Academia.eduAcademia.edu

Figure 8 – uploaded by Lionel Tarassenko

See full PDFdownloadDownload figure
The three Tables show that the ‘generic system’ trained on subject B achieves a level of performance in the analysis of the test recordings from subject A which is comparable to that of the patient-specific classifier trained on that same subject. There is some degrada- tion when subject B is the test subject (with a ‘generic system’ trained on subject A), but that is probably due to the fact that there are only 26 recordings in the data- base for subject A (compared with 73 recordings for subject B) and hence a less comprehensive coverage of spike patterns in that training set. Not surprisingly, the biggest drop in accuracy is seen when group D pro- vides the test subjects for the ‘generic system’ trained

Table 8 The three Tables show that the ‘generic system’ trained on subject B achieves a level of performance in the analysis of the test recordings from subject A which is comparable to that of the patient-specific classifier trained on that same subject. There is some degrada- tion when subject B is the test subject (with a ‘generic system’ trained on subject A), but that is probably due to the fact that there are only 26 recordings in the data- base for subject A (compared with 73 recordings for subject B) and hence a less comprehensive coverage of spike patterns in that training set. Not surprisingly, the biggest drop in accuracy is seen when group D pro- vides the test subjects for the ‘generic system’ trained

Related Figures (12)

In this paper, we extend our work on the neural net- work analysis of the EEG to consider the detection of inter-ictal spikes in epileptic patients. Epilepsy does not refer to a specific disease, but rather to a group of symptoms due to varied causes. It is defined as a chronic brain disorder of various aetiologies character- ised by recurrent seizures (ictal disturbances) due to excessive discharge of cerebral neurons. Between sei- zures, the EEG of subjects who suffer from epilepsy is usually characterised by occasional inter-ictal activity in the form of spikes or spike and wave complexes. When the EEG is recorded in the investigation of epi- lepsy, it is almost always an inter-ictal recording which is made. Occasionally, a seizure may also occur by chance during the recording but, in the majority of instances when epilepsy is confirmed by analysis of the EEG, it is on the basis of inter-ictal rather than ictal activity [7]. Fig.1 Inter-ictal spikes (shown with rectangles) from one of the EEG recordings analysed in this paper A. spike, as shown in Fig. 1, is a sudden burst of electrical activity, lasting for up to 70ms, in an otherwise normal background. A sharp wave is simply a spike of longer duration, usually between 70 and 200ms. Spikes and sharp waves are sometimes followed by a slower wave, hence the description of ‘spike and wave complex’.
In this paper, we extend our work on the neural net- work analysis of the EEG to consider the detection of inter-ictal spikes in epileptic patients. Epilepsy does not refer to a specific disease, but rather to a group of symptoms due to varied causes. It is defined as a chronic brain disorder of various aetiologies character- ised by recurrent seizures (ictal disturbances) due to excessive discharge of cerebral neurons. Between sei- zures, the EEG of subjects who suffer from epilepsy is usually characterised by occasional inter-ictal activity in the form of spikes or spike and wave complexes. When the EEG is recorded in the investigation of epi- lepsy, it is almost always an inter-ictal recording which is made. Occasionally, a seizure may also occur by chance during the recording but, in the majority of instances when epilepsy is confirmed by analysis of the EEG, it is on the basis of inter-ictal rather than ictal activity [7]. Fig.1 Inter-ictal spikes (shown with rectangles) from one of the EEG recordings analysed in this paper A. spike, as shown in Fig. 1, is a sudden burst of electrical activity, lasting for up to 70ms, in an otherwise normal background. A sharp wave is simply a spike of longer duration, usually between 70 and 200ms. Spikes and sharp waves are sometimes followed by a slower wave, hence the description of ‘spike and wave complex’.
The time-domain descriptors, mobility (44) and com- plexity (C) are estimated as follows:
The time-domain descriptors, mobility (44) and com- plexity (C) are estimated as follows:
Since a spike is a sharp transient, it will be character- ised by high values of first and second derivatives. In t S m cto Oo he EEG literature, these parameters are known as the ope, dv/dt (V/s), and sharpness, @v/dt? (V/s*), where y is the time-varying EEG signal. An alternative set of time-domain descriptors have been proposed by Hjorth 4] and subsequently used by Walmsley [15], namely [ activity (4), mobility (4 and complexity (C). These haracterise the amplitude, slope and slope spread of he EEG as described below and the last two could herefore be used as alternatives to the more traditional ope and sharpness parameters. All of the time-domain parameters are calculated using a 3-sample window, as shown in Fig. 2. Let the slope from vo to vy; be so and that from y, to v2 be sy. We have:
