Electroencephalographic (EEG) recordings were obtained from patients undergoing outpatient or in- patient long term video monitoring (LTM) for a sus-.
Proceedings of the 2005 IEEE Engineering in Medicine and Biology 27th Annual Conference Shanghai, China, September 1-4, 2005
Neural Codes in Human Extracranial EEG: Identification of Epilepsy Features Taufik Valiante§∗ , Alan W. L. Chiu† and Berj L. Bardakjian† § Department of Surgery, ∗ Toronto Western Research Institute, † Institute of Biomaterials and Biomedical Engineering, Edward S. Rogers Sr. Department of Electrical and Computer Engineering, University of Toronto, Toronto, Canada abstract- Features of epilepsy from human extracranial EEG recordings were obtained using the wavelet artificial neural network (WANN). The WANN is also a robust signal processing tool for the estimation of nonlinear time-frequency relation and it had previously been shown to be able to classify and predict state transitions in the in-vitro hippocampal slice model exhibiting spontaneous epilepsy. The variations in the power-frequency spectrum were analyzed. The accuracy of state classification was improved when more training data was used. The corresponding changes in synaptic weights between artificial neural units associated with more training data was studied to determine the correlations between learning in WANN and frequency information in human epilepsy. Keywords- human extracranial recording; neural code; wavelet transform; artificial neural network; spontaneous seizures I. INTRODUCTION Epilepsy is characterized by transient interruptions of brain function caused by abnormal temporal and spatial coherent firing of a neuronal population, often referred to as a seizure or ictal event. A typical seizure transpires as the high complexity possibly chaotic (HPC) electrical brain activity changes to low complexity possibly rhythmic (LPR) behaviour. The neural codes are embedded in these dynamic manifolds and the transitions between them. We explore the neural codes in human extracranial EEG exhibiting seizure behaviour. The time-varying frequency information of the neuronal activity can be obtained using wavelet transforms. We isolate the different states of seizure episodes into multiple frequency bands, and observe the dynamical changes of each band during progression into seizure events.
0-7803-8740-6/05/$20.00 ©2005 IEEE.
By incorporating features extracted from the wavelet transforms into an artificial neural network, state classification and transitions of seizure-like events (SLEs) can be captured [1][2]. The difference of synaptic weights distribution in the WANN may also provide clues on the relationship between frequency and learning. Lastly, the WANN implementation can be used as a classifier and predictor for state transitions in human extracranial EEG recordings. It was shown that the WANN can accurately classify over 85% of the preictal and ictal activities. II. HUMAN EXTRACRANIAL RECORDINGS Electroencephalographic (EEG) recordings were obtained from patients undergoing outpatient or inpatient long term video monitoring (LTM) for a suspected seizure disorder. The standard 10-20 system of extra-cranial electrode placement was used. Analogy signals were digitized at 200Hz following bandpass filtering of 0.1 - 100 Hz using standard commercial hardware and software for EEG acquisition and review (XLTEK, Oakville Canada). Baseline periods and seizure events were identified by visual inspection of the EEG and correlated to clinical manifestations captured on video and push-button events generated by the patients. Signals from electrodes manifesting the stereotypic baseline and seizure activity periods were exported to ASCII formatted files for further analysis. Each seizure episode contains two states: preictal and ictal. The wavelet transforms of the human extracranial recordings are shown in FIG. 1. III. METHODOLOGY A) Progression into Seizures Morlet wavelet transform on all recordings was performed continuously over time to obtain time-frequency profiles of the signals up to 48Hz. Quantitative analysis on the time-frequency information was obtained by summing the computed
432
Figure 2: General structure of the WANN for human epilepsy prediction. Time series was recorded and wavelet transform was performed on successive 5 sec intervals. frequency bands up to 48Hz were selected as input sequences to the neural network of size 200 by 2. The output units gave the final output likelihoods. Figure 1: Examples of 20 sec EEG segments from one patient during the preictal state (top trace) and the ictal state (bottom trace) as well as their time frequency profile up to 20 Hz are shown.
