PaperlayerView exact revision on arXiv

Source-linked AI summary

Compressed Sensing for Energy-Efficient Wireless Telemonitoring of Noninvasive Fetal ECG via Block Sparse Bayesian Learning

Zhilin Zhang, Tzyy-Ping Jung, Scott Makeig, Bhaskar D. Rao

arXiv:1205.1287v7stat.MLcs.LGstat.AP

TL;DR

FECG telemonitoring needs low-energy compression despite raw recordings being non-sparse and strongly contaminated, which limits conventional CS. This work applies BSBL to reconstruct such recordings while preserving multichannel interdependence for ICA, and reports high-quality reconstruction with sparse binary sensing matrices that can use two nonzero entries per column.

  • Problem

    Raw FECG recordings are non-sparse and strongly contaminated, causing current CS algorithms to generally fail in low-energy FECG telemonitoring.

  • Method

    The study uses the BSBL framework to compress and reconstruct raw FECG recordings while exploiting their unknown block and correlation structure.

  • Results

    The framework reconstructs raw recordings with high quality and preserves multichannel interdependence sufficiently for high-fidelity ICA decomposition.

  • Takeaways & Limitations

    BSBL supports wireless FECG telemonitoring with sparse binary sensing matrices having as few as two nonzero entries per column, reducing compression-stage CPU execution.

Abstract

from arXiv · show

Fetal ECG (FECG) telemonitoring is an important branch in telemedicine. The design of a telemonitoring system via a wireless body-area network with low energy consumption for ambulatory use is highly desirable. As an emerging technique, compressed sensing (CS) shows great promise in compressing/reconstructing data with low energy consumption. However, due to some specific characteristics of raw FECG recordings such as non-sparsity and strong noise contamination, current CS algorithms generally fail in this application. This work proposes to use the block sparse Bayesian learning (BSBL) framework to compress/reconstruct non-sparse raw FECG recordings. Experimental results show that the framework can reconstruct the raw recordings with high quality. Especially, the reconstruction does not destroy the interdependence relation among the multichannel recordings. This ensures that the independent component analysis decomposition of the reconstructed recordings has high fidelity. Furthermore, the framework allows the use of a sparse binary sensing matrix with much fewer nonzero entries to compress recordings. Particularly, each column of the matrix can contain only two nonzero entries. This shows the framework, compared to other algorithms such as current CS algorithms and wavelet algorithms, can greatly reduce code execution in CPU in the data compression stage.

I. INTRODUCTION

FECG telemonitoring aims to support ambulatory fetal-health monitoring, but wireless systems face strict energy constraints while raw FECG recordings are non-sparse and heavily contaminated. The study proposes BSBL to reconstruct these recordings and exploit their structure.

  • FECG heart-beat rate, morphology, and dynamic behavior support assessment of fetal development and disease.Abnormal fetal heart-beat-rate patterns can indicate fetal asphyxia.
  • Wireless body-area networks could enable ambulatory FECG monitoring at home and reduce hospital visits, time, and medical expenses.Sensors communicate through low-power radios and nearby handheld devices with remote healthcare terminals.
  • Energy consumption is a primary WBAN design constraint, requiring minimal on-sensor computation and data compression before transmission.Wavelet-based compression can dissipate substantial energy.
  • Compressed sensing uses a simple linear sensing matrix to compress data before transmission and reconstruct it at remote terminals by exploiting sparsity.Sparse binary sensing matrices can extend sensor lifetime and reduce energy consumption.
  • FECG telemonitoring challenges CS because strict energy limits conflict with the non-sparsity and strong contamination of raw recordings.Most CS algorithms struggle to reconstruct such signals directly.
  • The study applies BSBL because it can reconstruct non-sparse signals and exploit unknown spatial, temporal, and dynamic structure.These properties are presented as the basis for successful wireless FECG telemonitoring.

