Source-linked AI summary
Compressed Sensing of EEG for Wireless Telemonitoring with Low Energy Consumption and Inexpensive Hardware
Zhilin Zhang, Tzyy-Ping Jung, Scott Makeig, Bhaskar D. Rao
TL;DR
EEG telemonitoring must meet energy, compression, and hardware-cost constraints, but EEG is nonsparse and difficult for conventional CS recovery. The paper applies BSBL to EEG telemonitoring and reports better recovery than state-of-the-art CS algorithms, with quality sufficient for practical use.
Problem
EEG is not sparse in the time or transformed domains, making it difficult for current CS algorithms to recover signals with quality suitable for clinical and engineering applications.
Method
The study applies Block Sparse Bayesian Learning to compress and recover EEG for wireless telemonitoring, evaluating fidelity through subsequent signal processing.
Results
BSBL-BO recovered EEG with better quality than representative CS algorithms, preserving characteristic signal features and supporting subsequent analysis.
Takeaways & Limitations
BSBL is promising for wireless EEG telemonitoring and cognitive neuroscience or engineering applications involving nonsparse physiological signals.
Abstract
from arXiv · showhide
Telemonitoring of electroencephalogram (EEG) through wireless body-area networks is an evolving direction in personalized medicine. Among various constraints in designing such a system, three important constraints are energy consumption, data compression, and device cost. Conventional data compression methodologies, although effective in data compression, consumes significant energy and cannot reduce device cost. Compressed sensing (CS), as an emerging data compression methodology, is promising in catering to these constraints. However, EEG is non-sparse in the time domain and also non-sparse in transformed domains (such as the wavelet domain). Therefore, it is extremely difficult for current CS algorithms to recover EEG with the quality that satisfies the requirements of clinical diagnosis and engineering applications. Recently, Block Sparse Bayesian Learning (BSBL) was proposed as a new method to the CS problem. This study introduces the technique to the telemonitoring of EEG. Experimental results show that its recovery quality is better than state-of-the-art CS algorithms, and sufficient for practical use. These results suggest that BSBL is very promising for telemonitoring of EEG and other non-sparse physiological signals.
I. INTRODUCTION
EEG telemonitoring through WBANs must balance energy consumption, compression, and hardware cost while preserving recovery quality for practical use. The study proposes BSBL to recover EEG despite its lack of sparse structure.
- Telemonitoring context: WBAN-based EEG telemonitoring can support continuous, ubiquitous monitoring without frequent hospital visits.Sensors collect and compress EEG, transmit it through a smartphone and the Internet, and a remote computer recovers the original signal.
- System constraints: Energy consumption is a primary constraint because battery limitations require low-power operation.Lower energy consumption permits smaller batteries and sensors, improves patient comfort, and extends battery and sensor lifetime.
- System constraints: Compression is necessary because ultra-low-power radios have limited capacity and smartphone transit points can be overloaded.The constraint becomes more important with multiple sensors or high sampling rates.
- System constraints: Low hardware cost requires low-complexity sensor compression and recovery that does not require sensor-side preprocessing.Economic viability and individual acceptance depend partly on keeping hardware costs low.
- Motivation and approach: Conventional compression cannot satisfy all constraints simultaneously, while current CS methods struggle with nonsparse physiological signals such as EEG.The study therefore explores BSBL for high-quality EEG compression and recovery, assessing fidelity through subsequent signal processing.
II. COMPRESSED SENSING AND BLOCK SPARSE BAYESIAN LEARNING
Compressed sensing reduces transmitted data using a sensing matrix, but conventional recovery depends on sparsity that EEG often lacks. BSBL addresses this challenge through block-structured Bayesian modeling and modified pruning.
- Compressed sensing: Compressed sensing maps an N-length signal to M measurements with M ≪ N using a full row-rank sensing matrix.Recovery algorithms use the compressed data and known sensing matrix to estimate the original signal.
- Compressed sensing: When the signal is not sparse, CS can use a dictionary D so that x = Dz and z is sparse.The recovery proceeds by estimating z from y and ΦD, then reconstructing x as Dz.
- Compressed sensing: Sparse binary sensing matrices can reduce energy and hardware complexity while retaining competitive compression relative to wavelet compression.Binary-matrix multiplication requires accumulator registers, whereas wavelet compression uses real-number multiplications and coefficient searches.
- Challenge: EEG is not sparse in time or sufficiently sparse in transformed domains, limiting conventional CS performance for telemonitoring.BSBL is presented as a framework for addressing this nonsparse-signal problem.