were adjusted via supervised back-propagation learning. The effective synaptic weights were then compared to the EEG frequency components when they were traced back to each of the input neural wavelet coefficients in frequency bins of 3.2Hz width unit. Hence, the relationship between learning of in 5 sec intervals. The time-dependent variation of WANN and frequency information of epilepsy can be each bin is monitored to determine if there exists any studied. observable trend using statistical comparison. This study is different from those described in [3] when C) Classification of Seizures frequency components of each extracellular bursts In order to utilize all frequency information, the were analyzed. Whereas, the current technique was WANN was modified so that frequency up to 48Hz applied continuously on the EEG waveforms. were used as input. Seizure episodes not included in the training set were used to validate the WANN. Using different combination of input training data, true B) Synaptic Weights and Learning Wavelet artificial neural network (as shown in FIG. positives (TP) denoting correct classification of pre2) had been designed to classify states and predict ictal state and true negatives (TN) denoting correct state transitions into ictal onsets [1][2]. Utilizing the classification of ictal state were computed. Any imwavelet coefficients as the input domains, the WANN provements in the WANN performance were recorded was also able to provide a nonlinear estimate of the and compared with respect to the amount of training episodes used. frequency information.
Six consecutive extracranial seizure episodes (denoted by Epi1 to Epi6 were used for this analysis. Six WANN (denoted by N et1 to N et6 ) were designed and each WANN used a different combination of episodes as training data. In general, the training set for N eti consisted of Epi1 to Epii . Alternatively, N et6 was trained using all episodes, whereas N et1 used only the first episode. After the networks had been trained and the synaptic weights connecting artificial neural units
IV. RESULTS A) Progression into Seizures Only frequency up to 13Hz was shown (in FIG. 3) because the higher frequency components are extremely small compared to the sub-13Hz activities. No particular trend was observed for activities above 13Hz. In contrary to the analysis of in-vitro hippocampal slice model extracellular field recordings
433
Figure 4:
The absolute values of synaptic weights with respect to input frequencies for N et2 , N et3 and N et6 respectively are shown. The weights in the frequency range of 3-5Hz and 13-18Hz increased most significantly as more training data was used.
Figure 3: The progressions of power-frequency spectrum (up to 13Hz) of the preictal (top trace) and inctal states (bottom trace) are captured. Different frequency bands are shown: 03.2Hz (blue solid), 3.2-6.4Hz (dark green dash), 6.4-9.6Hz (red dots) and 9.6-12.8Hz (light green dot dash).
Model N et2 N et3 N et4 N et5
TP 85±7% 82±8% 80±10% 81±10%
TN 86±7% 84±5% 86±5% 86±6%
[3][4] where extraburst (within burst) frequency components were analyzed, here, no dominate changes Table 1: A summary of seizure classification using in any frequency bands were observed. In addition, WANN. the standard deviations of each frequency band were large compared to extracellular extraburst studies. frequency analysis and features of epilepsy were studied. The use of wavelet transforms has previously B) Synaptic Weights and Learning The changes in effective synaptic weights were been used to determine the start times of seizures studied with respect to the number of training in extracellular field recordings in the in-vitro hipepisodes used. In FIG. 4, the absolute values of pocampal slice model. It was suggested that the imthe synaptic weights were plotted as the function portance of TF methods can be used as tools for the of input frequencies. Input frequencies above 18Hz study of spontaneous and induced changes in oscillawere not shown because their variations were small. tory states. Inspired by these results, wavelet transMore significant increase in the synaptic strength forms are incorporated as a first layer nonlinearity was observed around 3-5 and 13-18Hz. It suggested into our WANN predictor design. Previous studies had demonstrated that strength that theta- and beta- oscillations may be important of high frequency oscillations (HFOs) may be an imin the neuronal codes of epilepsy features. portant indicator for SLE onsets [4]. However, the lack of HFOs ( 320Hz) in human extracranial EEG C) Classification of Seizures recordings due to the relatively low sampling rate The preliminary offline analysis of state classifica(200Hz) did not appear to hinder the performance tion using WANN gave satisfactory results. Classifiof the epileptic state classification. The progression cation can be achieved with over 80% accuracy. of neuronal activities in different frequency bands, leading to seizures and during seizures, were studied. V. DISCUSSION Contrary to the in-vitro model, no major trend in Neural codes were explored in the context of time- power-frequency components were found.
