Article Preview
TopIntroduction
There is a trend towards increasing monitoring of physiological parameters such as the electrocardiogram, ECG, in home care settings. During the last decade, textile electrodes have been developed and tested for ECG recordings (Axisa et al., 2005; Catrysse et al., 2004; Rattfalt et al., 2007; Scilingo et al., 2005; Westbroek et al., 2006). One large benefit of textile electrodes is that they can be included in a garment. A disadvantage is that they are sensitive to disturbances due to motion. A way of reducing those is by signal processing and effective such methods are highly needed.
Common disturbances in the ECG are baseline drift, electrode movement disturbances (EMD), power line interference and myoelectric activity. Baseline drift can depend on changes in electrode potentials. By using linear filtering techniques, some of the disturbances can be reduced but to some extent they have the same frequency content as the ECG, which demands other techniques to separate them from the ECG (Sörnmo & Laguna, 2005).
EMD are caused by a change in the electrode potential and deformation potentials of the skin (Ödman, 1982). When using regular skin electrodes, EMD are reduced by having an adhesive collar and a gel filled cup to ensure as stable skin–electrode interface as possible. With textile electrodes, an adhesion is normally not possible. Instead the skin contact has to be secured by a tight fit but this makes the electrodes more sensitive motion disturbances. One way to overcome this is to obtain redundancy in signals by introducing several electrodes, assuming that some of them contain non-corrupted data at each point in time. However, this introduces yet another step; identifying which channel that contain useful information and in which channels the noise level is high.
Previously, multivariate methods have been proposed by applying independent component analysis (Ragnarsson, Östlund, & Wiklund, 2005) or blind source separation with canonical correlation to analyze the multi-channel ECG (Rattfalt et al., 2006). However, the output from these two methods consists of statistically derived components from the original signal set and it is not known which component that contains the heartbeat information.
The aim of this study was to develop and evaluate a robust heartbeat detector. We suggest a method based on weighted correlation in a multi-channel ECG to construct a heartbeat detector based on knowledge of the signal at hand. It uses an extracted segment from the ECG of a QRS complex as a filter kernel together with a weight function based on the weighted mean of consecutive heartbeats’ variances. The algorithm’s performance was compared by the well-established Pan Tompkins heartbeat detector.