. An Electrocardiogram (ECG) is a recording of the electrical activity of the heart with features that correspond to stages in the cardiac conduction system. Earlier works have shown that predicting the origin of VT using these features is possible using machine learning techniques such as support vector machines. However there are variations among each patient such as heart geometry and scar characteristics which are not accounted for by these methods. This thesis aims to explore the use of multi-task learning (MTL) to treat the predictive modeling for different patients as separate but related tasks, where we can model the similarities and differences across patients. While traditional MTL approach enforces all tasks to share something in common, we hypothesize that clustering the patients into subgroups during multi-task learning may improve the performance by considering the heterogeneity of the patient group. Unexpectedly, results obtained on 39 patients suggested that sharing information across patient-specific models -- whether or not to consider automatic sub-grouping of the patients -- had little effect on the accuracy of the models. We conclude the thesis by speculating the potential reasons and future explorations for this unexpected result.

Publication Date

7-26-2019

Document Type

Thesis

Student Type

Graduate

Degree Name

Computer Science (MS)

Department, Program, or Center

Computer Science (GCCIS)

Advisor

Linwei Wang

Advisor/Committee Member

Richard Zanibbi

Advisor/Committee Member

Joe Geigel

Campus

RIT – Main Campus

Share

COinS