434
ACKNOWLEDGEMENT The intrinsic characteristics of WANN allow easy elimination of the common frequency noise. An arThis work has been supported by the Natural Scitificial neural network in general can be treated as ences and Engineering Council of Canada (NSERC) an estimator of nonlinear relationship between the and the Canadian Institutes of Health Research variables of the input units. In this application, (CIHR). since time-frequency information is used as input, the WANN can generate a nonlinear estimate of epilepREFERENCES tic frequency components. It is a robust technique to estimate the nonlinear relationship between the [1] A. W. L. Chiu, S. Daniel, H. Khosravani, P. L. Carlen, and time-dependent frequency information in any sys- B. L. Bardakjian, “Prediction of seizure onsets in an in-vitro tem. While this nonlinear estimate is not unique, hippocampal slice model of epilepsy using gaussian-based and it nevertheless may give us some insight on the fre- wavelet-based artificial neural networks,” Annals of Biomedical Engineering, vol. 33, no. 6, pp. 798-810, 2005. quency contents in which the neural code is estabA. W. L. Chiu, E. E. Kang, M. Derchansky, P. L. Carlen, lished. As the “memory” associated with an artificial [2] and B. L. Bardakjian, “Online prediction of onsets of seizureneural network was represented by the strength of the like events in hippocampal neural networks using wavelet arsynaptic weights, the learning process of these “mem- tificial neural networks,” revised submission to the Annals of ory” can be treated as the modification of synaptic Biomedical Engineering, 2005. weights when training cases were presented. We com- [3] H. Khosravani, C. R. Pinnegar, J. R. Mitchell, B. L. Barhigh frepared the effective synaptic weights of the WANN dakjian, P. Federico, and P. L. Carlen, “Increased quency oscillations precede in vitro low mg 2+ seizures,” revised with respect to the input frequency information and submission to Brain, 2005 (in press). found that the “memory” of specific frequency ranges [4] A. W. L. Chiu, M. Cotic, S. S. Jahromi, H. Khosravani, P. L. tended to be strengthen with more training cases. Carlen, and B. L. Bardakjian, “The effects of high frequency The WANN had been shown to classify as well as oscillations in hippocampal electrical activities on the classfipredict the time of ictal onset [1][2]. Here we applied cation of epileptiform events using artificial neural networks,” revised submission to the Journal of Neural Engineering, 2005. the WANN for the classification of human extracranial EEG data without detailed understanding of the exact system dynamics. For a neuronal system, this is particularly desirable since generally this information is incomplete. The WANN system is also patientindependent. Its robustness also makes it ideal because it captures the underlying dynamics of seizure episodes instead of focusing on finding the optimal parameters for individual subjects. Future work will incorporate spatial information into the training process of the WANN using nonlinear coherence functions so that it will be able to capture the dynamics of seizure episodes in the spatiotemporal-frequency domain. VI. CONCLUSIONS We studied the features of epilepsy using human extracranial EEG recording to help us understand the coding process of neurons in the frequency domain. Three time-frequency analyses were conducted. It was found that unlike the in-vitro hippocampal slice model, there was no apparent trend in any of the frequency bands prior to seizures. The effective synaptic weights corresponding to the frequency ranges of 3-5 and 13-18Hz in the trained WANN strengthened as more seizure episodes were used for the “learning” of WANN. Finally we applied the WANN for the classification of epileptic states and achieved over 85% accuracy.
435
