GB2404251A

GB2404251A - Brain signal analysis to detect spikes and rhythmic patterns

Info

Publication number: GB2404251A
Application number: GB0316957A
Authority: GB
Inventors: David Lowe; Christopher J James
Original assignee: Aston University
Current assignee: Aston University
Priority date: 2003-07-19
Filing date: 2003-07-19
Publication date: 2005-01-26
Also published as: GB0316957D0

Abstract

Brain signals are analysed to detect transient features of interest in the signals (e.g. spikes, anomalous activity) by representing multichannel EEG data as multidimensional EEG vectors, comparing the EEG vectors with pre-calibrated reference vectors, and identifying the reference vector that is the closest match to the EEG vector. The reference vectors are associated with a probability of representing a feature of interest. If the probability associated with the identified matching vector exceeds a threshold the ECG data is determined to have a region of interest. The analysed data can be used to assist in the diagnosis of epilepsy. Also disclosed is the analysis of EEG data to identify rhythmic patterns associated with an epileptic seizure, where a data sample window is used.

Description

240425 1

BRAIN SIGNAL ANALYSIS METHOD

The present invention relates to a method of analysing brain signals, particularly, but not exclusively for assisting clinicians in the diagnosis of epi lepsy.

Epilepsy is the name for a group of functional disorders of the brain that are characterised by repetitive seizures. Epilepsy is caused by abnormal, excessive electrical discharges of the neurons in the brain. Epileptic seizures are responses of the brain to many different kinds of insult. Since epilepsies have a multitude of different causes, there are many different types of seizure.

Overall about 5% of people will report a seizure at some time in their lives. However, not everyone who experiences a seizure will develop epilepsy. For the diagnosis of epilepsy to be made, recurrent and unprovoked seizures must be observed. Up to 75% of those diagnosed with epilepsy may become seizure free within five years of diagnosis, following correct, early and uninterrupted (drug) therapy. EEG is of vital diagnostic value to epileptologists. In some cases diagnosis requires recording the seizures during in-patent EEG monitoring in specialised epilepsy monitoring units. This usually requires EEG recording over several days. These patients may be those in whom clinical diagnosis is obscure or those who require precise seizure localisation for epilepsy surgery to take place. EEG is also used as confirmation for a diagnosis.

Once a particular type of procedure has been proposed, the presence and location of activity recorded between seizures (known as interictal activity - 2 - or "spikes") can confirm the diagnosis. Detection of spikes may require many hours of recording in order to obtain just a few instances of this spurious brain activity. Additionally, analysis has traditionally required a highly trained clinician to study potentially hours of multi- channel EEG recording. Not only is this extremely time demanding, but it is inherently prone to error due to the monotonous nature of hours of "normal" EEG punctuated by occasional anomalous activity.

Attempts to automate the analysis to the extent that specific regions of the EEG recording are flagged as being of potential interest to the clinician have been made. One such approach involves an expert defining putative spike behaviour and characterising that behaviour by a discrete number of statistical parameters (such as amplitude, sharpness, etc). for each channel of the EEG (and there are typically 16 to 32). If the threshold for the statistical parameter is exceeded then a spike is logged for that particular channel at a given point in the time series. If sufficient channels log a spike at the same point in time (manually set threshold) then that region is flagged as being of interest. Such systems suffer from a number of problems: the algorithms are slow and generally not very robust. There is a lot of natural variability in the raw EEG data and, depending on the settings, the defined parameter/thresholding technique results in a large number of false positives ("artefacts") requiring a process of artefact rejection normally carried out by the clinician. On the other hand, if attempts are made to reduce the number of artefacts, there is serious danger of missing genuine spikes. Thirdly, the methods are patient specific and require tuning to specific patients. There is thus a need for an improved system which is both quick and reliable. - 3 -

It is the object of the present invention to provide an improved brain signal analysis method which overcomes one or more disadvantages of the known methods and which is preferably faster than real time and robust.

According to a first aspect of the present invention, there is provided a method of analysing brain signals for transient features of interest, said method comprising the steps of: (i) representing multichannel EEG data as a series of correspondingly dimensioned EEG vectors, for each EEG vector; (ii) comparing the EEG vector with a pre-calibrated set of correspondingly dimensioned reference vectors, each reference vector being associated with a probability of representing a feature of interest, (iii) identifying which reference vector corresponds most closely with the EEG vector, (iv) returning the probability associated with the reference vector identified in step (iii), and (v) comparing the probability obtained in step (iv) with a predetermined threshold, such that if the threshold is exceeded, the location (i.e. time) in the EEG data is identified as having a feature of interest.

Preferably, the method also includes an additional step of reporting the locations in the EEG data where the threshold value is exceeded and/or displaying the relevant portion(s) of the EEG data. - 4 -

The method can be applied to pre-recorded EEG data ("off-line") or the method can include the step of obtaining the EEG data and analysing the signals as the EEG data is collected ("on-line").

