The coordinated activity of neurons organized into networks is believed to act as the major computational units in the brain. In order to understand how the spatio-temporal organization of activity in networks of neurons transform information it is necessary to use high resolution monitoring techniques. Extracellular electrophysiology is currently the best tool for performing high-resolution recording from neural tissue in an awake animal. It offers information about the spiking (output) and synaptic activity (input) of neurons in the recorded area. The interpretation of the data that is gathered using extracellular recording must be patiently interpreted, however, as there can be ambiguities. As more research is carried out it will be possible to better interpret extracellular recording data in terms of neural network activity and to begin to unfold the way that network activities work in concert to transform information.

Introduction

Recording electrical oscillations in the brain began in 1875, when Richard Caton discovered that currents could be recorded from deep inside the brain (Caton, 1875). Extracellular recording and our understanding of the brain have progressed significantly over the past 125 years. Since then, neuroscientists have come to believe that neurons are the fundamental building blocks of the nervous system which integrate incoming signals to generate electrical signals that propagate down their length and influence other neurons via direct electrical or via chemical synapses. These neurons are understood to be organized into networks, layers, maps, and systems in which their coordinated activity mediate the emergence of a large diversity of behaviors.

It has been recognized that understanding the process by which neuronal networks cooperate to perform complex functions will require monitoring the activity of many neurons over time in an awake animal (Buzsaki et al., 1992, Wilson & McNaughton, 1993). This is because the computation in neuronal networks is thought to involve the concerted activity of many neurons working in synchrony over both near (< 1mm) and far (> 1mm) distances. Understanding how the neurons in a network coordinate their activity in space and time will be an essential step in understanding how these networks cooperate to transform sensory information into internal representations that can be used to generate functional behavioral output.

Extracellular electrophysiology is currently the best technique for monitoring the activity of small populations of neurons in an awake animal. This technique offers two types of information about the activity of neurons in a population. By recording from electrodes placed near cells it is possible to record action potentials – the output of neurons – at the millisecond temporal resolution that they occur. Using many electrodes in this manner it will potentially be possible to record from hundreds of neurons simultaneously in tissue volumes less than 1mm³ with little damage to the tissue. Extracellular recording also offers the ability to record information about the cooperating synaptic inputs into a cell population. With both of these types of information in combination with detailed anatomical information about the tissue that is recorded from, it may be possible to be begin to understand the input-output function that determines transformation of information in a neuronal network.

Procedure/Results

In its most simple form, extracellular recording can be performed by placing a single wire in the brain that has insulation covering all but its very tip. Fluctuations in the voltage measured between this wire and a reference wire (commonly attached to a skull screw) can then be measured. Since the fluctuations in the local field potential that occur in the brain are commonly less than a millivolt, the signal must be amplified so that it may be detected and recorded by a common personal computer. In the process of amplifying the signal, it is useful to filter the signal to remove very low (<1Hz) and very high (>3kHz) frequencies.

Using this basic strategy, two types of information can be gathered. Specifically, if the wire is placed near to a neuronal cell body (less than 140 microns according to Henze et al. (2000)), action potentials fired by that cell may be recorded. This is possible because to fire an action potential the neuron transiently opens sodium channels allowing positively charged sodium ions to rush down the voltage gradient into the cell. This movement of ions into the cell creates a negative fluctuation in voltage in the immediately surrounding area relative to distant locations. This leads to a transient change in voltage between the extracellular recording electrode and the distant reference wire. The general method described above can be optimized for the identification and recording from individual neurons. By using multiple fine wires wound together into a tight bundle or silicon probes that have multiple recording sites in tight proximity (less than 50 microns), activity of single neurons may be better isolated (Gray et al., 1995, Harris et al., 2000). This is due to the fact that if a single recording electrode were flanked by two neurons of the same type at equal distance from the recording electrode, it is likely that the voltage fluctuations caused by one neuron firing an action potential would be statistically indistinguishable from the other neuron’s action potentials. However, if multiple nearby sites are recorded from, the two neurons won’t be the same distance from all of the recording sites. Because the voltage fluctuations created by an action potential drop off with distance from the cell, each cell will have a unique signature for the set of recording sites and these signatures can be used to uniquely identify each recorded neuron (though see discussion for potential problems). From a single "tetrode" (four wire recording electrode) it is possible to record from nearly 20 identifiable neurons in densely populated populations. By recording from 128 channels or more in this manner, it is conceivable to record from hundreds of neurons simultaneously in tissue volumes less than a cubic millimeter.

