Christian W. Hesse

Independent component analysis for biomedical signals

Independent component analysis (ICA) is increasing in popularity in the field of biomedical signal processing. It is generally used when it is required to separate measured multi-channel biomedical signals into their constituent underlying components. The use of ICA has been facilitated in part by the free availability of toolboxes that implement popular flavours of the techniques. Fundamentally ICA in biomedicine involves the extraction and separation of statistically independent sources underlying multiple measurements of biomedical signals. Technical advances in algorithmic developments implementing ICA are reviewed along with new directions in the field. These advances are specifically summarized with applications to biomedical signals in mind. The basic assumptions that are made when applying ICA are discussed, along with their implications when applied particularly to biomedical signals. ICA as a specific embodiment of blind source separation (BSS) is also discussed, and as a consequence the criterion used for establishing independence between sources is reviewed and this leads to the introduction of ICA/BSS techniques based on time, frequency and joint time-frequency decomposition of the data. Finally, advanced implementations of ICA are illustrated as applied to neurophysiologic signals in the form of electro-magnetic brain signals data.

The FastICA algorithm with spatial constraints

In many blind source separation (BSS) applications, especially for biomedical signal processing, there are specific expectations regarding the spatial and temporal characteristics of some sources, but post-hoc comparisons between source estimates and anticipated outcomes can be complicated and unreliable. One alternative is to incorporate additional prior knowledge, e.g., about the spatial topography of selected source sensor projections, into the BSS approach by means of constraints. This letter describes a modified version of the FastICA algorithm for spatially constrained BSS, where the estimates of selected columns of the mixing matrix are constrained with reference to predetermined source sensor projections.

On Semi-Blind Source Separation Using Spatial Constraints With Applications in EEG Analysis

Blind source separation (BSS) techniques, such as independent component analysis (ICA), are increasingly being used in biomedical signal processing applications, including the analysis of multichannel electroencephalogram (EEG) and magnetoencephalogram (MEG) signals. These methods estimate a set of sources from the observed data, which reflect the underlying physiological signal generating and mixing processes, noise and artifacts. In practice, BSS methods are often applied in the context of additional information and expectations regarding the spatial or temporal characteristics of some sources of interest, whose identification requires complicated post-hoc analysis or, more commonly, manual selection by human experts. An alternative would be to incorporate any available prior knowledge about the source signals or locations into a semi-blind source separation (SBSS) approach, effectively by imposing temporal or spatial constraints on the underlying source mixture model. This work is concerned with biomedical applications of SBSS using spatial constraints, particularly for artifact removal and source tracking in EEG analysis, and provides definitions of different types of spatial constraint along with general guidelines on how these can be implemented in conjunction with conventional BSS methods

Interpreting single trial data using groupwise regularisation

Univariate statistical approaches are often used for the analysis of neuroimaging data but are unable to detect subtle interactions between different components of brain activity. In contrast, multivariate approaches that use classification as a basis are well-suited to detect such interactions, allowing the analysis of neuroimaging data on the single trial level. However, multivariate approaches typically assign a non-zero contribution to every component, making interpretation of the results troublesome. This paper introduces groupwise regularisation as a novel method for finding sparse, and therefore easy to interpret, models that are able to predict the experimental condition to which single trials belong. Furthermore, the obtained models can be constrained in various ways by placing features extracted from the data that are thought to belong together into groups. In order to learn models from data, we introduce a new algorithm that makes use of stability conditions that have been derived in this paper. The algorithm is used to classify multisensor EEG signals recorded for a motor imagery task using (groupwise) regularised logistic regression as the underlying classifier. We show that regularisation dramatically reduces the number of features without reducing the classification rate. This improves model interpretability as it finds features in the data such as mu and beta desynchronisation in the motor cortex contralateral to the imagined movement. By choosing particular groupings we can constrain the regularised solutions such that a lower number of sensors is used or a model is obtained that generalises well over subjects. The identification of a small number of groups of features that best explain the data make groupwise regularisation a useful new tool for single trial analysis.

A novel semiblind signal extraction approach for the removal of eye-blink artifact from EEGs

A novel blind signal extraction (BSE) scheme for the removal of eye-blink artifact from electroencephalogram (EEG) signals is proposed. In this method, in order to remove the artifact, the source extraction algorithm is provided with an estimation of the column of the mixing matrix corresponding to the point source eye-blink artifact. The eye-blink source is first extracted and then cleaned, artifact-removed EEGs are subsequently reconstructed by a deflation method. The a priori knowledge, namely, the vector, corresponding to the spatial distribution of the eye-blink factor, is identified by fitting a space-time-frequency (STF) model to the EEG measurements using the parallel factor (PARAFAC) analysis method. Hence, we call the BSE approach semiblind signal extraction (SBSE). This approach introduces the possibility of incorporating PARAFAC within the blind source extraction framework for single trial EEG processing applications and the respected formulations. Moreover, aiming at extracting the eye-blink artifact, it exploits the spatial as well as temporal prior information during the extraction procedure. Experiments on synthetic data and real EEG measurements confirm that the proposed algorithm effectively identifies and removes the eye-blink artifact from raw EEG measurements.

