Irregular optogenetic stimulation waveforms can induce naturalistic patterns of hippocampal spectral activity.

Objective. Therapeutic brain stimulation is conventionally delivered using constant-frequency stimulation pulses. Several recent clinical studies have explored how unconventional and irregular temporal stimulation patterns could enable better therapy. However, it is challenging to understand which irregular patterns are most effective for different therapeutic applications given the massively high-dimensional parameter space.Approach. Here we applied many irregular stimulation patterns in a single neural circuit to demonstrate how they can enable new dimensions of neural control compared to conventional stimulation, to guide future exploration of novel stimulation patterns in translational settings. We optogenetically excited the septohippocampal circuit with constant-frequency, nested pulse, sinusoidal, and randomized stimulation waveforms, systematically varying their amplitude and frequency parameters.Main results.We first found equal entrainment of hippocampal oscillations: all waveforms provided similar gamma-power increase, whereas no parameters increased theta-band power above baseline (despite the mechanistic role of the medial septum in driving hippocampal theta oscillations). We then compared each of the effects of each waveform on high-dimensional multi-band activity states using dimensionality reduction methods. Strikingly, we found that conventional stimulation drove predominantly 'artificial' (different from behavioral activity) effects, whereas all irregular waveforms induced activity patterns that more closely resembled behavioral activity.Significance. Our findings suggest that irregular stimulation patterns are not useful when the desired mechanism is to suppress or enhance a single frequency band. However, novel stimulation patterns may provide the greatest benefit for neural control applications where entraining a particular mixture of bands (e.g. if they are associated with different symptoms) or behaviorally-relevant activity is desired.

The Entrainment of the Substantia Nigra During Temporal Lobe Seizures is Dependent on the Seizure Onset Pattern.

Rationale: Temporal lobe (TL) epilepsy is the most common form of drug-resistant epilepsy. While the limbic circuit and the structures composing the TL have been a major focus of human and animal studies on TL seizures, there is also evidence suggesting that the basal ganglia have an active role in the propagation and control of TL seizures. Studies in patients have shown that TL seizures can cause changes in the oscillatory activity of the basal ganglia when the seizures spread to extratemporal structures. Preclinical studies have found that inhibition of the substantia nigra pars reticulata (SN), a major output structure of the basal ganglia, can reduce the duration and severity of TL seizures in animal models. These findings suggest the SN plays a role critical in the maintenance or propagation of TL seizures. Two stereotyped onset patterns commonly observed in TL seizures are low-amplitude fast (LAF) and high-amplitude slow (HAS). Both onset patterns can arise from the same ictogenic circuit, however seizures with LAF onset pattern typically spread farther and have a larger onset zone than HAS. Therefore, we would expect LAF seizures to entrain the SN more so than HAS seizures. Here, we use a nonhuman primate (NHP) model of TL seizures to confirm the implication of the SN in TL seizure and to characterize the relationship between TL seizure onset pattern and the entrainment of the SN. Methods: Recording electrodes were implanted in the hippocampus (HPC) and SN in 2 NHPs. One subject was also implanted with extradural screws for recording activity in the somatosensory cortex (SI). Neural activity from both structures was recorded at a 2 kHz sampling rate. Seizures were induced by intrahippocampal injection of penicillin, which produced multiple spontaneous, nonconvulsive seizures over 3-5 hours. The seizure onset patterns were manually classified as LAF, HAS or other/undetermined. Across all seizures, spectral power and coherence were calculated for the frequency bands 1-7 Hz, 8-12 Hz and 13-25 Hz from/between both structures and compared between the 3 seconds before the seizure, the first 3 seconds of the seizure, and the 3 seconds before seizure offset. These changes were then compared between the LAF and HAS onset patterns. Results: During temporal lobe seizures, the 8-12 Hz and 13-25 Hz power in the SN along with the 1-7 Hz and 13-15 Hz power in the SI was significantly higher during onset than before the seizure. Both the SN and SI had an increase in coherence with the HPC in the 13-25 Hz and 1-7 Hz frequency ranges, respectively. Comparing these differences between LAF and HAS, both were associated with the increase in the HPC/SI coherence, while the increase in HPC/SN increase was specific to LAF. Conclusion: Our findings suggest that the SN may be entrained by temporal lobe seizures secondary to the SI during the farther spreading of LAF seizures, which supports the theory that the SN plays a role in the generalization and/or maintenance of temporal lobe seizures and helps explains the anti-ictogenic effect of SN inhibition.

Meta-Bayesian Optimization for Deep Brain Stimulation.

