Exploiting prior knowledge in compressed sensing wireless ECG systems
Date
2015
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Wireless body area networks promise to revolutionize health monitoring by allowing
the transition from centralized health care services to ubiquitous and pervasive
health monitoring in every-day life. One of the major challenges in the design of such
systems is the energy consumption as wireless body area networks are battery-powered.
Recent results in telecardiology show that compressed sensing (CS) is a promising tool
to lower energy consumption in wireless body area networks for electrocardiogram
(ECG) monitoring. However, the performance of current CS-based algorithms, in
terms of compression rate and reconstruction quality of the ECG, still falls behind
the performance attained by state-of-the-art wavelet-based algorithms. This is mainly
because current CS-based algorithms exploit only the sparsity of the signal, ignoring
important signal structure information that can be known a priori and lead to
enhanced reconstruction results.
This dissertation presents methods to exploit prior knowledge of the ECG in
order to improve the reconstruction quality and to increase the compression rates
offered by current CS-based algorithms. First, we describe an algorithm that exploits
prior information about the wavelet dependencies across scales and the high fraction
of common support of the wavelet coefficients of consecutive ECG segments.
One of the main challenges in the reconstruction of ECG signals via CS-based
algorithms is the recovery of the small-magnitude wavelet coefficients. This dissertation
also presents a weighted ℓ1 minimization algorithm, based on a maximum a posteriori
(MAP) approach, that exploits the exponentially decaying magnitude of the detail
coefficients across scales and the accumulation of signal energy in the approximation
subband.
In real scenarios, ECG recordings are often corrupted by artifacts. This dissertation
also presents a robust reconstruction method for ECG signals in the presence of
electromyographic noise. To achieve this objective, robust statistics are used to develop
appropriate methods addressing the problem of electromyographic noise, which can be
modeled as impulsive noise.
Most prior work in CS ECG has employed analytical sparsifying transforms
such as wavelets. Another contribution of this dissertation is to adaptively learn a
sparsifying transform (overcomplete dictionary) that exploits the multi-scale sparse
representation of ECG signals. By calculating subdictionaries at different data scales,
we are able to exploit the correlation within each wavelet subband and, subsequently,
represent the data in a more efficient manner.
Generic sparsity models that are not tied to a specified structure are also explored
in this dissertation. More precisely, restricted Boltzmann machines and deep
belief networks are employed to model the sparsity pattern of ECG signals with the
goal of exploiting higher-order statistical dependencies between sparse coefficients.
The effectiveness of the proposed algorithms is demonstrated on real ECG signals
from the MIT-BIH Arrhythmia Database. Results show that the proposed algorithms
require fewer measurements and offer superior reconstruction accuracy than
existing CS-based methods for ECG compression.