Alzheimer's disease (AD) is a complex and widespread form of dementia whose prevalence in the world is expected to double in the next 20 years.1 There is no known cure for the disease, but a number of medications have been developed to delay its symptoms and causes. Much Alzheimer's research focuses on selecting biomarkers that can detect Alzheimer's early, predict future clinical decline, and measure the impact of drug treatments. Biomarkers measured using various techniques can reveal changes in the body over time due to the disease (e.g., structural imaging to measure brain atrophy, functional imaging to quantify hypometabolism, and collecting cerebrospinal fluid to detect protein levels). However, most of these clinical examinations are costly, invasive, not clinically valuable at the onset of Alzheimer's, and cannot be used remotely to continuously monitor a patient living at home over time.
The use of electroencephalography (EEG) has been proposed as an approach for mass screening the population at risk for Alzheimer's2, 3 because EEG is noninvasive, repeatable, and can be easily executed at home through wireless body area networks as a personalized medical tool.4 EEG is also a direct correlate of brain function that is used clinically to monitor brain activity. Wireless body area networks use sensors to collect and compress EEG signal data transmitted to a smartphone via Bluetooth, and then sent to a remote telehealth provider via the Internet where the original EEG is recovered. Because wireless networks have limited bandwidth, the compression ratio of the EEG signals must be high to remove as much redundancy as possible and limit the spatial-temporal throughput rate bottleneck.
Compressive sensing (CS)5,6 is an emerging compression technique that takes advantage of a signal's sparsity to sample and compress the signal at the same time. The technique exploits signal structure to acquire data at a rate proportional to the true information rate, rather than to the Nyquist rate. As a result, the effective sampling rate is lowered. While many physiological signals, such as EEG, are not sparse in nature, neither in the time domain nor in the transformed domain, the EEG signals of Alzheimer's patients (AD-EEG) exhibits two characteristics that make them good candidates for signal compression: they are typically slower (i.e., contain more low-frequency power) and less complex (i.e., more predictable) than those of healthy subjects.2, 3
To investigate the correlation between the compressibility and complexity of AD-EEG, we analyzed EEG signals from healthy control (HC) subjects, pre-dementia patients (i.e., individuals with mild cognitive impairment, MCI), and patients with AD. Individuals with MCI are not compromised in their daily living, but they have a subclinical and isolated cognitive deficit, and are potentially at risk of developing AD. Our results showed that MCI patients exhibited an EEG signal that fell between the signals collected from healthy individuals and subjects with AD (see Figure 1). The reconstruction of the original signal from a reduced (sub-Nyquist) set of samples is more accurate for the AD-EEG signal, which is slower and less complex. Although Figure 1 was reconstructed using a wavelet coding approach that exploits sparsity, the CS technique can be used to determine a small collection of linear projections of the approximately sparse EEG signals containing enough information for accurate reconstruction (e.g., AD-EEG).
Figure 1. Original and reconstructed EEG recordings for three groups: healthy controls (HC), individuals with mild cognitive impairment (MCI), and subjects with Alzheimer's disease (AD).
Next, we compared the reconstruction root mean square (RMS) error with the number of cancelled components in the reconstructed representation of the EEG signal for the three subject groups (see Figure 2). The RMS error increased as the number of zeroed coefficients increased, but the HC RMS error was always highest for a given number of components. This enhanced compressibility property indicates that CS can be used for improved quality of recovery in Alzheimer's patients and as an additional marker for AD-EEG discrimination.
Figure 2. Root mean square (RMS) reconstruction error as a function of zeroed reconstructed signal components shows EEG signal compressibility for three groups: healthy controls (HC), individuals with mild cognitive impairment (MCI), and subjects with Alzheimer's disease (AD).
In recent work, we proposed applying a block sparse Bayesian learning algorithm to help with the recovery of the EEG signals at the receiver.7 This method exploits the ability of machine learning approaches to induce sparseness at the block level. Using this method, we found that AD-EEG became more compressible with the progression of the disease. This method could be improved by incorporating intra-channel and inter-channel correlations of the EEG signals.
We also introduced the concept of permutation entropy to measure the complexity of EEG signals and showed that AD-EEG is less complex on average than the EEG of HCs (EEG-HC).2 Our recent work revealed permutation entropy values ranging from 0.65–0.70 for HCs, 0.63–0.66 for those with MCI, and 0.61–0.64 for individuals with AD.7 This indicates that the EEG-AD signal recordings are typically slower and less complex than EEG-HC recordings. Our results provide evidence of an inverse correlation between the compressibility measure and the complexity measure. The evolution of these markers in time may yield supporting evidence for the progression of MCI to AD.
Compressed sensing can be used to distinguish Alzheimer's from other forms of dementia, and to differentiate individuals with AD from those with MCI and HCs. In addition, the compression factor can serve as an additional marker of Alzheimer's evolution that can be inexpensively collected over time from people at risk. Compressed sensing can also be used to wirelessly monitor the progress of patients living at home with AD and the effect of treatments in mild to moderate Alzheimer's patients.
Future work will include applying our technique to a larger database of subjects; exploiting inter-channel correlations for making distributed compressed sensing of jointly sparse signals; and using this approach on different biomarker modalities, possibly within the Alzheimer's Disease Neuroimaging Initiative.
Francesco Carlo Morabito
University Mediterranea of Reggio Calabria
Reggio Calabria, Italy
Francesco C. Morabito is a professor of electrical engineering. Previously, he taught at the École Polytechnique Fédérale de Lausanne in Switzerland, and in Naples, Messina, and Cosenza. He is a foreign member of the Royal Academy of Doctors in Spain, was a former governor of the International Neural Network Society (INNS) for 12 years, and is a senior member of IEEE and INNS.
1. M. W. Weiner, D. P. Veitch, P. S. Aisen, L. A. Beckett, N. J. Cairns, R. C. Green, D. Harvey, The Alzheimer's Disease Neuroimaging Initiative: A review of papers published since its inception, Alzheimer's & Dementia 8(1 Suppl), p. S1-S68, 2012.
2. F. C. Morabito, D. Labate, F. La Foresta, A. Bramanti, G. Morabito, I. Palamara, Multivariate multi-scale permutation entropy for complexity analysis of Alzheimer's disease EEG, Entropy 14, p. 1186-1202, 2012.
3. J. Dauwels, K. Srinivasan, M. R. Reddy, T. Musha, F.-B. Vialatte, C. Latchoumane, J. Jeong, A. Cichocki, Slowing and loss of complexity in Alzheimer's EEG: two sides of the same coin?, Int'l. J. Alzheimer's Disease, p. 539621, 2011.
4. H. Szu, C. Hsu, J. Jenkins, J. Willey, J. Landa, Capturing significant events with neural networks, Neural Networks 36(29-30), p. 1-7, 2012.
5. D. L. Donoho, Compressed sensing, IEEE Trans. Inf. Theory 52(4), p. 1289-1306, 2006.
6. R. Baraniuk, Compressive sensing, IEEE Signal Process. Magazine 24(4), p. 118-121, 2007.
7. F. C. Morabito, G. Morabito, H. Szu, Monitoring and diagnosis of Alzheimer disease using noninvasive compressive sensing EEG, Proc. SPIE
8750, p.87500Y, 2013. doi:10.1117/12.2020886