On the analysis of single versus multiple channels of electromagnetic brain signals

Summary

Objective

When extracting information from electromagnetic (EM) brain function through recordings such as the electroencephalogram (EEG) it is often assumed that signal processing techniques must be applied to multiple simultaneous recordings in order to obtain useful results. However, sometimes only a single channel of EEG recording is available or desirable. In this paper we objectively assess a novel methodology which exploits only a single measurement channel to extract information of interest relatively independent of channel location (relative to the source of interest).

Methods

The method relies on a combination of a matrix of delay vectors constructed from the single channel measurement, along with constrained independent component analysis, which incorporates prior information into the process.

Materials

Here, we use synthetically generated seizure EEG, composed of real, normal multi-channel EEG onto which is superimposed synthetic epileptic “seizure-like” activity, at different signal-to-noise (SNR) levels, through an equivalent current dipole model.

Results

We show that the method can extract desired information from single channels with a reasonable accuracy even at very small SNR and from channels distant from the focus of the activity. This provides a powerful technique capable of extracting multiple sources underlying single channel recordings and will be useful in situations where only single channel EM recordings of brain function are desirable, such as would be the case in wearable or implantable recording devices.

Introduction

Electromagnetic (EM) brain function can be viewed through recordings of electroencephalographic (EEG) and magnetoencephalographic (MEG) activity. Information can be extracted through signal processing and analysis of the recorded data, and it is commonly assumed that this requires multiple simultaneous recordings. Generally this requires the extraction or identification of a signal of a specific morphology or containing a specific spectral ‘signature’. However, there are instances when only a single channel of recording is available, or when the signal of interest is only poorly represented in the multi-channel recording due to a poor signal-to-noise ratio (SNR). The difficulty of isolating signals of interest is then greatly increased.

In Ref. [1] we presented a novel methodology that exploits only a single measurement channel to extract information of interest, and demonstrated how a broad range of brain characterisations may be obtained from considering just a single channel. Whereas many methods attempt to extract activity using mimetic methods or broadband filtering methods to remove (or at least attenuate) those undesirable components in the recordings, we introduced a method whereby it is possible to break down single channel recordings of the EM brain signals into their underlying components, irrespective of the components’ origin (physiological or otherwise). In Ref. [1], the method we presented relies on the standard implementation of the technique of independent component analysis (ICA). Most methods in the literature applying ICA to biosignal analysis, rely on spatial (i.e., multichannel) analysis (see Ref. [2] where we applied ICA to epileptiform discharge analysis), however, the method we introduced is capable of isolating multiple underlying components using only the temporal information inherent in the single channel recordings (as clearly no spatial information is present, nor can be inferred, in single-channel recordings). In Ref. [3] we applied ICA to both multiple channel recordings of EEG as well as to the individual single channels; in this work we presented a two-dimensional cluster representation of all the resulting (temporal) independent components (ICs) in order to compare both approaches. The use of ICA in this way is a much more powerful technique than simply filtering the recording channels at the frequency band of interest as the independence assumption allows for the separation of sources with some overlap in their frequency content, it has further been shown many times in the literature that the assumption of statistical independence seems to translate well into ‘neurophysiological independence’ i.e. the independent sources extracted are generally quite neurophysiologically meaningful [4].

This paper provides an innovation and an objective assessment of these techniques used to process single channels of data, which has been made possible through the novel introduction of constrained ICA (cICA). cICA allows for prior information about the signal under analysis to be included in the overall analysis and can be used to extract only a single IC. In this framework it allows for the automation of the single channel analysis process. In this paper we pose two questions:

• Is a single channel sufficient to extract temporal activity of interest in EM brain signal analysis? (i.e., can it produce results comparable to those that could be obtained based on multi-channel analysis?)

• If single channel analysis is useful in this way, what is the effect on performance, of the location of the single channel under analysis, relative to the location of any focus of activity of interest?

We must be very clear at this point to state that it is impossible to infer spatial information from a single channel recording—other than information that can be obtained through the knowledge of the actual spatial location of the single channel in question. Given that single channel analysis will only provide knowledge of a temporal waveform, we then ask, what can be inferred from this temporal waveform in the context of a brain source of known morphology and spatial distribution? We are not attempting to solve the inverse problem – i.e. to determine where in the brain a particular source is located, rather we want to know if – given a known spatio-temporal source – single channel analysis can detect the presence of that source and how much this is influenced by the relative locations of the recording channel and the known source.

We attempt to objectively answer these questions using synthetically generated ictal EEG generated from a segment of real multichannel EEG onto which is superimposed known (synthetic) activity, at different mixture levels, through an equivalent current dipole (ECD) in a spherical model of the head. In this context the questions we pose are related to a possible single-channel wearable system that would be used to detect the presence of known epileptiform activity. Whereas, in such a system, the spatio-temporal characteristics of the ictal activity may be assumed to be approximately known, the location of the recording channel may not necessarily be in the most optimal position to detect that particular source.

As the activity is known a priori it is possible to objectively compare the extracted temporal components to the known source. As our particular interest in neurophysiological analysis lies in the analysis of epileptiform activity, the synthetic data we use in this study consists of a sine-wave with a frequency randomly assigned to be between 2 and 10 Hz. The analysis is repeated multiple times, each time using an ECD located in a different region of the head model and with a different orientation and base frequency—this will test the method with different relevant locations between the recording channels and the known source each time. Whilst we have chosen to apply these methods to synthetically generated EEG, it should be apparent that this also readily applies to MEG signals as the underlying model is the same. It should also be noted that here we also assume that the seizure frequency of interest is known a priori, this is quite feasible as it is generally possible to analyse the seizure activity of a given patient on one test recording to obtain these a priori parameters (for each seizure ‘template’ for a given patient) to be used on other recordings. This a priori information is used as the constraint in cICA where a sinusoid of the ‘known seizure frequency’ is used as a temporal constraint.

In the next section, we discuss the major points of our analysis and give an overview of each step of the process. The methods used to measure performance are indicated and in the Section 3 a typical set of results is depicted as well as the overall results after repeatedly applying the method to 100 trials of randomly generated synthetic seizure (ictal) scalp EEG.