**4. Assessment of the long-term EEG changes**

The use of telemetry to capture continuous recordings has the advantage of allowing the detection of SRS and long term changes in circadian brain oscillations. However, telemetry results in a large accumulation of data. A large volume of data can result in analysis delay, frustration and poor EEG interpretation. Unique tools capable of performing efficient spectral analyses (Rossetti *et al*., 2006; Romcy-Pereira *et al*., 2008; Lehmkuhle *et al*., 2009), seizure estimation, and spike detection (Saab and Gotman, 2005; White *et al*., 2006; Casson *et al*., 2007; Jacquin *et al*., 2007; Hopfengärtner *et al*., 2007) has been used in several studies on epilepsy. Artificial neural networks have proven to be the most reliable tool (Gabor *et al*., 1996; Gabor, 1998; Nigam and Graupe, 2004; Kiymik *et al*., 2004; Tzallas *et al*., 2007; Srinivasan *et al*., 2007; Patnaik and Manyam, 2008) but require tremendous computational power in order to be time effective when analyzing large data sets. Another alternative is the use of commercial software designed for seizure detection. However, most often, this type of software is "tuned" to specific parameters for human subjects, such as sleep stages and spike and wave activity. In some cases, these parameters must be changed between

Use of Telemetric EEG in Brain Injury 211

Also, when dealing with prolonged EEG recordings, usually the recordings are split in several files due to the operational system file size limitation. In this case, it is important to limit the file size (for example, 100 MB), so if a data corruption/loss occurs, there is still a chance to recover some or most of the EEG epochs. However, if the file size is too small (for example 1 MB), several files will be generated making copying of files to another unit for

The objectives of the experiment will influence the tools used for EEG analysis. Parameters, such as changes in power spectra over time, seizure duration and frequency, and number of channels recorded are very important. For example, if more than one channel is recorded, then a coherence and cross-correlation analysis may be performed (electrodes must be

Also, the choice of using commercial EEG software or open source software (or a combination of both) will depend on one's budget and technical background. It is common to find commercial software that allows the processing of large datasets using relatively little memory, but it is not easy to find an open source code with this feature. The graphical user interface (GUI) is another important feature to take into account. Commercial software often includes a user-friendly GUI, while the open-source software GUI is often less functional for the implementation of useful tools. For example, several functions are normally run from the command line when dealing with open source software. Also, the open-source GUI is normally just used to semi-automate the use of certain functions or to visually screen biopotential recordings, such as the EEG. Commercial code is often more stable than open-source code which is constantly aggregated with new functions and therefore, more subject to unexpected errors. Thus, open-source software is not approved for clinical use and should be used only experimentally. User support is another important factor when choosing what type of tool one may use. Commercial software has dedicated personnel for user support, while the help provided by open-source code teams is

Open source software, like the EEGLAB (Delorme and Makeig, 2004) and Chronux Analysis Software (http://chronux.org), is able to open several different file types, while commercial software is restricted to handling a small number of formats. The European data format is open and very flexible (EDF; B Kemp *et al*. 1992; Bob Kemp and Olivan 2003). In fact, EEG MATLAB toolboxes (de Araujo Furtado *et al*., 2009) can benefit from this open file format. Also, the number of functions present in open-source software is virtually unlimited, since one can always add new functions. Thus, the scientific community can always help to implement new analysis approaches. On the other hand, current commercial software is limited since the user cannot add new complex analysis options, since the source code is not available. Normally, the language adopted for open source software is MATLAB. Although it requires an initial investment in a MATLAB license, it still does not compare in terms of price to commercial code that does not have the same flexibility, and functions as MATLAB. Also, a compiled version of a MATLAB toolbox does not require MATLAB. Using MATLAB, one can test new tools that are not present yet in commercial software. Through open source software and MATLAB, one can create figures (in different formats) that are immediately suitable for publications. On the other hand, in commercial software, graphics can only be saved using specific formats. It is very clear that commercial EEG software and

further analysis very slow.

bipolar to run cross-correlogram).

dependent on its availability.

**4.2 Choosing the tools to analyze the EEG data** 

subjects, bringing bias to the analysis. Therefore, in several situations, the use of third-party software tools for the evaluation of large data sets (for example, EEG acquired during longterm pharmacological studies) may be contaminated by bias if the software was originally designed to address a dissimilar problem. However, several groups have invested time in creating tools that permit users, without previous programming experience, to run complex EEG analysis algorithms (Delorme and Makeig, 2004; Mørup *et al*., 2007; Romcy-Pereira *et al*., 2008; de Araujo Furtado *et al*., 2009). Such tools are quite reliable, and some of them are now adjusted for large data sets with multiple parameters, such as EEG, EMG, temperature and gross motor activity.

