Manipulation of neuronal activity by an artificial spiking neural network implemented on a closed-loop brain-computer interface in non-human primates

Abstract

Objective.Closed-loop brain-computer interfaces can be used to bridge, modulate, or repair damaged connections within the brain to restore functional deficits. Towards this goal, we demonstrate that small artificial spiking neural networks can be bidirectionally interfaced with single neurons (SNs) in the neocortex of non-human primates (NHPs) to create artificial connections between the SNs to manipulate their activity in predictable ways.Approach.Spikes from a small group of SNs were recorded from primary motor cortex of two awake NHPs during rest. The SNs were then interfaced with a small network of integrate-and-fire units (IFUs) that were programmed on a custom clBCI. Spikes from the SNs evoked excitatory and/or inhibitory postsynaptic potentials in the IFUs, which themselves spiked when their membrane potentials exceeded a predetermined threshold. Spikes from the IFUs triggered single pulses of intracortical microstimulation (ICMS) to modulate the activity of the cortical SNs.Main results.We show that the altered closed-loop dynamics within the cortex depends on several factors including the connectivity between the SNs and IFUs, as well as the precise timing of the ICMS. We additionally show that the closed-loop dynamics can reliably be modeled from open-loop measurements.Significance.Our results demonstrate a new type of hybrid biological-artificial neural system based on a clBCI that interfaces SNs in the brain with artificial IFUs to modulate biological activity in the brain. Our model of the closed-loop dynamics may be leveraged in the future to develop training algorithms that shape the closed-loop dynamics of networks in the brain to correct aberrant neural activity and rehabilitate damaged neural circuits.

Keywords: brain-computer interface; closed-loop; intracortical microstimulation; spiking neural networks.

Figure 1.

Experimental design. (A) Diagram showing placement of Utah array in primary motor cortex. The stimulation channels (S) were always adjacent to the recording channels (R), as shown in the example here, to reliably evoke spikes in the SNs. (B) Example of a spontaneous and evoked spike. Shaded area shows stimulation artifact, which is followed 2 ms later by an evoked spike. (C) Example showing how spikes recorded from two SNs (triangles in left panel) on Neurochip3 Recording Events (R1 and R2) cause PSPs in artificial neuron S1 (S1_v in right panel), which spikes when its membrane potential reaches a fixed threshold. When S1 spikes, it sends a trigger to stimulate one of the two SNs (in this example, the top SN in the left panel) after a delay t_stim. (D) Typical experimental timeline, starting with open-loop Poisson distributed stimulation to build a library of SN responses to ICMS, followed by the recording of spontaneous SN activity (Pre). The SNs were then interfaced with the IFUs in an clBCI implemented on the clBCI during closed-loop stimulation. This was sometimes followed by a recording of spontaneous SN activity (Post) to measure any lasting effects from the ICMS. (E) Example of a feedback loop generated by the clBCI. Spontaneous spikes from the SNs can trigger stimulation via the clBCI to evoke new spikes in SNs (second spike in R1 in the right panel). The spontaneous spikes that trigger stimulation are grouped together with the evoked spikes to create an ‘elementary pattern’ (blue box and orange box). Elementary patterns with overlapping spikes are grouped together to identify the complete feedback loop and its size (the number of spikes in the loop). The size of the feedback loop in the example shown here is 5 spikes.

Figure 2.

Simulating the closed-loop dynamics between SNs and IFUs. (A) Markov model to simulate spike trains from SNs. For every interspike interval, ISI₁, from an SN, all possible interspike intervals, ISI₂, that follow ISI₁ are stored in an array. To generate a spike train, a seed is randomly selected from the ISI₁’s, after which an ISI₂ is randomly chosen from the array. This approximates sampling from the probability distribution P(ISI₂ | ISI₁) (B) Stimulation response model. The probability of evoking a spike after stimulation is dependent on the time interval Δt_prev between the previous spike time and stimulation onset and is measured from the data obtained in the OL epoch. Similarly, rebound spikes, which are observed at time intervals Δt_rebound after evoked spikes, are randomly sampled from the distribution of responses measured experimentally. (C) Example PSTH of R3 spike times locked to stimulation onset (S3 → R3 from figure 5) collected from the OL epoch. (D). PSTH of evoked spikes (S3 → R3 from figure 5). (E) Probability distribution of evoking a spike (S3 → R3 from figure 5) as a function of Δt_prev. Grey curve shows unsmoothed distribution, black curve shows the distribution smoothed with a moving average over 10 ms. (F) Original and oversampled distributions of rebound spikes (S3 → R3 from figure 5) following stimulation.

