National Technical University of Athens (NTUA)
Ανάπτυξη μεθοδολογιών βασισμένων σε μεγέθη μη - γραμμικής δυναμικής για την ανάλυση δεδομένων επιληψίας
Abstract
dc:descriptionIn the current thesis, methodologies based on tools derived from the theory of non-linear dynamics and chaos are developed for analyzing the electroencephalogram (EEG) of epileptic patients. Both scalp and intracranial long-term EEG recordings are used, which are acquired during presurgical evaluation for epilepsy surgery. In the first part of the thesis, a methodology for the automatic detection of epileptic seizures is developed, using scalp EEG. The methodology is based on the method of k-nearest neighbor (k-NN) for fractal dimension (FD) estimation, which is proposed in this study for this purpose and is compared to two other methods of FD estimation (methods of Katz and of Higuchi). The scalp EEG data is acquired from 8 patients with mesial temporal lobe epilepsy, has a total length of 553.14 hours and contains 55 seizures. All three methods are time domain methods (reconstruction of the phase space of the system is not required). Their accuracy of estimation and robustness in the presence of noise is moreover assessed and the values of parameters involved in the calculations are selected, based on synthetic signals of known FD (Weierstrass functions). The k-NN method outperformed the other FD estimation methods based on both synthetic signal analysis (higher robustness in the presence of noise) and EEG analysis (higher seizure detection sensitivity and specificity), and provided seizure detection results which compared favorably to those of the literature. In the second part of the thesis, bivariate synchronization methods are studied towards discriminating the amplitude distributions of the synchronization values of the preictal and interictal periods, using scalp EEG of different frequency bands. For this purpose Receiver Operating Characteristic (ROC) curves are utilized and the statistical significance of the results is assessed using appropriately designed surrogates analysis (seizure time surrogates). One linear (cross-correlation) and two non-linear (average amount of mutual information and mean phase coherence) methods are used, while for the instantaneous phase estimation two approaches are utilized: a common one (the analytical representation of the signal based on the Hilbert transformation) and a new method which, according to its creators, provides high accuracy of estimation of the real phase. Different parameters are studied, as the duration of the preictal period, the level of smoothing of the synchronization time series and different hypotheses for the production of the ROC curves (preictal increase or decrease of synchronization values). It is shown that for each of the patients and each methodology, a statistically significant discrimination of the interictal and preictal values is achieved for some channel pairs and combinations of parameters. Nevertheless, a unique parameters’ combination was not found for which a statistically significant mean result value could be achieved in the patients’ ensemble. In the third part of the thesis, a non-linear multivariate synchronization approach towards producing alarms for epileptic seizure prediction, based on eigenvalue analysis of the mean phase coherence matrix, is developed and applied to simultaneous scalp and intracranial EEG. The v methodology is compared to a corresponding multivariate linear approach, which is based on covariance. The simultaneous scalp and intracranial EEG data is acquired from 6 patients, has a total length of 464.97 hours and contains 28 seizures. The results are statistically assessed using both an appropriately designed analytical approach (based on Poisson distributions) and appropriately designed surrogate analysis approach (alarm times surrogates), which both constitute random seizure prediction methods. Different parameters are studied, i.e. the duration of the preictal period and the level of smoothing of the synchronization time series. The prediction performance achieved by the non-linear methodology as applied to the intracranial EEG is satisfactory (in terms of both sensitivity and specificity), for different combinations of parameters, for four out of six patients. Moreover it exceeds the performance of the analytical random predictor. Nevertheless, the same does not stand in the case of statistical assessment via seizure time surrogates, the performance of which is never exceeded by the performance of the original time series. Application of the methodology to the simultaneous scalp EEG did not provide in any case a performance superior to the performances of any of the random prediction methods. The performance of the linear method was poorer, when compared to the non-linear one. Thus, overall, a statistically significant seizure prediction result was not achieved, despite the initially optimistic indications. Finally, in the fourth part of the thesis, an approach is suggested for associating the genetic substrate of epileptic patients, as defined based on a genetic study of epilepsy in Greece which is currently in progress, with the morphology of EEG of epileptic seizures, as quantified using the non-linear methods of analysis that are studied in the thesis
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2011
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Polychronaki, Georgia-Stavroula
- Πολυχρονάκη, Γεωργία-Σταυρούλα
Subjects
dc:subject × 23- Μη-γραμμική δυναμική ανάλυση
- Μορφοκλασματική διάσταση
- Συγχρονισμός
- Πολυμεταβλητή ανάλυση
- Ανάλυση με υποκατάστατα
- Ηλεκτροεγκεφαλογράφημα
- Επιληψία
- Επιληπτική κρίση
- Ανίχνευση κρίσεων
- Non-linear dynamics
- Fractal dimensions
- Synchronization
- Multivariate analysis
- Surrogate analysis
- Electroencephalogram (EEG)
- Epilepsy
- Epileptic seizure
- Seizure detection
- Seizure prediction
- Ιατρική και Επιστήμες Υγείας
- Βασική Ιατρική
- Medical and Health Sciences
- Basic Medicine
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/26301