#### **4.1 Choosing the parameters of acquisition**

Prior to the start of any experiment, it is fundamental to choose the proper parameters of acquisition to optimize further analysis. The objectives, maximum frequency of interest, duration of the experiment, number of channels and available disk storage are key factors in determining the sampling rate and pre-filtering options. Obviously, according to the Nyquist Theorem, the signal must be sampled at least twice the maximum frequency of interest to extract all of the information from the bandwidth and represent the original biopotential (Drongelen, 2006). For example, in order to observe massive oscillations such as hippocampal ripples (Buzsáki *et al*., 1987; Buzsáki *et al*., 2003; ~200 Hz) a sampling rate of over 1 KHz is recommended for practical purposes. Also, if one wants to verify whether electrographic seizures have a behavioral correlate or not, synchronous video should be recorded. The use of 2 EEG channels (250 Hz each) plus temperature (250 Hz), activity (0.1 Hz) and signal strength (16 Hz) recorded in a Data Systems International system (DSI, Arden Hills, MN) results in approximately 175 MB per day. If one performs a 30-day experiment, it will be necessary to reserve approximately 5.2 GB per subject. In this particular example, the animals were placed in individual cages, each positioned on AM radio receiving pads (RPC-1; Data Systems International - DSI, Arden Hills, MN) that detect signals from an implanted transmitter (F40-EET) and send them to an input exchange matrix. Each analog input matrix is capable of receiving input from up to four receivers. A PCI-card model number CQ2240 (Data Systems International - DSI, Arden Hills, MN) receives data input from an exchange matrix. The signal is sent to a computer and telemetry data (up to 16 animals) are recorded through Dataquest ART 4.1 (Acquisition software; Data Systems International - DSI, Arden Hills, MN). The DSI transmitter uses a voltage-controlled oscillator which converts the biopotential difference into a frequency signal. The biopotential channels are encoded in pulse-to-pulse intervals that are transmitted by the F40-EET as RF waves. The relationship between the transmitted interval in microseconds and the input signal in millivolts is described by the calibration entered into the Dataquest ART 4.1 in units of microseconds per millivolts. Attenuation of the signal is very low due to the close proximity of the transmitter to the receiver. The filtering at the device level for the system (implant and acquisition system) is described as less than 3dB attenuation at 1 Hz and 50 Hz in the case of the F40-EET. The filtering within the implanted transmitter is nominally 0.6 Hz (−3dB) for the high-pass filter and 60 Hz (−3dB) for the low-pass filter. It is generated by one-pole of a high-pass filtering and one-pole of low-pass filtering. The activity of each animal is derived from the strength of the signal. When the signal strength changes by a set amount, the data exchange matrix generates an activity count. The number of counts is proportional to both distance and speed of movement. However, the activity is a relative measure, not the distance traveled (de Araujo Furtado *et al*. (2009).

subjects, bringing bias to the analysis. Therefore, in several situations, the use of third-party software tools for the evaluation of large data sets (for example, EEG acquired during longterm pharmacological studies) may be contaminated by bias if the software was originally designed to address a dissimilar problem. However, several groups have invested time in creating tools that permit users, without previous programming experience, to run complex EEG analysis algorithms (Delorme and Makeig, 2004; Mørup *et al*., 2007; Romcy-Pereira *et al*., 2008; de Araujo Furtado *et al*., 2009). Such tools are quite reliable, and some of them are now adjusted for large data sets with multiple parameters, such as EEG, EMG, temperature