Figure 3.

Spike dynamics during Pre epoch, CL epoch, and simulations for a clBCI with three SNs and three IFUs tested in monkey K. The top panel shows the correlograms across the Pre epoch (black), CL epoch (orange), and simulation of the CL dynamics (blue). The Pearson correlation coefficients between the correlograms during the Pre and CL epochs, as well as between the CL epoch and simulations, are shown for each condition. Across all correlograms, the Pearson correlation coefficients were consistently higher between the CL epoch and simulations than between the Pre and CL epochs, indicating that the simulated clBCI dynamics more closely captured the spike dynamics observed during closed-loop operation than those present during baseline activity. Note that the correlations between the correlograms are symmetric (i.e. R1 → R2 = R2 → R1), so we did not include the correlation coefficients for the correlograms along the lower diagonal. The bottom panels show the network architecture (green and red arrows show excitatory and inhibitory connections, respectively), IFU connection weights (defined as the fraction of the PSP amplitude to IFU activation threshold), locations of recording/stimulation channels on the Utah Array, and the rate of feedback loops of size N and their average sizes during the CL epoch and simulations.

Figure 4.

Same format as figure 3. Experiment performed in monkey J. The black arrows in the autocorrelograms of R3 and R4 show features in correlograms arising from feedback loops created by the clBCI.

Figure 5.

Same format as figure 3. Experiment performed in monkey J. The black arrows in the cross-correlograms of R1 and R2, as well as R3 and R4 show features in the correlograms arising from the ICMS exciting evoked spikes in both SNs with the same pulse.

Figure 6.

Same format as figure 3. Experiment performed in monkey J. Black arrow in cross-correlogram of R1 and R2 show co-excitation of the SNs from ICMS. Black arrow in autocorrelogram of R3 shows low amount of excitation due to the low probability of evoking a spike at small delays between the previous spike time and stimulation onset.

Figure 7.

Effects of inhibition strength and stimulation delays on closed-loop spike dynamics. (A) Color plot of average feedback loop size showing monotonic effects of IPSP magnitude and decay time constant. Simulations were 5 min each and were performed using the clBCI in figure 4. Both variables significantly predicted the change in average feedback loop size in a linear regression model (p < 10⁻¹⁶). (B) Smoothed autocorrelograms (3 ms moving average) of R3 and R4 showing changes in feedback loops with increased IPSP magnitude (red) and increased IPSP decay time constant (gray). (C) Examples of how each corresponding feedback loop from (B) were generated. Times shown under arrows correspond to the peaks observed in (B) due to the feedback loops. (D) Probability of evoking a spike as a function of Δt_prev for the clBCI shown in figure 6. (E) Number of feedback loop sizes between 0-15 spikes for the three different delay combinations used in the three simulations of the clBCI in figure 6 (see text for details). Feedback loops in 21/20 ms clBCI have the largest feedback loop sizes due to R2 and R3 having much larger probabilities of being evoked at larger values of Δt_prev. Feedback loop size distributions differed significantly across stimulation delay conditions (Kruskal–Wallis, p < 0.001), with all pairwise comparisons significant by post-hoc Wilcoxon rank-sum tests (all p < 0.001). (F) Autocorrelograms of R1 and R3 for each of the simulations in (E). (G) Examples of feedback loops for each of the three simulations in (E).

Figure 8.

Quantifying the effects of stimulation artifacts. (A) Percentage of missed spikes as a function of EPSP magnitude (measured as % of activation threshold of IFUs) and network size (number of SNs and IFUs) for the simulations with no evoked spikes (left) and with evoked spikes (right). Linear regression revealed that both connection strength and network size significantly predicted the fraction of missed spikes in both simulation conditions (all p < 10⁻³), with R² = 93.1% without evoked spikes and R² = 96.1% with evoked spikes. (B) Ratio of missed spikes that were evoked spikes over all spikes (evoked and spontaneous spikes) for the simulations with evoked spikes. (C) Relationship between the firing rate of the SNs and the number of spikes that were missed due to stimulation artifacts. An ANCOVA-style analysis further showed that evoked spikes increased the baseline missed spike rate (p = 4.9 × 10⁻⁴), but did not significantly alter the relationship between firing rate and spike loss (p = 0.23). The slightly lower R² in the evoked spike condition at high connection strengths reflects increased variability in evoked spike probabilities and inhibition duration, which disrupted the otherwise monotonic relationship between network parameters and spike obstruction.