Advanced Heartbeat Classification Models for Reliable Electrocardiogram Analysis in Ambulatory Health Monitoring.

摘要

The past decade has witnessed a significant boost of interest in wearable health monitoring systems. The advance in bio-sensing and wearable computing technologies has enabled the continuous monitoring of vital signals, e.g., electrocardiogram (ECG) and blood pressure, in ambulatory environments, for better health management of a variety of at-risk populations, including elderly, people suffering from chronic diseases, etc., during their daily life.;In this thesis, we focus on developing advanced signal processing and machine learning algorithms for reliable and robust automatic analysis of ECG signals in ambulatory health monitoring scenarios, for the timely detection of various abnormal cardiac conditions. The major challenge for automatic ECG analysis comes from the significant variations in ECG signals, which can be divided into two categories, i.e., inter-person variations and intra-person variations. Inter-person variations refer to significant variations in morphologies of ECG signals among different subjects, whilst intra-person variations occur when a person experiences changes in heart conditions or physical state.;We investigate the construction of subject-customized classification models, in order to address the inter-person variations. The proposed subject-customized models consist of two categories of models, namely, general models and specific models, focusing on representing the general population knowledge and the specific knowledge of a particular subject, respectively. Besides, we develop a multi-view based semi-supervised learning approach to fully automate the construction of individual-specific models. Towards handling the intra-person variations, we propose rhythm context-aware models that model and incorporate valuable underlying rhythm context information, based on a time series analysis approach. In addition, we explore a probabilistic graphical model-based framework to provide a representation of important knowledge, and merge information from different models to make the final decision. We also investigate the potential of an ECG-based biometric solution for patient identification, which can be used as an automatic login solution for related health monitoring devices, offering security and convenience. Finally, the proposed algorithm framework for automatic ECG signal analysis exhibits a significant improvement over the state-of-the-art methods, based on the evaluation of a benchmark database, providing a promising solution for enhanced personalized ECG analysis.