Prior to the start of any experiment, it is fundamental to choose the proper parameters of acquisition to optimize further analysis. The objectives, maximum frequency of interest, duration of the experiment, number of channels and available disk storage are key factors in determining the sampling rate and pre-filtering options. Obviously, according to the Nyquist Theorem, the signal must be sampled at least twice the maximum frequency of interest to extract all of the information from the bandwidth and represent the original biopotential (Drongelen, 2006). For example, in order to observe massive oscillations such as hippocampal ripples (Buzsáki *et al*., 1987; Buzsáki *et al*., 2003; ~200 Hz) a sampling rate of over 1 KHz is recommended for practical purposes. Also, if one wants to verify whether electrographic seizures have a behavioral correlate or not, synchronous video should be recorded. The use of 2 EEG channels (250 Hz each) plus temperature (250 Hz), activity (0.1 Hz) and signal strength (16 Hz) recorded in a Data Systems International system (DSI, Arden Hills, MN) results in approximately 175 MB per day. If one performs a 30-day experiment, it will be necessary to reserve approximately 5.2 GB per subject. In this particular example, the animals were placed in individual cages, each positioned on AM radio receiving pads (RPC-1; Data Systems International - DSI, Arden Hills, MN) that detect signals from an implanted transmitter (F40-EET) and send them to an input exchange matrix. Each analog input matrix is capable of receiving input from up to four receivers. A PCI-card model number CQ2240 (Data Systems International - DSI, Arden Hills, MN) receives data input from an exchange matrix. The signal is sent to a computer and telemetry data (up to 16 animals) are recorded through Dataquest ART 4.1 (Acquisition software; Data Systems International - DSI, Arden Hills, MN). The DSI transmitter uses a voltage-controlled oscillator which converts the biopotential difference into a frequency signal. The biopotential channels are encoded in pulse-to-pulse intervals that are transmitted by the F40-EET as RF waves. The relationship between the transmitted interval in microseconds and the input signal in millivolts is described by the calibration entered into the Dataquest ART 4.1 in units of microseconds per millivolts. Attenuation of the signal is very low due to the close proximity of the transmitter to the receiver. The filtering at the device level for the system (implant and acquisition system) is described as less than 3dB attenuation at 1 Hz and 50 Hz in the case of the F40-EET. The filtering within the implanted transmitter is nominally 0.6 Hz (−3dB) for the high-pass filter and 60 Hz (−3dB) for the low-pass filter. It is generated by one-pole of a high-pass filtering and one-pole of low-pass filtering. The activity of each animal is derived from the strength of the signal. When the signal strength changes by a set amount, the data exchange matrix generates an activity count. The number of counts is proportional to both distance and speed of movement. However, the activity is a

relative measure, not the distance traveled (de Araujo Furtado *et al*. (2009).

and gross motor activity.

**4.1 Choosing the parameters of acquisition** 

Also, when dealing with prolonged EEG recordings, usually the recordings are split in several files due to the operational system file size limitation. In this case, it is important to limit the file size (for example, 100 MB), so if a data corruption/loss occurs, there is still a chance to recover some or most of the EEG epochs. However, if the file size is too small (for example 1 MB), several files will be generated making copying of files to another unit for further analysis very slow.

#### **4.2 Choosing the tools to analyze the EEG data**

The objectives of the experiment will influence the tools used for EEG analysis. Parameters, such as changes in power spectra over time, seizure duration and frequency, and number of channels recorded are very important. For example, if more than one channel is recorded, then a coherence and cross-correlation analysis may be performed (electrodes must be bipolar to run cross-correlogram).

Also, the choice of using commercial EEG software or open source software (or a combination of both) will depend on one's budget and technical background. It is common to find commercial software that allows the processing of large datasets using relatively little memory, but it is not easy to find an open source code with this feature. The graphical user interface (GUI) is another important feature to take into account. Commercial software often includes a user-friendly GUI, while the open-source software GUI is often less functional for the implementation of useful tools. For example, several functions are normally run from the command line when dealing with open source software. Also, the open-source GUI is normally just used to semi-automate the use of certain functions or to visually screen biopotential recordings, such as the EEG. Commercial code is often more stable than open-source code which is constantly aggregated with new functions and therefore, more subject to unexpected errors. Thus, open-source software is not approved for clinical use and should be used only experimentally. User support is another important factor when choosing what type of tool one may use. Commercial software has dedicated personnel for user support, while the help provided by open-source code teams is dependent on its availability.

Open source software, like the EEGLAB (Delorme and Makeig, 2004) and Chronux Analysis Software (http://chronux.org), is able to open several different file types, while commercial software is restricted to handling a small number of formats. The European data format is open and very flexible (EDF; B Kemp *et al*. 1992; Bob Kemp and Olivan 2003). In fact, EEG MATLAB toolboxes (de Araujo Furtado *et al*., 2009) can benefit from this open file format. Also, the number of functions present in open-source software is virtually unlimited, since one can always add new functions. Thus, the scientific community can always help to implement new analysis approaches. On the other hand, current commercial software is limited since the user cannot add new complex analysis options, since the source code is not available. Normally, the language adopted for open source software is MATLAB. Although it requires an initial investment in a MATLAB license, it still does not compare in terms of price to commercial code that does not have the same flexibility, and functions as MATLAB. Also, a compiled version of a MATLAB toolbox does not require MATLAB. Using MATLAB, one can test new tools that are not present yet in commercial software. Through open source software and MATLAB, one can create figures (in different formats) that are immediately suitable for publications. On the other hand, in commercial software, graphics can only be saved using specific formats. It is very clear that commercial EEG software and