As defined herein "transient" features of interest are non-regularly occurring anomalous brain activity of relatively short duration. The method is particularly suited to detecting interictal "spikes" which typically occur over a period of about 200 ms.

The method may include a step of filtering the EEG data through a multichannel spatial information filter whereby to make a preliminary assessment of likelihood of the time point corresponding to the data representing a transient feature of interest. Only the data passing through the filter need then be processed according to steps (ii) to (v). Where the feature of interest is a transient spike, a spatial complexity filter is conveniently used, only data being below a predetermined complexity threshold being likely to represent a spike.

Preferably, the method comprises the additional step of pre-processing the EEG data prior to step (i). Such pre-processing preferably involves data reduction, for example, reducing the dimensionality of the EEG data, which will typically have from 16 to 32 channels of data to less than 10 dimensions and more preferably to about 5 dimensions, although it will be understood that the number of dimensions chosen is a balance between speed and information loss. Itwill be understood that where - 5- data reduction takes place, the reference vectors will have the same dimensionality as the reduced data set.

Preferably, the pre-processing is achieved by independent component analysis or singular value decomposition. An advantage of these methods of data reduction is that they provide information on the complexity of the data. Thus, the pre-processing conveniently includes data reduction and filtering. It will be understood that the combination of data reduction and filtering greatly increases the speed at which the EEG data can be analysed.

The first aspect of the invention also resides in a system for carrying out the method, said system comprising: (i) an input for receiving multichannel EEG data, (ii) means for converting multichannel EEG data into multidimensional EEG vectors, (iii) data storage means comprising a look up table comprising reference vectors of the same dimensionality as the EEG vectors, each reference vector having an associated probability of representing an event of interest, (iv) means for comparing each EEG vector with the reference vectors, and identifying the closest match, and (v) an output for outputting the probability value associated with the closest matching vector for each EEG vector. - 6-

Preferably, the system also comprises a display for displaying numerically or graphically the probability values obtained or, more preferably regions of the input BEG data identified as being of interest.

Clearly, the robustness of the method depends on the accuracy of the probabilities associated with the reference vectors. Preferably, the look up table is a trained neural network, each node in the network representing a unique reference vector.

In a preferred embodiment, the neural network is self-organising and the training unsupervised. The training comprises two distinct steps: (i) labelling each node of the neural network with a reference vector, and, (ii) calibrating each reference vector with a probability.

The labelling step is carried out by passing data through the network which utilises a clustering algorithm. The training process is unsupervised in that the input data is not pre-classified as having features of interest or not.

The calibration step is carried out by passing well-classified data (derived from EEG data both having features of interest and not) through the trained network and assigning a probability to each node according to which node corresponds most closely with any given input data. For example, in a preferred embodiment, each node contains a responsibility weighting function (such as a Gaussian) which is used to convert the "distance" of a - 7 - data point to a probability of "belonging" to each node. Conveniently, the neural network comprises 100 nodes.

According to a second aspect of the present invention, there is provided a method of analysing EEG data for rhythmic patterns associated with seizure, said method comprising the steps of, for each channel of EEG data: (i) defining a data sample window comprising two half windows of appropriate size, (ii) separately determining an information measure for each half of the data sample window, (iii) comparing the information measures obtained, and (iv) determining the likelihood of said data sample window containing a rhythmic seizure pattern based on the result of step (iii).

It will be understood that the method will be repeated for each data point in the EEG, thereby providing an indication of where in time along the EEG the seizure pattern was located. Preferably, the method comprises a step of flagging or otherwise highlighting where in the EEG the rhythmic behaviour was determined. Since the method identifies regions where the rhythmic seizure pattern has already started, the method preferably flags and/or displays a portion of the EEG beginning a short time (e.g. 5 seconds) prior to the seizure region. Which channels of the EEG show the rhythmic behaviour will depend on the location of the seizure.

Preferably, a region is only flagged if a minimum threshold of channels exhibiting the behaviour is reached. Typically this will be about 20% of the channels. - 8-

Rhythmic patterns in the brain can arise for a number of reasons e.g. chewing and when resting (so-called alpha waves). However, such rhythmic patterns generally occur at different frequencies. The data sample window is chosen accordingly. For example, a particularly useful window size is 100 ms. For a data sample rate of 5 ms, this corresponds to 20 data samples (10 per half window). Using such a window size, non relevant rhythmic patterns will generally not be located. It will be understood that although the known probable frequency of the seizure wave is a primary factor in determining the window size, other factors, such as noise, may be taken into account.

When the brain enters the characteristic wave associated with a seizure, the information content of the two half windows will tend to be equal.

Steps (ii) and (iii) may be achieved by determining the complexity of the data in each half window and then comparing the two half windows. I Where both windows are below a complexity threshold, the region is determined as a rhythmic seizure pattern.