In addition to action potentials, it is also possible to record the coordinated synaptic activity in a population of neurons (though see discussion). This is possible because when a cell receives an excitatory synaptic input it responds by opening ion channels that allow current to flow into the cell. Likewise, an inhibitory synaptic input often results in current flow out of the cell (though shunting inhibition creates no outward current flow). These events summate and the field potential recorded at any given site at any given time reflects the linear sum of fields generated by current sources (e.g. EPSPs) and current sinks (e.g. IPSPs). The recorded fluctuations thus can reflect the synchronized synaptic activity of the population of neurons in the local area around the recording electrode. Fast sodium action potentials don’t significantly contribute to this field recording because of the low-pass (capacitative) filtering properties of the extracellular environment separating the recording electrode and the cell (Nadasdy et al., 1998). By recording from many sites of accurate spacing in a population of neurons and performing a current source density analysis, it is possible to visualize the spatio-temporal order of synaptic activity exhibited in the population. The current source density analysis uses the conductive properties of the neuronal tissue, to convert the recorded voltage fluctuations into current flow. This is important if the conductivity of the tissue isn’t constant throughout because it can lead to sham sources and sinks at the interface between layers (Petsche et al., 1984).

Discussion

There are several benefits to recording neuronal unit activity from extracellular recording electrodes. Probably the largest benefit is that it can be done in an awake, behaving animal. This is both advantageous because recording can be done free of anesthetic and because it is possible to attempt to relate neuronal activity to functional behavior. Another large benefit is that with improving micro-machining techniques and improving personal computers it will be possible to record from hundreds of neurons in a population simultaneously.

However, there are certainly problems with recording unit activity extracellularly. One major problem is that there is no way to know the anatomy of the recorded neuron. In this way the technique limits the ability to make inferences about the structure/function relationship needed to uncover the computational rules governing a network. This problem can be overcome somewhat by the fact that some distinctions can be made between different neuronal types based on the shape of the waveform and firing characteristics of the neuron (Ranck, 1973). However, many neuronal types cannot yet be distinguished by this method (Freund & Buzsaki, 1996). It is possible that simultaneous extracellular and intracellular recordings may reveal physiological characteristics that allow the anatomy to be determined.

The other major problem with recording unit activity extracellularly is that it is difficult to make strong claims about the unit isolation. While the use of tetrodes and other multisite probes help isolation considerably, there are still circumstances in which isolation errors are high. What’s worse is that these conditions have to do with the state of the neuron and therefore may occlude from view the function of the neuron in particular states. For example, by performing simultaneous intracellular and extracellular recording, Harris et al. (2000) found that during a sharp-wave associated ripple event error rates in the isolation of units in the hippocampal CA1 pyramidal layer could exceed 50%. It is also the case that the momentary membrane potential of a neuron can affect the amplitude of the action potential (Nadasdy et al., 1998). Since unit isolation is largely based on action potential amplitude, changes in the momentary membrane potential can have deleterious effects on unit isolation. This is problematic since during the "theta" state in the hippocampus, the action potentials of both pyramidal cells and pyramidal layer interneurons are initiated from more depolarized states (Henze et al., 2000). This could lead to a single neuron in different states to be misrepresented as two neurons. Likewise, the baseline membrane potential of neuron increases over the course of a complex spike burst and thus leads to a decreased amplitude of action potentials late in the burst. This change in action potential amplitude leads to false exclusion of spikes (Harris et al., 2000). Changes in the generation of a dendritic spike upon generation of an action potential can also greatly affect the extracellularly recorded spike. In the same neuron an action potential may travel down different dendrites in an unpredictable manner, thus altering the extracellular signal (Nadasdy et al., 1998). It is possible that this variability is due to dendritic inhibition, which could again lead to systematic errors in unit separation (Buzsaki et al., 1996).

Recording local field potentials from extracellular electrodes can provide 3D information about the cooperating synaptic inputs into a recorded area. When neurons in a network simultaneously receive many concerted inputs, this population event can be recorded as local voltage fluctuations in the extracellular field. If these population events occur regularly, the resulting voltage fluctuations can be seen to rhythmically oscillate. Recording field potentials has revealed that network activity in the brain regularly shifts between two functional states and shows different types of population synchrony within these states (Buzsaki & Traub, 1997).

Despite the useful information that has been gained from recording local field potentials, the interpretations of these data are limited. These limitations are in the present ability to describe the voltage fluctuations in terms of the underlying changes in neuronal firing patterns. In addition to IPSPs and EPSPs, there are a number of possible causes of voltage fluctuations in the brain. For example, one non-synaptic contributor to the local field is the spike afterpotential, which especially contributes when it is slow and long lasting like the calcium-mediated potassium current (Nadasdy et al., 1998). Likewise, it is now believed that intrinsic neuronal currents across the cell membrane can show oscillatory behavior and cause changes in the recorded local field potential (Buzsaki & Traub, 1997). Similarly it has been found that when synaptic activity is blocked, large neuronal populations can show emergent activity that is associated with large (mV) fluctuations in extracellular potentials. While the precise source of these fluctuations isn’t fully understood, it’s clear that this non-synaptic activity can greatly affect the recorded field potential (Buzsaki & Traub, 1997).