Use of Telemetric EEG in Brain Injury 213

Fig. 5. Graphical user interface created using MATLAB that allows users to view various properties of EEG signal. Representative cortical electroencephalograms showing electrical

As mentioned, when more than one channel is used, we can investigate the relationship between them. Connectivity between different brain circuits can be evaluated by determining the temporal relationship between brain signals from distinct brain regions. Cross-correlogram may be a better choice to analyze epileptiform patterns, when the propagation of temporal defined events has an important role. The coherence is preferred when the background activity between different areas is comparable (Drongelen, 2006). According to Drongelen (2006), the coherence C between two different signals (x and y) can be defined as Sxy normalized by power spectra Sxx and Syy, respectively. Then, in order to determine coherence, a number (without dimension) between 0 and 1, Sxy is squared. It can

�(�) <sup>=</sup> |���(�)�|

���(���� = ���(��)�(��)} However, the electrodes must be bipolar if ones decide to run a cross-correlogram, so the measurements on each channel will correspond to a better defined and small

In summary, the choice of tools used for signal processing must be carefully evaluated, taking into account the goals and methodological limitations of the study. A reasonable background not only in neuroscience, but mathematics and computer programming may be

The cross-correlation Rxy, between two time series x and y, can be represented by:

neuroanatomical area, something that does not happen with monopolar electrodes.

���(�)���(�)

activity during SE (B). Adapted from de Araujo Furtado and co-workers (2009).

be represented by:

necessary depending on the objectives.

open-source software have advantages and disadvantages and the software selection for EEG analysis is usually not easy.

We record EEG, activity and temperature during the entire course of an experiment and for extended periods beyond SE. A set of MATLAB algorithms was developed (de Araujo Furtado *et al*., 2009) to remove artifacts and measure the characteristics of long-term EEG recordings. The algorithms use short-time Fourier transforms to calculate the power spectrum of the signal for 2-sec intervals. The FFT can be represented by the expression:

$$X(k) = \begin{array}{c} \sum\_{j=1}^{N} \mathcal{X}(j)\omega\_{N}^{(j-1)(k-1)}, \end{array}$$

$$\text{where } a\_N = e^{(2\pi t)/N}$$

The spectrum is then divided into the delta (1-4 Hz), theta (4.1-8 Hz), alpha (8.1-12 Hz), and beta (12.1-25 Hz) bands. Using the MATLAB function "robustfit.m", a linear fit to the power spectrum is used to indicate the likehood of normal EEG activity versus artifacts and high amplitude spike-wave activity. Changes in seizure frequency and duration over a prolonged period are a powerful indicator of the effects of potential neuroprotectants against seizures. The algorithm is very sensitive and, combined with further visual inspection, can give a reliable measurement of both SE duration and SRS frequency. Batch-processing is also used, which is a considerable advantage if a large number of subjects is used (such as in experimental pharmacology), because it increases the level of automation, allowing us to focus on other tasks, while the data is being processed.

Another way to evaluate spectral changes in the EEG signal is through the short-time Fourier transforms (STFT) that can be represented by the expression:

����(�� �) = ���(�) � � (���)�������, where W represents the sliding window used to divide the signal (Romcy-Pereira *et al*., 2008). Normally, when the STFT of a signal is calculated, one can represent it in a spectrogram, where the power spectrum is calculated for different times (t), segmented by the window W. In essence, a spectrogram pictures the distribution of the energy of the signal in the time and in frequency domain. However the size and type of the time window are determinants in the analysis and one must keep in mind the limitations of the STFT: as the time resolution increases (shorter window), the accuracy in which the frequency component is measured is diminished. An alternative is the use of wavelets, but it takes a longer time to process these compared to FFT.

Among the several types of wavelets, the Morlet wavelet is one of the most used to create a time frequency representation of EEG signals that may have epileptiform patterns*.* The Morlet wavelet is represented according to the expression:

�(��� �) = �� � �������� � � �������, where �� = ������ is the time of the wavelet and �� is its frequency (Romcy-Pereira *et al* (2008).