Advantageously, the method is not patient specific and no training of a neural network is required. Furthermore, the nature of the LEG equipment used and the number of channels of data is not relevant. In addition, because the method identifies which channels are exhibiting the t seizure behaviour, the method can be used to identify spatially where any seizures occurred. - 9 -

Also according to the second aspect of the present invention, there is provided a system for carrying out the method, said system comprising an input and an output, data storage means and one or more processors for carrying out steps (ii) to (iv).

Preferably, said system is portable.

The method of the second aspect may also be regarded as a method for predicting the onset of a brain seizure. The algorithms are sufficiently efficient to operate in real time. The brain enters the rhythmic sine wave pattern characteristic of a seizure shortly before clinical symptoms of the seizure are evident. The period between the brain entering the seizure behaviour and clinical onset may be several seconds, and may therefore be long enough to provide useful warning to an epileptic, for example to enable medication to be taken.

The present invention also resides in a carrier medium carrying a computer executable software program for controlling a computer to carry out the method of the first or second aspects of the present invention.

Preferably, the carrier medium is a storage medium, such as a floppy disk, CD-ROM, DVD or a computer hard drive. Although it will be understood that the carrier medium may also be a transient carrier e.g. an electrical or optical signal.

Claims

Claims 1. A method of analysing brain signals for transient features of

interest, said method comprising the steps of: (i) representing multichannel EEG data as a series of correspondingly dimensioned EEG vectors, for each EEG vectors; (ii) comparing the EEG vector with a pre-calibrated set of correspondingly dimensioned reference vectors, each reference vector being associated with a probability of representing a feature of interest, (iii) identifying which reference vector corresponds most closely with the EEG vector, (iv) returning the probability associated with the reference vector identified in step (iii), and (v) comparing the probability obtained in step (iv) with a predetermined threshold, such that if the threshold is exceeded, the location (i.e. time) in the EEG data is identified as having a feature of interest.
2. The method of claim 1 comprising an additional step of reporting the locations in the EEG data where the threshold value is exceeded and/or displaying the relevant portion(s) of the EEG data.
3. The method of claim 1 or 2, comprising the step of obtaining the EEG data and analysing the signals as the EEG data is collected.
4. The method of any preceding claim, comprising a step of filtering the EEG data through a multichannel spatial information filter whereby to make a preliminary assessment of likelihood of the time point corresponding to the data representing a transient feature of interest.
5. The method of claim 4, wherein only the data passing through the filter is processed according to steps (ii) and (v).
6. The method of claim 4 or 5, wherein the feature of interest is a transient spike and the special information filter is a spatial complexity filter with only data being below a predetermined complexity threshold being likely to represent a spike.
7. The method of any preceding claim, comprising the additional step of pre-processing the EEG data prior to step (i).
8. The method of claim 7, wherein said pre-processing involves data reduction.
9. The method of claim 7 or 8, wherein the pre-processing is achieved by independent component analysis or singular value decomposition.
10. The method of any one of claims 7 to 9, wherein the pre-processing includes data reduction and filtering.
11. A system for carrying out the method of any one of claims 1 to 10 said system comprising: (i) an input for receiving multichannel BEG data, I (ii) means for converting multichannel EEG data into multidimensional EEG vectors, (iii) data storage means comprising a look up table comprising reference vectors of the same dimensionality as the EEG vectors, each reference vector having an associated probability of representing an event of interest, (iv) means for comparing each EEG vector with the reference vectors, and identifying the closest match, and (v) an output for outputting the probability value associated with the closest matching vector for each EEG vector.
12. The system of claim 11 additionally comprising a display for displaying numerically or graphically the regions of the input EEG data identified as being of interest.
13. The system of claim 11 or 12, wherein the look up table is a trained neural network, each node in the network representing a unique reference vector.
14. The system of claim 13, wherein the neural network is self-organising and the training unsupervised.
15. A method of analysing EEG data for rhythmic patterns associated with seizure, said method comprising the steps of, for each channel of EEG data: - (i) defining a data sample window comprising two half windows of appropriate size, (ii) separately determining an information measure for each half of the data sample window, (iii) comparing the information measures obtained, and (iv) determining the likelihood of said data sample window containing a rhythmic seizure pattern based on the results of step (iii).
16. The method of claim 15 comprising a step of flagging or otherwise highlighting where in the BEG the rhythmic behaviour was determined.
17. The method of claim 16, wherein a region is only flagged if a minimum threshold of channels exhibiting the behaviour is reached.
18. The method of any one of claims 15 to 17, wherein steps (ii) and (iii) are achieved by determining the complexity of the data in each half window and then comparing the two half windows, the region being determined as a rhythmic seizure pattern where both windows are below a complexity threshold.
19. A system for carrying out the method of any one of claims 15 to 18, said system comprising an input and an output, data storage means and one or more processors for carrying out steps (ii) and (iv).
20. The system of claim 19 which is portable.