Tracking Epileptiform Activity in the Multichannel Ictal EEG using Spatially Constrained Independent Component Analysis

Blind source separation (BSS) methods such as independent component analysis (ICA) are increasingly being used in biomedical signal processing for decomposition of multivariate time-series, such as the multichannel electroencephalogram (EEG), into a set of underlying sources, some of which may reflect clinically relevant neurophysiological activity such as epileptic seizures or spikes. Tracking and detecting signals of interest fundamentally requires at least some a priori knowledge or assumptions regarding the spatial and/or temporal characteristics of the target sources. While such prior information is conventionally used during post-processing, it seems equally sensible to incorporate any available information into the data decomposition process from the outset. This work presents an alternative approach to source tracking in multichannel EEG, which exploits prior knowledge of the spatial topographies of the scalp voltage distributions associated with the target sources. The predetermined target topographies are used in conjunction with spatially constrained ICA to extract target source waveforms which are uncontaminated by contributions from coactive and spatially correlated brain and artifact sources. These signals can then be further analyzed in terms of their morphological, spectral or statistical properties. As illustrated in the context of epileptiform EEG, this method is useful for tracking seizures

Extraction of Task-Related Activation From Multi-Echo BOLD fMRI

The development of parallel imaging technology has made possible the acquisition of multiple T2 *-weighted MRI images after a single excitation. This has opened new possibilities for functional MRI using the blood oxygenation level dependent (BOLD) contrast mechanism, which has conventionally acquired a single image at a fixed echo time TE. Regarding the multi-echo functional magnetic resonance imaging (fMRI) time-series at each voxel as a simultaneously sampled multichannel signal facilitates the application of established multichannel source extraction methods, which could provide improved estimates of the underlying signal component reflecting task-related BOLD. This work considers ten methods reflecting three different source extraction approaches in which either the TE dependence of the BOLD contrast is exploited, the correlation with an expected response (or design matrix) is maximized, or a maximally task-related component is selected from a statistical signal decomposition. The performance of these methods in extracting task-related BOLD activation minimally contaminated by head motion artifacts is examined in the context of an fMRI experiment in which the multi-echo data are systematically corrupted with varying degrees of artificially induced head motion. The best results were obtained with least-squares methods applied to log-transformed data, namely, adaptive beamforming using only the echo-times, and Wiener filtering using the design matrix.

Tracking and detection of epileptiform activity in multichannel ictal EEG using signal subspace correlation of seizure source scalp topographies

Conventional methods for monitoring clinical (epileptiform) multichannel electroencephalogram (EEG) signals often involve morphological, spectral or time–frequency analysis on individual channels to determine waveform features for detecting and classifying ictal events (seizures) and inter-ictal spikes. Blind source separation (BSS) methods, such as independent component analysis (ICA), are increasingly being used in biomedical signal processing and EEG analysis for extracting a set of underlying source waveforms and sensor projections from multivariate time-series data, some of which reflect clinically relevant neurophysiological (epileptiform) activity. The work presents an alternative spatial approach to source tracking and detection in multichannel EEG that exploits prior knowledge of the spatial topographies of the sensor projections associated with the target sources. The target source sensor projections are obtained by ICA decomposition of data segments containing representative examples of target source activity, e.g. a seizure or ocular artifact. Source tracking and detection are then based on the subspace correlation between individual target sensor projections and the signal subspace over a moving window. Different window lengths and subspace correlation threshold criteria reflect transient or sustained target source activity. To study the behaviour and potential application of this spatial source tracking and detection approach, the method was used to detect (transient) ocular artifacts and (sustained) seizure activity in two segments of 25-channel EEG data recorded from one epilepsy patient on two separate occasions, with promising and intuitive results.

Mapping scalp topographies of rhythmic EEG activity using temporal decorrelation based constrained ICA

Independent component analysis (ICA) methods are being increasingly applied to the analysis of electromagnetic (EM) brain signals. However, these powerful techniques still generally require subjective a posteriori analysis in order to visualise neurophysiologically meaningful components in the outputs. Standard implementations of ICA are restrictive mainly due to the square mixing assumption (i.e., as many sources as measurement channels) - this is especially so with large multichannel recordings. There are many instances in neurophysiological analysis where there is strong a priori information about the signals being sought; as in tracking the changing scalp topographies of rhythmic activities. Through constraining the ICA solution it is possible to extract signals that are statistically independent, yet which are similar to some reference signal which incorporates the a priori information. We demonstrate this method on a multichannel recording of an epileptiform electroencephalogram (EEG), where we automate the repeated simultaneous extraction of both rhythmic seizure activity, as well as alpha-band activity, over an epoch of EEG. Subjective analysis of the results shows scalp topographies with realistic spatial distributions which conform to our neurophysiologic expectations. This work shows that constraining ICA can be a very useful technique, especially in automated systems and we demonstrate that this can be successfully applied to EM brain signal analysis.

A comparison of time structure and statistically based BSS methods in the context of long-term epileptiform EEG recordings

Blind source separation (BSS) techniques are increasingly being applied to the analysis of biomedical signals in general and electroencephalographic (EEG) signals in particular. The analysis of the long-term monitored epileptiform EEG presents characteristic problems for the implementation of BSS techniques because the ongoing EEG has time varying frequency content which can be both slowly varying and yet also include short bursts of neurophysiologically meaningful activity. Since statistically based BSS methods rely on sample- estimates, which generally require larger window sizes, these methods may extract neurophysiologically uninformative components over short data segments. Here we show that BSS techniques using signal time structure succeed in extracting neurophysiologically meaningful components where their statistical counterparts fail. To this end we use an algorithm that extracts linear mixtures of nonstationary sources, without pre- whitening, through joint diagonalisation of a number of windowed, lagged cross-covariance matrices. We show that this is extremely useful in tracking seizure onset in the epileptiform EEG.