II. COMPRESSED SENSING AND BLOCK SPARSE BAYESIAN LEARNING

The paper formulates compressed sensing as linear measurement and reconstruction of FECG segments, then models raw recordings as noisy, structured signals with unknown block organization. Transform-domain sparsity is inadequate when small coefficients matter for downstream ICA.

  • In the noisy CS model, y is formed from the sensing matrix Φ, signal x, and compression noise v; in this application v can be ignored.The signal x is a raw FECG segment and y is transmitted through a WBAN.
  • CS reconstructs x from compressed data y at a remote terminal using the designed sensing matrix Φ.The inverse problem is underdetermined and is typically addressed by exploiting signal sparsity.
  • When x is sparse only after transformation, CS reconstructs sparse coefficients θ using Ω=ΦΨ and then recovers x through x=Ψθ.This approach depends on sufficient sparsity in the transformed representation.
  • Raw FECG recordings are not sparse in the time domain, with every entry in the illustrated segment nonzero.The segment contains visible maternal and fetal QRS complexes, including overlapping contamination.
  • Raw FECG recordings can be viewed as block-sparse signals contaminated by signal noise, with the block partition unknown during practical telemonitoring.Illustrative segments contain significant nonzero regions despite the overall recording being far from sparse.
  • Existing CS methods struggle with unknown block partitions, whereas BSBL was developed to reconstruct non-sparse but structured signals.The paper presents BSBL as having superior performance over existing algorithms for this signal class.

B. Block Sparse Bayesian Learning

BSBL represents signals as correlated Gaussian blocks and estimates their block activity and within-block correlation. BSBL-BO preserves non-sparse structure through a small pruning threshold and does not require the user-defined partition to match the true partition.

  • BSBL models each block x_i with a multivariate Gaussian prior N(0, γ_iB_i).The parameter γ_i controls block sparsity, while B_i captures within-block correlation.
  • A zero γ_i makes the corresponding block zero, while the block-diagonal prior covariance combines the separate block covariances.The model assumes blocks are mutually uncorrelated.
  • BSBL estimates γ_i, B_i, and λ through Type-II maximum likelihood before obtaining the signal estimate by maximum-a-posteriori estimation.The parameter estimates determine the posterior reconstruction.
  • BSBL-BO reconstructs non-sparse signals by setting the γ_i-pruning threshold small, thereby pruning fewer blocks during iterations.A smaller threshold allows fewer γ_i values to be removed and fewer blocks to become zero.
  • BSBL-BO estimates matrices B_i to exploit intra-block correlation and improve reconstruction performance.This correlation modeling is presented as a distinct ability of BSBL-derived algorithms.
  • The user-defined block partition need not match the true block partition.The partition functions as a regularization for covariance estimation rather than a required exact description of the signal.

III. EXPERIMENTS

Experiments used two diverse real-world raw FECG datasets and compared representative CS algorithms under a common sensing matrix. Because the sensing matrix was shared, the paper reports reconstruction results rather than energy-consumption differences.

  • Two real-world FECG datasets were selected, including one where FECG was barely visible and another where it was invisible.The datasets were intended to test the algorithm under diverse recording conditions.
  • Ten representative CS algorithms were compared, with each representing a family of algorithms and having top-tier performance within that family.The authors state that this supports generalization of comparison conclusions to related algorithms.
  • All CS algorithms used the same sensing matrix to compress the FECG recordings.This design equalized the compression setup across algorithms.
  • Energy consumption was the same across the compared CS algorithms under the shared sensing matrix, so only reconstruction results were presented.The paper therefore does not directly analyze energy consumption in these experiments.
  • Raw-recording reconstruction was evaluated in the context of downstream FECG extraction rather than solely by mean square error.The reconstructed recordings were further processed using techniques such as BSS/ICA and nonlinear filtering.

A. The DaISy Dataset

On the DaISy dataset, BSBL-BO reconstructed non-sparse FECG recordings more faithfully when exploiting intra-block correlation, preserving downstream FECG extraction quality.