To compound the difficulty associated with interpreting the changes in local current seen by a current source density analysis, it is impossible to a priori determine the passive or active nature of the sources and sinks of current (Buzsaki & Traub, 1997). Specifically, if neurons in a population organized in a laminar fashion receive excitatory input to their cell bodies, the result would be for current to flow into the cell bodies and produce an active sink in measured current flow. In a simple case this current would then passively travel down the dendrites and out of the open resting potassium channels, leading to a source of current at the level of the dendrites. This source, however, will be due to a passive return current. It’s possible that this scenario could be confused with a scenario in which the dendrites received an inhibitory input and became an active source, leading the cell bodies to produce a passive sink. This confusion in combination with the multiple possible causes of local field potential fluctuations leaves field potential recording taken at face value to be merely a gross neural correlate.

However, it is possible to push beyond this gross correlation to begin to understand the complex pattern of neural activity in a network that gives rise to the 3D patterns of local field fluctuations gathered from extracellular recording. In particular, when the detailed anatomy of a network is well understood it may be possible to make many inferences about the possible causes of current flow in a region. This information can be combined with information about the activity of interneurons and principle cells recorded simultaneously with the local field potential. This information would be most elegantly gathered from simultaneous intracellular and extracellular recording, where the anatomy of the recorded neurons could be determined. However, the generality of these findings to an awake animal would need to be verified by extracellular recording of units and field potential. Together these recording techniques may allow for the high-resolution recording of local field potentials over a small area of tissue to reveal the complex interactions that are taking place in the network to mediate a functional computation. Although this process is very tedious, it allows for understanding of network activity to be built up slowly and ultimately to be brought into focus of a single technique so that many network operations can be seen at once. Since it’s likely that each operation and the precision timing between them is very important in governing function of a network, developing a systematic method for observing them will be essential to developing an understanding of the functional interplay in and between networks of neurons.

Appendix

Data Acquisition:

Extracellular recording is most commonly performed using tetrodes. These are created by winding four very fine (12 micron) insulated wires around one another and slightly melting the insulation to stiffen them (without creating cross-talk between the wires). This process gives four conductors in a tight spacing. Several of these tetrodes are then attached to a connector and mounted in a drive, which allows them to be independently moved with great accuracy. This drive is then affixed to an animal’s head with skull screws and cement such that the wires are centered over the brain region of interest. Attached to two of these skull screws are wires that serve as a reference wire to perform differential recording and to electrically ground the animal. A grounded wire mesh surrounding the driver can serve as a Faraday cage to reduce external sources of noise. When the animal has recovered from surgery, a cable can be plugged in to carry electrical signals from the recording wires to an amplifier. Commonly a small field effect transistor is used on the animal end of these wires to increase the current that carries the signal to the amplifier so that the signal isn’t washed out by external noise. The signal is then usually amplified by approximately 5000x and filtered below 1Hz and above 3kHz. The signal is then large enough and regular enough in amplitude to be effectively captured by a high throughput A/D converter (preferably at a 20kHz or higher sampling rate per channel) on a personal computer. It is also common to visualize the signals on the computer screen or on an oscilloscope and to listen to them played through a speaker. Both of these online tools help to guide the experimenter in adjusting the depth of the electrode in the brain to achieve the desired recording location.

Data Analysis:

Once the recording has been completed there are some basic procedures that must be performed on the data before further analysis can be performed. Processing of field potential data involves two basic processes. As mentioned before, a current source density analysis can be performed on the data in order to convert the individual voltage fluctuations into current flow into and out of the recording locations. If a particular field event is of interest, one must selectively extract these events using either automatic or manual methods. Similarly, if one wishes to analyze unit data one must extract it from the rest of the data. To do this a copy of the trace is digitally filtered at about 500Hz and a spike discriminator window is passed over each recording trace to extract the data points surrounding each spike and the time of its occurrence. The characteristics of each spike on each tetrode are then transformed using a principal components analysis and the first 3 PCs are used to assign spikes to different neurons. For each tetrode, a 12 dimensional representation of each spike is then generated and these are then statistically clustered using an automatic spike sorting algorithm or viewed as sequential 2 dimensional representations and sorted manually by attempting to find the natural borders to the groups of spikes. Once all the spikes are grouped together based on the cells that elicited them, these data can be analyzed for timing of spikes relative to spikes of other cells, relative to field events, and relative to the behavior/sensory stimulation of the animal.