Catálogo de publicaciones - libros

The routine electroencephalogram (rEEG) is a useful diagnostic test for neurologists. But this test is frequently misinterpreted by neurologists due to a lack of systematic understanding of paroxysmal electroencephalographic discharges (PEDs), one of the most important features of EEG. A heuristic algorithm is described which uses conventional blind signal source separation (BSSS) algorithms to detect PEDs in a routine EEG recording. This algorithm treats BSSS as a ‘black box’ and applies it in a computationally-intensive multitaper algorithm in order to detect PEDs without a pre-specification of signal morphology or scalp distribution. The algorithm also attempts to overcome some of the limitations of conventional BSSS as applied to the study of neurophysiology datasets, specifically the ‘over-completeness problem’ and the ‘non-stationarity problem’.

A temporally-constrained blind-source-separation algorithm was used to analyse auditory evoked potentials, evoked from impulse trains with inter-stimulus rates of 95 and 198 Hz. A nonstationarity of variance contrast function was used, and a simulation run showing its ability to extract sources based on a simple convolved model of auditory brainstem and middle latency responses. For a stimulus rate of 95 Hz, where no neural adaptation occurs, this approach was partially successful for experimental data. For the higher rate of 198 Hz particularly poor results were observed for brainstem responses. It is hypothesised that this may be due to the neural adaptation process and/or an inappropriate choice of source model.

This paper introduces gradient based method for robust assessment of the sparse pulse sources, such as motor unit innervation pulse trains in the filed of electromyography. The method employs multichannel recordings and is based on Convolution Kernel Compensation (CKC). In the first step, the unknown mixing channels (convolution kernels) are compensated, while in the second step the natural gradient algorithm is used to blindly optimize the estimated source pulse trains. The method was tested on the simulated mixtures with random mixing matrices, on synthetic surface electromyograms and on real surface electromyograms, recorded from the external anal sphincter muscle. The results prove the method is highly robust to noise and enables complete reconstruction of up to 10 concurrently active motor units.

Functional Magnetic Resonance Imaging (FMRI) allows indirect observation of brain activity through changes in blood oxygenation, which are driven by neural activity. ICA has become a popular exploratory analysis approach due its advantages over regression methods in accounting for structured noise as well as signals of interest. However, standard ICA in FMRI ignores some of the spatial and temporal structure contained in such data. Using prior knowledge that the Blood Oxygenation Level Dependent (BOLD) response is spatially smooth and manifests itself on certain spatial scales, we estimate the unmixing matrix using only the coarse coefficients of a 3D Discrete Wavelet Transform (DWT). We utilise prior biophysical knowledge that the BOLD response manifests itself mainly at the spatial scales we use for unmixing. Tests on realistic synthetic FMRI data show improved accuracy, greater robustness to misspecification of underlying dimensionality, and an approximate fourfold speed increase; in addition the algorithm becomes parallelizable.

A multivariate non-parametric approach for the processing of fMRI group data is important to address variability of hemodynamic responses across subjects, sessions, and brain regions. Independent component analysis (ICA) has a limitation during the inference of group effects due to a permutation problem of independent components. In order to address this limitation, we present an independent vector analysis (IVA) for the processing of fMRI group data. Compared to the ICA, the IVA offers an extra dimension for the dependent parameters, which can be assigned for the automated grouping of dependent activation patterns across subjects. The IVA was applied to the fMRI data obtained from 12 subjects performing a left-hand motor task. In comparison with conventional univariate methods, IVA successfully characterized the group-representative activation time courses (as component vectors) without extra data processing schemes to circumvent the permutation problem, while effectively detecting the areas with hemodynamic responses deviating from canonical, model-driven ones.

This paper deals with the problem of estimating atrial activity during atrial fibrillation periods in the electrocardiogram (ECG). Since the signal of interest differs in kurtosis sign from the dominant sources in the ECG, we propose an independent component analysis method for source extraction based on the different kurtosis sign and extend it with a constraint of spectral concentration in the 3-12Hz frequency band. Results show that we are able to estimate the atrial fibrillation with a single algorithm having low computational complexity ((7n-7)T).

Exploratory matrix factorization methods like PCA, ICA and sparseNMF are applied to identify marker genes and classify gene expression data sets into different categories for diagnostic purposes or group genes into functional categories for further investigation of related regulatory pathways. Gene expression levels of either human breast cancer (HBC) cell lines [6] or the famous leucemia data set [10] are considered.

Visual perception of transformation invariance, such as translation, rotation and scaling, is one of the important functions of processing visual information in the Brain. To simulate this perception property, we propose a computational model for perception of transformation. First, we briefly introduce the transformation-invariant basis functions learned from natural scenes using Independent Component Analysis (ICA). Then we use these basis functions to construct the perceptual model. By using the correlation coefficients of two neural responses as the measure of transformation-invariance, the model is able to perform the task of perception of transformation. Comparisons with Bilinear Sparse Coding presented by Grimes and Rao and Topo-ICA by Hayvarinen show that the proposed perceptual model has some advantages such as simple to implement and more robust to transformation invariance. Computer simulation results demonstrate that the model successfully simulates the mechanism for visual perception of transformation invariance.

In contrast to the traditional hypothesis-driven methods, independent component analysis (ICA) is commonly used in functional magnetic resonance imaging (fMRI) studies to identify, in a blind manner, spatially independent elements of functional brain activity. ICA is particularly useful in studies with multi-modal stimuli or natural environments, where the brain responses are poorly predictable, and their individual elements may not be directly relatable to the given stimuli. This paper extends earlier work on analyzing the consistency of ICA estimates, by focusing on the spatial variability of the components, and presents a novel method for reliably identifying subspaces of functionally related independent components. Furthermore, two approaches are considered for refining the decomposition within the subspaces. Blind refinement is based on clustering all estimates in the subspace to reveal its internal structure. Guided refinement, incorporating the temporal dynamics of the stimulation, finds particular projections that maximally correlate with the stimuli.

A multi-modal linear mixing model is suggested for simultaneously measured MEG and EEG data. On the basis of this model an ICA decomposition is calculated for a combined MEG and EEG signal vector using the TDSEP algorithm. A single modality demixing procedure is developed to classify ICA components to be multi-modality sources detected by EEG and MEG simultaneously or to be single mode sources. Under this premise, data from 10 subjects are analysed and four exemplary types of sources are selected. We found that these sources represent physically meaningful multi- and single-mode signals: Alpha oscillations, heart activity, eye blinks, and slow signal drifts.