Electric Stimulation with Sinusoids and White Noise for Neural Prostheses

We are investigating the use of novel stimulus waveforms in neural prostheses to determine whether they can provide more precise control over the temporal and spatial pattern of elicited activity as compared to conventional pulsatile stimulation. To study this, we measured the response of retinal ganglion cells to both sinusoidal and white noise waveforms. The use of cell-attached and whole cell patch clamp recordings allowed the responses to be observed without significant obstruction from the stimulus artifact. Electric stimulation with sinusoids elicited robust responses. White noise analysis was used to derive the linear kernel for the ganglion cell's spiking response as well as for the underlying excitatory currents. These results suggest that in response to electric stimulation, presynaptic retinal neurons exhibit bandpass filtering characteristics with a peak response that occurs 25 ms after onset. The experimental approach demonstrated here may be useful for studying the temporal response properties of other neurons in the CNS.


INTRODUCTION
Neural prostheses generally use electric stimulation with pulse trains to modulate neural activity. However, pulses do not allow fi ne control over the pattern of elicited activity. For example, in retinal prostheses the incidental stimulation of axons on the retinal surface (Jensen et al., 2003) diminishes the fi delity over the spatial pattern of activation. In addition, the temporal resolution of elicited spike trains through activation of the synaptic network with pulsatile stimulation has been quite limited (Jensen and Rizzo III, 2007). The use of alternative stimulus waveforms to improve control over neural activation has not been well studied, possibly because the electrical artifact associated with such waveforms obscures the recorded spikes. As a result, only those spikes that appear after the electric stimulus is turned off can be studied. Here, we demonstrate that patch clamp recordings minimize the stimulus artifact, allowing the neuronal response during stimulation to be studied. We were able to observe both the spiking response and synaptic currents that were elicited by sinusoidal and white noise stimulation.
The choice of stimulus waveform for use in neural prostheses should be made with some understanding of the temporal response properties of the neurons being activated. For example, if the goal is to use sinusoidal stimulation to excite a group of neurons, it is fi rst necessary to determine the timescale over which these neurons respond in order to guide the choice of stimulus frequency. While this could be estimated by varying the frequency of sinusoidal stimulation until the optimal frequency range is found, such an approach is time consuming, and these recordings are often diffi cult to maintain for extended periods of time. We present a more effi cient method using white noise analysis to estimate the temporal response properties of the target neuron. Such analysis is commonly applied to study neuronal responses in the visual (Chichilnisky, 2001) and Electric stimulation with sinusoids and white noise for neural prostheses the exposed area was conical with an approximate height of 100 µm and base diameter of 35 µm. The height of the stimulating electrode remained fi xed at 25 µm above the inner limiting membrane, which is a thin transparent membrane that comprises the inner surface of the retina. The current injected through the stimulating electrode was either a sinusoid of 5-100 Hz, or a zero-mean, pseudo-random binary noise sequence (peak-to-peak, 3.0 µA, unless noted otherwise), where the stimulus level was randomly fl uctuated between −1.5 and 1.5 µA at a time resolution of 1 ms.