Since a spike is a sharp transient, it will be character- ised by high values of first and second derivatives. In t S m cto Oo he EEG literature, these parameters are known as the ope, dv/dt (V/s), and sharpness, @v/dt? (V/s*), where y is the time-varying EEG signal. An alternative set of time-domain descriptors have been proposed by Hjorth 4] and subsequently used by Walmsley [15], namely [ activity (4), mobility (4 and complexity (C). These haracterise the amplitude, slope and slope spread of he EEG as described below and the last two could herefore be used as alternatives to the more traditional ope and sharpness parameters. All of the time-domain parameters are calculated using a 3-sample window, as shown in Fig. 2. Let the slope from vo to vy; be so and that from y, to v2 be sy. We have:
Fig.3 = Histogram plots of the four time-domain parameters for the two classes (normal or epileptic spike) , The normalisation factor is the average of the parameter value in the previous five normal segments. Results are shown for subject B (177 773 normal patterns, 4582 epileptic spikes) a Slope; b Sharpness; ¢ Mobility; d Complexity Although a sharp transient such as a spike can be iden- tified in the time domain, its high-frequency content should also make it detectable in the frequency domain. In all of our previous work on EEG analysis (mostly in the context of sleep disorders), we have made use of autoregressive modelling. AR modelling is best known as an alternative to the discrete Fourier Each of the above discrimination factors was investigated using the database collected from subject B. For the first two indices, it was possible to make use of all the normal patterns (218523) and all the spike patterns (17782). For the last two, the number of available patterns is reduced, since there is a
Fig.3 = Histogram plots of the four time-domain parameters for the two classes (normal or epileptic spike) , The normalisation factor is the average of the parameter value in the previous five normal segments. Results are shown for subject B (177 773 normal patterns, 4582 epileptic spikes) a Slope; b Sharpness; ¢ Mobility; d Complexity Although a sharp transient such as a spike can be iden- tified in the time domain, its high-frequency content should also make it detectable in the frequency domain. In all of our previous work on EEG analysis (mostly in the context of sleep disorders), we have made use of autoregressive modelling. AR modelling is best known as an alternative to the discrete Fourier Each of the above discrimination factors was investigated using the database collected from subject B. For the first two indices, it was possible to make use of all the normal patterns (218523) and all the spike patterns (17782). For the last two, the number of available patterns is reduced, since there is a
Fig.4 Test classification error (%) against model order for reflection coefficients as input features (subject A) 4.2.1 Choice of model order: When autoregtes- sive modelling is used for spectral estimation, a suffi- cient number of poles must be chosen to approximate the shape of the spectrum. The criterion for selecting the optimal order when AR or reflection coefficients are used as a feature vector at the input to a neural network is quite different. If the model order is too low, part of the formation which would discriminate between the two classes will not be present in the input feature vector. If the model order is too high, some of the noise in the EEG signal will also be fitted. Hence the optimal model order is the minimum order which adequately encodes the discriminatory information and this can be established from simple classification exper- iments. Neural network classifiers are trained using a balanced database containing equal numbers of normal and spike patterns from subject A and partitioned into training, validation and test sets. For each model order, the network architecture is determined from performance on the validation set (see below) and the classification error is then evaluated for this architecture on the test set. When this classification etror is plotted as a function of model order, the plot shown in Fig. 4 is obtained. This shows that a model order of 6 is optimal. A similar result is obtained if the expermment is repeated for a balanced database of nor- mal and spike patterns for subject B.
Fig.4 Test classification error (%) against model order for reflection coefficients as input features (subject A) 4.2.1 Choice of model order: When autoregtes- sive modelling is used for spectral estimation, a suffi- cient number of poles must be chosen to approximate the shape of the spectrum. The criterion for selecting the optimal order when AR or reflection coefficients are used as a feature vector at the input to a neural network is quite different. If the model order is too low, part of the formation which would discriminate between the two classes will not be present in the input feature vector. If the model order is too high, some of the noise in the EEG signal will also be fitted. Hence the optimal model order is the minimum order which adequately encodes the discriminatory information and this can be established from simple classification exper- iments. Neural network classifiers are trained using a balanced database containing equal numbers of normal and spike patterns from subject A and partitioned into training, validation and test sets. For each model order, the network architecture is determined from performance on the validation set (see below) and the classification error is then evaluated for this architecture on the test set. When this classification etror is plotted as a function of model order, the plot shown in Fig. 4 is obtained. This shows that a model order of 6 is optimal. A similar result is obtained if the expermment is repeated for a balanced database of nor- mal and spike patterns for subject B.
