PREDICTABILITY OF EPILEPTIC SEIZURE FROM
Section 4.4 Experim ents 81 235 236 237 238 239 240 241 242 243 244
- I __________L_ 235 236 237 236 239 240 241 242 243 244 245 - I __________I__________ I__________ L_ 235 236 237 238 239 240 241 242 243 244 245 “1---T
F ig u re 4.1. Intracranial EEG recordings from patient 1.
- I __________ L_ 235 236 237 238 239 240 241 242 243 244 245 -I--- l_ 236 237 238 239 240 241 242 243 244 245 “i r 235 236 237 238 239 240 241 242 243 244 245 - ■8*
S ection 4 .4 . Experiments 82
S T L m a x o f Intracranial EEG
The data from three patients were tested and the corresponding STL max calculation results are shown in Fig. 4.3. It can be seen th at STL max values are positive, which indicates that the system is chaotic. Because the data used in the experiments were known to have the seizure at the end of the data segment, one can expect the STLmax to show a downward trend at the end of the segment. It can be seen that the STLmax in Fig. 4.3 also present a gradual downward trend prior to the occurrence of seizure, especially in the last two patients. The results are in accordance with all previous research findings using nonlinear dynamic methods.
3.5 1 2.5 Patient 1 100 200 Time (sec) 300 4.5 2.5 Patient 2 100 200 Time (sec) 300 3.5 2.5 X Patient 3 0.5 100 200 300 Time (sec)
F ig u re 4.3. STLmax values from intracranial EEG recordings. The dash line shows the linear approximation of the trend of STLmax.
S T L m a x o f E stim a ted S o u r ces U sin g SO B I
The STLmax values calculated from the estim ated sources by using SOBI are given in Fig. 4.4. For patient 1, there is one source with a minimum value of STLmax of around 250 sec, and another around 150 sec. For patient 2, STLmax of source 2 and source 4 gradually drop before the seizure onset. For patient 3, only source 1 presents a very clear downward trend prior to the seizure.
S ection 4 .4 . E xperim ents 83 Tim e ( S e c o n d s ) T im e ( S e c o n d s ) T im e (S e c o n d s ) 4.5 4 3 2.5 2 100 150 200 250 300 350 T im e ( S e c o n d s ) 0 50 to 150 200 250 300 350 0 50 100
S ection 4 .4 . E xperim ents 84
S T L m ax o f t h e E stim a te d S o u r c e s U sin g J A D E
The STLmax values calculated from the estim ated sources by using JADE are given in Fig. 4.5. For patient 1, source 1 shows the minimum STLmax near the end of the segment. For patient 2, it is seen that source 1 and source 2 present a gradual drop before the seizure onset.
For patient 3, also, there are two sources showing a downward trend prior to the seizure. Compared with the results from SOBI, for patient 2 and patient 3, the STLmax value from JADE seems to produce clearer nonlinear dynamical transitions during the evolution of seizure.
8 Time (Seconds) T im e ( S e c o n d s ) T im e ( S e c o n d s ) a T im e (S e co n d s) T im e (S e co n d s)
S ection 4 .4 . E xperim ents 85
4.4.3 Experiment II
In the second experiment, CTICA was applied to the same scalp EEG data. A component which had the highest correlation with the reference signal was selected for the further nonlinear analysis. The component selected by this way is the source closest to the reference signal. As explained in Chapter 3, the reference signal in CTICA was formed by averaging and filtering the signals from the epileptic area, since the epileptic area and the frequency band w ith seizure were usually known a priori. Therefore, the source closest to the reference will contain more information from seizure area th an the other sources.
Applying the seizure detection techniques to identify seizure com ponents from the ICA outputs is not feasible for seizure prediction , as in long EEG recordings seizure maybe present or absent from time to time. Identification of seizure components in the each EEG segment is not only extremely time consuming, bu t even may not be necessary as seizure maybe absent in some segments. On the other hand, the component closest to the reference is always capable to reflect the dy namic changes in the epileptic area irrespective of the occurrence of the seizure, thereby, it is suitable to be applied for nonlinear dynamic analysis to investigate the predictability of seizure.
In this experiment, the reference signal was selected as the aver
age of F8 and F7 since the epileptic zone was known near the frontal
area, followed by bandpass fi tering w ith frequency range 3 Hz - 15 Hz. The CTICA was performed for each consecutive EEG segment and one source which has the highest correlation with the reference signal was selected. Then, STLmax was estim ated from this source. The main ad vantage of applying CTICA is th a t STLmax can be calculated for one
Section 4 .5 . Conclusion 8 6
source only and no overlapping window is needed, therefore the com putational cost is much less and this approach is more efficient than th at in the first experiment.
Fig. 4.6 gives the results of STLmax obtained from same patients as the above experiment. It is noticed th a t, in all cases, a gradual drop is observed prior to seizure, in accordance with the results from the intracranial EEGs (as in Fig. 4.3).
3.5 3 o' 2.5 I 2 .5 1 1 0.5 0 100 200 T im e ( s e c ) 300 3.5. 3 2.5 2 2 1.5 1 0 100 200 T im e ( s e c ) 300 o' 2.5 T im e ( s e c )
F ig u re 4.6. STLmax values of the seizure source obtained from
CTICA. The dash line shows the linear approximation of the trend of STLmax.
4.5 Conclusion
Predictability of epileptic seizure has been investigated in this chap ter. By using BSS techniques, the seizure components can be extracted from the scalp EEG background. The results of nonlinear quantifica tion present a similar downward trend as for the intracranial recordings, which suggests predictability of seizure from scalp EEG. By incorpo rating the prior spatial and spectral information about the seizure as the constraint, the CTICA algorithm proves to be an effective and su perior method for seizure source separation, which not only can extract the seizure source, but also is less computationally expensive than the
S ection 4 .5 . C onclusion 87