The GUI shown in Fig. 5 simultaneously plots the EEG in the time domain, the FFT, Morlet wavelet transform, allowing the confirmation or rejection of seizures by the user. Although many artifacts are clipped automatically, the user has the chance to verify if artifacts are still present. Gross motor activity and temperature are compared with EEG changes.

It is very important to identify and, if possible, remove artifacts in the EEG recording. Either manually or semi-automatically, artifacts should be rejected prior to any interpretation of the results. Delorme and co-workers (2007) presented an elegant method that uses independent component analysis (ICA) decomposition as a tool to isolate different artifacts from the EEG, although muscle artifact detection was not very effective.

open-source software have advantages and disadvantages and the software selection for

We record EEG, activity and temperature during the entire course of an experiment and for extended periods beyond SE. A set of MATLAB algorithms was developed (de Araujo Furtado *et al*., 2009) to remove artifacts and measure the characteristics of long-term EEG recordings. The algorithms use short-time Fourier transforms to calculate the power spectrum of the signal for 2-sec intervals. The FFT can be represented by the expression:

where �� = �(���)�� The spectrum is then divided into the delta (1-4 Hz), theta (4.1-8 Hz), alpha (8.1-12 Hz), and beta (12.1-25 Hz) bands. Using the MATLAB function "robustfit.m", a linear fit to the power spectrum is used to indicate the likehood of normal EEG activity versus artifacts and high amplitude spike-wave activity. Changes in seizure frequency and duration over a prolonged period are a powerful indicator of the effects of potential neuroprotectants against seizures. The algorithm is very sensitive and, combined with further visual inspection, can give a reliable measurement of both SE duration and SRS frequency. Batch-processing is also used, which is a considerable advantage if a large number of subjects is used (such as in experimental pharmacology), because it increases the level of automation, allowing us to

Another way to evaluate spectral changes in the EEG signal is through the short-time

����(�� �) = ���(�) � � (���)�������, where W represents the sliding window used to divide the signal (Romcy-Pereira *et al*., 2008). Normally, when the STFT of a signal is calculated, one can represent it in a spectrogram, where the power spectrum is calculated for different times (t), segmented by the window W. In essence, a spectrogram pictures the distribution of the energy of the signal in the time and in frequency domain. However the size and type of the time window are determinants in the analysis and one must keep in mind the limitations of the STFT: as the time resolution increases (shorter window), the accuracy in which the frequency component is measured is diminished. An alternative is the

Among the several types of wavelets, the Morlet wavelet is one of the most used to create a time frequency representation of EEG signals that may have epileptiform patterns*.* The

The GUI shown in Fig. 5 simultaneously plots the EEG in the time domain, the FFT, Morlet wavelet transform, allowing the confirmation or rejection of seizures by the user. Although many artifacts are clipped automatically, the user has the chance to verify if artifacts are still

It is very important to identify and, if possible, remove artifacts in the EEG recording. Either manually or semi-automatically, artifacts should be rejected prior to any interpretation of the results. Delorme and co-workers (2007) presented an elegant method that uses independent component analysis (ICA) decomposition as a tool to isolate different artifacts

� �������, where �� = ������ is the time of the wavelet and �� is its

� (���)(���) ��� ,

�(�) = ∑ �(�)��

focus on other tasks, while the data is being processed.

Morlet wavelet is represented according to the expression:

�

frequency (Romcy-Pereira *et al* (2008).

�(��� �) = �� � ��������

Fourier transforms (STFT) that can be represented by the expression:

use of wavelets, but it takes a longer time to process these compared to FFT.

present. Gross motor activity and temperature are compared with EEG changes.

from the EEG, although muscle artifact detection was not very effective.

EEG analysis is usually not easy.

Fig. 5. Graphical user interface created using MATLAB that allows users to view various properties of EEG signal. Representative cortical electroencephalograms showing electrical activity during SE (B). Adapted from de Araujo Furtado and co-workers (2009).

As mentioned, when more than one channel is used, we can investigate the relationship between them. Connectivity between different brain circuits can be evaluated by determining the temporal relationship between brain signals from distinct brain regions. Cross-correlogram may be a better choice to analyze epileptiform patterns, when the propagation of temporal defined events has an important role. The coherence is preferred when the background activity between different areas is comparable (Drongelen, 2006).