Table 2: Classification error rates with 6 reflection coeffi- cients and normalised prediction error as input features (balanced databases) The error rates in Tables 1 and 2 represent average classification error rates on the validation sets and the corresponding error rates subsequently evaluated on the test sets with the same MLP. (The test error rates are averages from several experiments, as the S-fold cross-validation procedure is repeated several times, with different random allocations of patterns to the S subsets in each case.)
Table 2: Classification error rates with 6 reflection coeffi- cients and normalised prediction error as input features (balanced databases) The error rates in Tables 1 and 2 represent average classification error rates on the validation sets and the corresponding error rates subsequently evaluated on the test sets with the same MLP. (The test error rates are averages from several experiments, as the S-fold cross-validation procedure is repeated several times, with different random allocations of patterns to the S subsets in each case.)
Duration is the interval between two successive minima which surround the point of maximum slope [12] Table 1: Classification error rates with normalised slope, sharpness and duration as input features (balanced databases)
Duration is the interval between two successive minima which surround the point of maximum slope [12] Table 1: Classification error rates with normalised slope, sharpness and duration as input features (balanced databases)
Table 3: Confusion matrix for MLP outputs In the case of a true positive (TP), the MLP identifies a spike pattern which was labelled as such by the expert. A false positive (FP) is the detection of a spike in an EEG segment which is labelled as normal by the expert. A false negative (FN) indicates that the MLP has missed a spike which the expert has identified in that segment. Finally, in the case of a true negative (TN), the MLP and the expert both agree that the EEG pattern is normal.
Table 3: Confusion matrix for MLP outputs In the case of a true positive (TP), the MLP identifies a spike pattern which was labelled as such by the expert. A false positive (FP) is the detection of a spike in an EEG segment which is labelled as normal by the expert. A false negative (FN) indicates that the MLP has missed a spike which the expert has identified in that segment. Finally, in the case of a true negative (TN), the MLP and the expert both agree that the EEG pattern is normal.
Table 5: Performance of the patient-specific classifiers on an independent test set of recordings for each sub- ject The results for Group C are significantly worse than those for the other subjects. Group C is a set of three different patients, all in a state of active wakefulness when the recordings were made. Their EEG recordings are the noisiest and contain more movement artefact than the other recordings, thereby making it more diffi- cult to detect spikes more accurately.
Table 5: Performance of the patient-specific classifiers on an independent test set of recordings for each sub- ject The results for Group C are significantly worse than those for the other subjects. Group C is a set of three different patients, all in a state of active wakefulness when the recordings were made. Their EEG recordings are the noisiest and contain more movement artefact than the other recordings, thereby making it more diffi- cult to detect spikes more accurately.
Figures are shown for the test sets only Table 4: Patient-specific classifiers for subjects A and B trained using ten-fold cross-validation and balanced databases 7.1 Final system test of patient-specific classifier
Figures are shown for the test sets only Table 4: Patient-specific classifiers for subjects A and B trained using ten-fold cross-validation and balanced databases 7.1 Final system test of patient-specific classifier
Table 7: Performance for the test recordings from sub- ject B for the patient-specific and generic MLP classifiers Table 8: Performance for the test recordings from group D for the patient-specific and generic MLP classifiers
Table 7: Performance for the test recordings from sub- ject B for the patient-specific and generic MLP classifiers Table 8: Performance for the test recordings from group D for the patient-specific and generic MLP classifiers
Table 6: Performance for the test recordings from sub- ject A for the patient-specific and generic MLP classifiers
Table 6: Performance for the test recordings from sub- ject A for the patient-specific and generic MLP classifiers
Related topics:
Electrical and Electronic Engineering
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved