ECG Classification with Patient-Dependent Normalization and Multi-step Classifier.
Electrocardiogram (ECG) is an important tool for monitoring abnormal heartbeats. Machine learning has been used to facilitate the process of identifying the beats from the ECG data. In this research, we undertook the challenging tasks of building an ECG heartbeat classifier under the inter-patient paradigm. Based on finding significant differences of feature importance in the intra- and inter-patient scenarios, a customised framework design is proposed. The customised framework design includes two novels contributions. These are the patient-dependent normalization to deal with the disparity within the ECG records and the cascading ensemble classifier framework to further increase classification accuracy. MIT-BIH dataset was used as a benchmark for a fair comparison with other algorithms. We followed the Association for the Advancement of Medical Instrumentation (AAMI) for handling and experimenting with the ECG data. Results from our system were comparable to those of similar systems with jk index of 0.728, normal beat’s F1 score of 95.80%, SVEB beat’s F1 score of 58.48%, and VEB beat’s F1 score of 87.57%. Feature analysis for different beat types showed that R-peak’s attributes largely helped with the distinction of these beat types. In the future, it may be wise to experiment with other datasets to demonstrate the generalizability of our system.
Advisor: Deng, Jeremiah
Degree Name: Master of Science
Degree Discipline: Information Science
Publisher: University of Otago
Keywords: Machine learning; ECG; Ensemble; Interpatient
Research Type: Thesis