  • The DaISy segment is far from sparse, with every entry non-zero, making reconstruction difficult for existing compressed-sensing algorithms.
  • Exploiting intra-block correlation enabled BSBL-BO to reconstruct the segment with high quality, whereas disabling it caused poor recovery and omitted an FECG QRS complex.
  • Only BSBL-BO reconstructed the segment satisfactorily among the compared algorithms; the other recovery results were not meaningful.
  • BSBL-BO reconstructed the whole DaISy dataset without visually observable distortions, including recordings affected by weak FECG signals and baseline wander.
  • FECG extracted from the reconstructed dataset retained the QRS complexes and residual-noise-free appearance of extraction from the original recordings, with Pearson correlation 0.931.

B. The OSET Dataset

On the more difficult OSET dataset, BSBL-BO recovered recordings containing strong baseline wander and very weak, noise-buried FECG, while preserving downstream ICA structure with little distortion.

  • The OSET dataset contains eight abdominal recordings with strong baseline wander and FECG that is weak or buried by noise and maternal ECG.
  • BSBL-BO recovered the dataset visually like the original, including the baseline wanders.
  • The ten previously evaluated CS algorithms again failed to reconstruct this dataset, while only CluSS-MCMC and BM-MAP-OMP results were shown.
  • After band-pass filtering and FastICA, the reconstructed recordings yielded independent components with very small distortion relative to components from the original dataset.

C. Reconstruction in the Wavelet Domain

Wavelet-domain reconstruction did not enable existing CS algorithms to recover raw FECG faithfully. Recovering large wavelet coefficients while missing many small ones disrupted ICA-based extraction.

  • C. Reconstruction in the Wavelet Domain: The wavelet approach was evaluated by reconstructing θ and then computing x = Ψθ using a Daubechies-4 wavelet basis.The sensing matrix was the same one used in the preceding experiment.
  • C. Reconstruction in the Wavelet Domain: All ten tested CS algorithms failed to extract FECG after wavelet-domain reconstruction.The SL0 reconstruction shown in Figure 9 also failed to produce an extracted FECG.
  • C. Reconstruction in the Wavelet Domain: Maintaining the ICA mixing structure requires accurate recovery of both large- and small-amplitude wavelet coefficients.The paper identifies this preservation as necessary for high-fidelity FECG extraction from reconstructed multichannel recordings.
  • C. Reconstruction in the Wavelet Domain: Wavelet reconstruction recovered large-amplitude coefficients but missed small-amplitude coefficients, causing ICA to fail.Raw FECG recordings contain many small-amplitude wavelet coefficients that are difficult for the CS algorithms to recover.

IV. PERFORMANCE ISSUES OF BSBL-BO

This section tests BSBL-BO under varying FECG, MECG, and noise strengths using Monte Carlo simulations. The reported results show high recovery quality even under the worst SINR conditions.

  • IV. PERFORMANCE ISSUES OF BSBL-BO: The experiments address whether BSBL-BO remains effective when MECG and noise strengths differ across datasets.Such variation can result from fetus position, pregnancy week, and random muscle movement.
  • IV. PERFORMANCE ISSUES OF BSBL-BO: The simulations varied the strengths of multichannel FECG, MECG, and noise in synthesized recordings.The recordings were modeled as the sum of multichannel FECG, MECG, and noise.
  • IV. PERFORMANCE ISSUES OF BSBL-BO: BSBL-BO maintained high recovery quality even in the worst signal-to-interference-and-noise scenarios.Figure 10 reports Pearson correlation averaged over 20 trials at different SINRs.
  • IV. PERFORMANCE ISSUES OF BSBL-BO: At SINR=-35dB, the extracted FECG from the recovered dataset still maintained high fidelity despite very strong noise.Figure 11 compares extracted FECGs from the synthesized and recovered datasets.

B. Effects of the Block Partition

BSBL-BO’s extraction quality remained nearly unchanged across a broad range of block sizes and compression settings. Smaller sensing matrices accelerated reconstruction, although they increased recording MSE at fixed compression ratio.