- BSBL framework: BSBL partitions signals into user-defined non-overlapping blocks and can treat that partition as covariance-estimation regularization rather than requiring true block alignment.Because EEG is nonsparse, the experiments disable BSBL's pruning of blocks with very small norms.
- BSBL framework: The experiments use the bound-optimization BSBL algorithm, BSBL-BO, among the three algorithms in the BSBL framework.The selected algorithm is compared with representative CS recovery methods in the experiments.
III. EXPERIMENTS OF COMPRESSED SENSING OF EEG
The experiments compare CS recovery quality under DCT and wavelet-based models using common sparse binary sensing matrices. NMSE and SSIM provide the primary recovery-quality measures.
- Experimental design: All compared CS algorithms use the same sensing matrix, so their energy consumption is equal and excluded from comparison.The experimental comparison therefore focuses on recovery quality.
- Evaluation measures: NMSE and SSIM measure recovery quality, with higher SSIM indicating better similarity to the original signal.SSIM is described as more suitable than NMSE for structured signals; identical signals have SSIM = 1.
- Experimental design: The first experiment uses an inverse DCT dictionary, while the second uses an inverse Daubechies-20 wavelet dictionary.Both experiments use sparse binary sensing matrices with 15 randomly located ones per column.
- Experiment 1: Table I reports averaged NMSE and SSIM for the three algorithms and for BSBL-BO without a dictionary matrix.The table summarizes performance across the whole dataset.
A. Experiment 1: Compressed Sensing with DCT
Using DCT-based compressed sensing on EEG epochs, BSBL-BO recovered signals with quality that preserved clinically relevant structure and supported subsequent ICA analysis.
- The dataset contained 32 EEG channels, each with 80 epochs of 384 points, and included muscle-movement artifacts.
- A 192 × 384 sparse binary sensing matrix and a 384 × 384 inverse DCT dictionary were used for epoch-by-epoch compression and recovery.
- Only BSBL-BO recovered characteristic EEG peaks, troughs, and oscillatory activity with good quality; DCT-based BSBL-BO performed best overall.
- DCT-based BSBL-BO required 0.105 second per epoch on average, compared with 0.271 second without a dictionary matrix.
- ICA decomposition used identical initialization for original and recovered datasets, and compared scalp maps, ERP images, and averaged ERPs.
- For two representative high-energy independent components, recovered and original scalp maps, ERP images, and averaged ERPs differed little.
B. Experiment 2: Compressed Sensing with WT
In the second experiment, BSBL-BO recovered event-related potentials from left- and right-direction EEG epochs more faithfully than the previous ℓ1 algorithm.
- The dataset contained 250 epochs for each of two events, ‘left direction’ and ‘right direction,’ with 256 sampling points per epoch.
- ERPs were computed by averaging the 250 recovered epochs for each event condition.
- The ℓ1-recovered ERPs were noisy, retained only main peaks, and did not clearly preserve the difference between event ERPs from the 160th to 250th time points.
- BSBL-BO preserved all details of the genuine ERPs with high fidelity, unlike the ℓ1 recovery.
- BSBL-BO achieved SSIM 0.97 and NMSE 0.008, versus SSIM 0.92 and NMSE 0.044 for the ℓ1 algorithm.
- BSBL-BO took 0.06 second per epoch on average on the stated computer.
IV. DISCUSSIONS
The discussion argues that BSBL can recover non-sparse EEG without requiring an optimal dictionary, while acknowledging that wavelet compression can achieve higher quality when energy consumption is unimportant.
- EEG coefficients remained non-sparse under popular dictionaries, making dictionary selection challenging and limiting current CS recovery quality for some applications.
- BSBL achieved sufficient recovery quality with general dictionary matrices, suggesting that seeking an optimal dictionary is not very crucial for EEG compression and recovery.
- When energy consumption was not an issue, wavelet compression produced better EEG recovery quality than CS algorithms including BSBL-BO.
- BSBL-BO nevertheless satisfied the ERP-analysis requirement, so the higher wavelet quality was not needed in that practical setting.
V. CONCLUSIONS
The conclusion presents BSBL as a promising approach for wireless EEG telemonitoring when EEG is non-sparse and subsequent signal analysis requires good recovery quality.
- EEG is non-sparse in both the time domain and transformed domains, making compression with current CS algorithms difficult.
- The study applied block sparse Bayesian learning to alleviate this recovery problem for EEG telemonitoring.
- Experimental results showed good EEG recovery quality that ensured subsequent signal analysis.
- The authors characterize BSBL as promising for wireless telemonitoring-based cognitive neuroscience studies and engineering applications.