According to Drongelen (2006), the coherence C between two different signals (x and y) can be defined as Sxy normalized by power spectra Sxx and Syy, respectively. Then, in order to determine coherence, a number (without dimension) between 0 and 1, Sxy is squared. It can be represented by:

$$C(\omega) = \frac{|\mathcal{S}\_{xy}(\omega)^2|}{\mathcal{S}\_{\infty}(\omega)\mathcal{S}\_{yy}(\omega)}$$

The cross-correlation Rxy, between two time series x and y, can be represented by:

$$R\_{xy}(t\_1 t\_2 = E\{\mathbf{x}(t\_1)y(t\_2)\})$$

However, the electrodes must be bipolar if ones decide to run a cross-correlogram, so the measurements on each channel will correspond to a better defined and small neuroanatomical area, something that does not happen with monopolar electrodes.

In summary, the choice of tools used for signal processing must be carefully evaluated, taking into account the goals and methodological limitations of the study. A reasonable background not only in neuroscience, but mathematics and computer programming may be necessary depending on the objectives.

Use of Telemetric EEG in Brain Injury 215

assistance of Andy Zheng, Michael Addis, Keenan Bailey and Soma Chanda is gratefully

Akulinichev IT, Baevskii RM (1964) Use of radiotelemetry in space medicine. Vestnik

Alsaadi TM, Thieman C, Shatzel A, Farias S (2004) Video-EEG telemetry can be a crucial tool

Bastlund JF, Jennum P, Mohapel P, Vogel V, Watson WP (2004) Measurement of cortical and

Ben-Ari Y, Lagowska J, Tremblay E, Le Gal La Salle G (1979) A new model of focal status

Budgett DM, Hu AP, Si P, Pallas WT, Donnelly MG, Broad JWT, Barrett CJ, Guild S-J,

Buzsáki G (1991) Network properties of memory trace formation in the hippocampus.

Buzsáki G, Buhl DL, Harris KD, Csicsvari J, Czéh B, Morozov a (2003) Hippocampal

Campbell IG (2009) EEG Recording and Analysis for Sleep Research In Current Protocols in

Canguilhem B, Boissin J (1996) [Is the animal in hibernal sleep awake?]. Neurophysiologie

Cantero JL, Atienza M (2005) The role of neural synchronization in the emergence of cognition across the wake-sleep cycle. Reviews in the Neurosciences 16:69-83. Carpentier P, Delamanche IS, Le Bert M, Blanchet G, Bouchaud C (1990) Seizure-related

Castro OW, Furtado MA, Tilelli CQ, Fernandes A, Pajolla GP, Garcia-Cairasco N (2011)

opening of the blood-brain barrier induced by soman: possible correlation with the acute neuropathology observed in poisoned rats. Neurotoxicology 11:493-508. Casson AJ, Yates DC, Patel S, Rodriguez-Villegas E (2007) Algorithm for AEEG data

selection leading to wireless and long term epilepsy monitoring. Conference proceedings : ... Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in Medicine and Biology Society.

Comparative neuroanatomical and temporal characterization of FluoroJade-

secondarily generalized convulsive seizures. Brain Research 163:176-9. Blanc C, Gravier H, Geier S (1967) Radiotelemetric recordings of the EEG of pilots in the

course of flights of long duration. Revue Neurologique 117:222-5. Bonnet MH, Arand DL (1995) We are chronically sleep deprived. Sleep 18:908-11.

Bollettino della Società Italiana di Biologia Sperimentale 67:817-35.

network patterns of activity in the mouse. Neuroscience 116:201-11. Buzsáki G, Haas HL, Anderson EG (1987) Long-term potentiation induced by physiologically relevant stimulus patterns. Brain Research 435:331-3. Caldwell JA, Lewis JA (1995) The feasibility of collecting in-flight EEG data from helicopter

pilots. Aviation, Space, and Environmental Medicine 66:883-9.

clinique = Clinical Neurophysiology 26:1-7.

Seizure : the Journal of the British Epilepsy Association 13:32-4.

radiotelemetry. Journal of neuroscience methods 138:65-72.

for neurologists experienced in epilepsy when diagnosing seizure disorders.

hippocampal epileptiform activity in freely moving rats by means of implantable

epilepticus: intra-amygdaloid application of kainic acid elicits repetitive

Malpas SC (2007) Novel technology for the provision of power to implantable physiological devices. Journal of Applied Physiology (Bethesda, Md.: 1985)

Akademii meditsinskikh nauk SSSR 19:60-6.

acknowledged.

**8. References** 

102:1658-63.

Neuroscience

Conference 2007:2456-9.