  • B. Effects of the Block Partition: Extraction quality was almost the same across block sizes h ranging from 4 to 90.The experiment used a 128 × 256 sparse binary sensing matrix with 12 ones per column and 20 repetitions.
  • C. Effect of Compression Ratio: When CR ≤60, extracted FECGs were satisfactory for clinical diagnosis, and CR=60 preserved identical detected R-wave peaks.The comparison used reconstructed and original datasets, with peak detection performed using the OSET toolbox.
  • C. Effect of Compression Ratio: A smaller sensing matrix produced slightly better extracted-FECG quality than N = 512 in the reported experiment.The smaller-matrix experiment used N = 256, with 12 ones in each column.
  • C. Effect of Compression Ratio: A smaller sensing matrix reduced reconstruction time, taking less than 0.6 seconds versus less than 1.4 seconds for a 512-point segment.The segment required 2.048 seconds to collect, and the smaller matrix enabled near-real-time reconstruction.
  • C. Effect of Compression Ratio: At fixed CR, smaller sensing matrices generally increased reconstruction MSE without correspondingly worsening extracted-FECG quality.This separates recording-level MSE from the downstream extraction measure.

D. Study on the Number of Nonzero Entries in Each Column of the Sensing Matrix

The study varied the number of nonzero entries per sensing-matrix column to assess compression sparsity. BSBL-BO’s reconstruction and extraction results were not affected by this parameter across the tested range.

  • D. Study on the Number of Nonzero Entries in Each Column of the Sensing Matrix: Both extracted-FECG correlation and reconstructed-dataset MSE were unaffected by d, the number of ones per sensing-matrix column.The experiment varied d from 2 to 14 over 20 repetitions per value.
  • D. Study on the Number of Nonzero Entries in Each Column of the Sensing Matrix: BSBL-BO was more robust to d than the basic ℓ1 CS algorithm reported in.The earlier algorithm’s performance was very sensitive to the number of nonzero entries per column.
  • D. Study on the Number of Nonzero Entries in Each Column of the Sensing Matrix: This robustness supports energy saving because the sensing matrix can use fewer nonzero entries per column without changing reported recovery measures.The tested sparse binary matrix had size 256×512 and remained full row-rank.

V. DISCUSSIONS

BSBL-BO directly reconstructs non-sparse raw FECG recordings by exploiting block structure and intra-block correlation, avoiding thresholding and transformed-domain reconstruction strategies.

  • Thresholding is unsuitable for FECG because weak amplitudes make an optimal threshold difficult to choose and can destroy multichannel ICA structure.
  • Transformed-domain reconstruction depends on sparse coefficients, but many small coefficients remain important for downstream processing such as ICA-based FECG extraction.
  • BSBL-BO directly reconstructs non-sparse signals without thresholding or transformed-domain reconstruction, enabling further signal processing or pattern recognition.
  • The method’s reconstruction relies on exploiting block structure and intra-block correlation in the signal.

C. Energy-Saving by the BSBL Framework

BSBL-BO can retain reconstruction performance with a much sparser sensing matrix, reducing compression computation and supporting broader implications for non-sparse signal recovery.

  • The work focuses on algorithms and does not itself analyze energy consumption or compare BSBL-BO with wavelet compression.
  • BSBL-BO can use d = 2 nonzero entries per sensing-matrix column while maintaining the same reconstruction performance and reducing energy consumption.
  • 768 additions are needed for 512-to-256 compression with d = 2, compared with 5888 additions for d = 12.
  • 11784 multiplications and 11272 additions are required by Daubechies-4 wavelet compression for the stated signal-compression example.
  • Experiments indicate that only BSBL-derived algorithms can use d = 2 without reconstruction failure or degraded quality, whereas other CS algorithms are sensitive to this setting.
  • Exploiting block structure and intra-block correlation can yield a solution close to a non-sparse true signal in an underdetermined problem.
  • The framework may also apply to other telemedicine signals, including adult ECG, wireless EEG, and electromyography.
Loading 1205.1287v7…