RESULTS
We used cell-attached patch clamp recordings to measure the spiking response of retinal ganglion cells to electric stimulation. We found that a relatively tight seal (∼40 MΩ) allowed large amplitude spikes to be recorded without obstruction from the stimulus artifact. Typical responses to sinusoidal stimulation and to stimulation with white noise are shown in Figures 1A,B, respectively. Robust spiking was observed at all sinusoidal stimulus frequencies tested (5-100 Hz), however cells were most sensitive to low frequencies.
In fact, some cells responded to 5-10 Hz stimulation at the lowest stimulus levels we could deliver (1 µA).
Electric stimulation with white noise was highly effective at activating some cells, while other cells did not elicit a response even at the highest stimulus amplitude tested (peak-to-peak 4.4 µA). Following stimulation for 60 s with white noise, the spike times were extracted and the spike-triggered average (STA) was computed by averaging the stimulus waveform that preceded each spike (Rieke et al., 1997). The STA represents the best linear description of the neuronal response to electric stimulation. A typical STA from an ON-ganglion cell is shown in Figure 1C. The STA has two peaks ( Figure 1C): an early peak (<5 ms) and a late phase peak (∼25 ms). The late phase and overshoot of the STA were abolished following application of the synaptic blocker cadmium chloride (not shown), suggesting that the early peak of the STA is mediated by direct excitation, while the late phase is mediated by activation of presynaptic neurons.
Retinal ganglion cells receive synaptic inputs from excitatory bipolar cells and inhibitory amacrine cells. By voltage clamping the ganglion cell membrane potential at −60 mV, we isolated the excitatory currents in response to electric stimulation. Examples are shown for stimulation with sinusoidal stimulation ( Figure 1D) and stimulation with noise ( Figure 1E). By convention, inward, FIGURE 1 | Spike recordings using cell-attached patch clamping allows visualization of the spikes (*) through the stimulus artifact, as shown for sinusoidal (A) and binary white noise (B) stimulation. Both stimulus waveforms elicited robust spiking responses. The spike-triggered average was estimated using 60 s of white noise stimulation (C), which is biphasic in shape with a peak near 25 ms. Excitatory currents were recorded in response to electric stimulation using sinusoids (D) and binary white noise (E). Both stimulus waveforms were highly effective at modulating the synaptic current to the ganglion cell. For stimulation with noise, the linear kernel (F) is estimated by cross-correlation of the stimulus and response. excitatory currents are depicted as negative, downward defl ections. Consistent with the spiking response, both sinusoidal and white noise stimulation elicited robust excitatory currents in the ganglion cell.

Frontiers in Neuroscience | Neuroprosthetics
White noise analysis was used to estimate the temporal relationship between the electric stimulus and the resulting current. The stimulus artifact is very small compared to the excitatory currents, as seen in Figure 1E where the rapid fl uctuations in the response (∼50 pA peak-to-peak) do not obstruct the input currents, which are hundreds of picoamps. After collecting 60 s of white noise data, the raw current response was cross-correlated with the stimulus waveform, giving an estimate of the linear kernel ( Figure 1F). As with the STA in Figure 1C, the peak of the kernel occurs at about 25 ms, followed by an overshoot. The early peak (<1 ms) of the kernel likely results from a capacitive transient, which is present for whole-cell recordings since the cell membrane now separates the stimulating electrode from the recording electrode.

DISCUSSION
Here, we show that both sinusoidal and white noise stimulation elicits robust activity in retinal neurons and that these responses can be studied via use of patch clamp recordings. Our results, although preliminary, raise the possibility that each class of retinal neuron may respond optimally to different frequencies of stimulation. If so, this would be highly advantageous as it would allow very specifi c patterns of neural activity to be elicited. Specifi cally, sinusoids may have the ability to be more selective in activating certain types of neurons because the energy in a sinusoid is confi ned to a single frequency. Conversely, pulses have energy distributed across a wide range of frequencies, thus activating all neurons within the vicinity of the stimulating electrode regardless of their temporal response properties.
The results presented here are consistent with previous studies that revealed two potential means of eliciting a response in ganglion cells with electric stimulation: (1) direct activation of the ganglion cell, (2) activation of neurons that are presynaptic to the ganglion cells (Jensen et al., 2005). Using white noise analysis, we showed the STA contained both a short-latency response (<1 ms), as well as a long-latency response (tens of milliseconds) for activation of neurons presynaptic to the ganglion cell. The long-latency component of the response was abolished with synaptic blockers, further confi rming that this component of the STA is mediated by the activation of presynaptic neurons. The presynaptic component of the STA, as well as the linear kernel estimated with excitatory currents, can be approximated as a bandpass fi lter with peak response near 25 ms.
These results suggest that white noise analysis techniques can be used in conjunction with patch clamp recordings to quickly estimate the temporal relationship between the electric stimulus and the neural response. Estimating the linear approximation to the system using this method has been commonly applied to retinal ganglion cell responses using stimulation with light (Chichilnisky, 2001;Kim and Rieke, 2001), but has yet to be used with electric stimulation. While this linear approximation does not fully account for the many nonlinear features in the system (e.g., rectifi cation or response saturation), the fi lter provides a general description of the time course of the response. Furthermore, white-noise analysis is more suited to characterize complex nonlinear systems than are pulse responses (Marmarelis and Naka, 1972). Using white noise to measure the system dynamics is also more effi cient than measuring the response to a wide range of sinusoidal frequencies and amplitudes. While much work remains in elucidating the underlying neural mechanisms responsive to pulsatile versus non-pulsatile stimulus waveforms, the methods described here provide a means to further explore these questions.