Speaker
Aron Wei-Hsiang Su
Title
An Online Nonlinear Bayesian Filtering Framework for Electrocardiogram Noise Removal
Abstract
The electrocardiogram (ECG) is a time-varying electrical signal that interprets the electrical activity of the heart. It is obtained by a non-invasive technique called surface electromyography (EMG), which is widely used in hospitals. The signals captured are usually corrupted with unwanted interference such as muscle artifacts, electrode artifacts, power line interference and respiration interference. Methods of noise reduction have decisive influence on the performance of all ECG signal-processing systems.
Many of the existing signal pre-processing techniques are able to remove most of the noise in the signal; however, one of the main problems with the existing signal processing techniques is the lack of reliable signal processing tools to extract the weak ECG components contaminated with background noise and permit the measurements of subtle features in the ECG. For example, the noise from EMG usually overlaps with the fundamental ECG cardiac components in the frequency domain in the range from 0.01 HZ to 100 HZ. It is impossible to reduce these interferences effectively without signal distortion using a linear filtering approach.
To solve this problem, a Bayesian filtering framework has been proposed, which is built upon an existing Bayesian filter framework. The previous framework tackled this problem with superior results compared with conventional ECG pre-processing approaches. One problem is that the framework has to be run offline. It is desirable for the framework to be processed online for most real world applications. The proposed framework developed in this thesis uses the theory from fixed-lag smoothing to modify an existing ECG dynamical model. The modified ECG dynamical model allows the ECG signal to be processed online, which could be used with the conventional extend Kalman filter. It was found the proposed framework had little loss in performance in terms of signal-to-noise ratio in comparison to the previous Bayesian filtering framework.