Deep brain stimulation (DBS) is becoming a fundamental tool for the treatment and study of neurological and psychiatric diseases and disorders. Recently developed DBS devices and electrodes have allowed for more flexible and precise stimulation. Densely packed stimulation contacts can be independently stimulated to shape the electric field, targeting pathways of interest, and avoiding those that may cause side-effects. However, this flexibility comes at a cost. Each additional stimulation setting causes an exponential increase in the number of potential stimulation settings. Recent works have addressed this problem using Bayesian optimization. However, this approach has a limited ability to learn from multiple subjects to improve performance. In this study we extend a recently developed meta-Bayesian optimization algorithm to the DBS domain. We evaluated this approach compared to classical Bayesian optimization and a random search using data collected from a nonhuman primate during stimulation of the subthalamic nucleus while recording evoked potentials in the motor cortex and locally within the subthalamic nucleus. On the task of finding the stimulation setting that maximized the evoked potential across a distribution of generated objective functions, meta-Bayesian optimization significantly outperformed the other approaches with a cumulative reward of 8.93±0.70, compared to 7.17±1.64 for Bayesian optimization (p < 10-9) and 6.89±1.56 for the random search (p < 10-9). Moreover, the algorithm outperformed Bayesian optimization when tested on an objective function not used during training. These results demonstrate that meta-Bayesian optimization can take advantage of the structure underlying a distribution of objective function and learn an optimal search strategy that can generalize beyond the objective functions that were not part of the training data. Clinical Relevance - This extends a meta-Bayesian optimization approach for optimizing DBS stimulation settings that outperforms state-of-art algorithms by 24.6%.

Multi-objective data-driven optimization for improving deep brain stimulation in Parkinson's disease.

Objective.Deep brain stimulation (DBS) is an effective treatment for Parkinson's disease (PD) but its success depends on a time-consuming process of trial-and-error to identify the optimal stimulation settings for each individual patient. Data-driven optimization algorithms have been proposed to efficiently find the stimulation setting that maximizes a quantitative biomarker of symptom relief. However, these algorithms cannot efficiently take into account stimulation settings that may control symptoms but also cause side effects. Here we demonstrate how multi-objective data-driven optimization can be used to find the optimal trade-off between maximizing symptom relief and minimizing side effects.Approach.Cortical and motor evoked potential data collected from PD patients during intraoperative stimulation of the subthalamic nucleus were used to construct a framework for designing and prototyping data-driven multi-objective optimization algorithms. Using this framework, we explored how these techniques can be applied clinically, and characterized the design features critical for solving this optimization problem. Our two optimization objectives were to maximize cortical evoked potentials, a putative biomarker of therapeutic benefit, and to minimize motor potentials, a biomarker of motor side effects.Main Results.Using thisin silicodesign framework, we demonstrated how the optimal trade-off between two objectives can substantially reduce the stimulation parameter space by 61 ± 19%. The best algorithm for identifying the optimal trade-off between the two objectives was a Bayesian optimization approach with an area under the receiver operating characteristic curve of up to 0.94 ± 0.02, which was possible with the use of a surrogate model and a well-tuned acquisition function to efficiently select which stimulation settings to sample.Significance.These findings show that multi-objective optimization is a promising approach for identifying the optimal trade-off between symptom relief and side effects in DBS. Moreover, these approaches can be readily extended to newly discovered biomarkers, adapted to DBS for disorders beyond PD, and can scale with the development of more complex DBS devices.

A framework for designing data-driven optimization systems for neural modulation.

Objective.Neural modulation is a fundamental tool for understanding and treating neurological and psychiatric diseases. However, due to the high-dimensional space, subject-specific responses, and variability within each subject, it is a major challenge to select the stimulation parameters that have the desired effect. Data-driven optimization provides a range of different algorithms and tools for addressing this challenge, but each of these algorithms has specific strengths and limitations, and therefore must be carefully designed for a given neural modulation problem. Here we present a framework for designing data-driven optimization algorithms for neural modulation.Approach.We develop this framework using an optogenetic medial septum stimulation model, where the goal is to find the stimulation parameters that modulate hippocampal gamma power to a desired value. This framework proceeds in four steps: (a) collecting stimulation data, (b) creating high-throughput simulation models, (c) prototyping a range of different data-driven optimization algorithms and evaluating their performance, and (d) deploying the best performing algorithmin vivo. Main results.Following this framework, we prototype and design an algorithm specifically for finding the medial septum optogenetic stimulation parameters that maximize hippocampal gamma power. Building on this, we then change our objective function to find the stimulation parameters that modulate gamma to a specific setpoint, use the framework to understand and anticipate the results before deployingin vivo. Significance.We show that this framework can be used to design an effective optimization solution for a specific neural modulation problem, and discuss how it can potentially be applied beyond the optogenetic medial septum stimulation model.

Optimizing neuromodulation based on surrogate neural states for seizure suppression in a rat temporal lobe epilepsy model.

OBJECTIVE: Developing a new neuromodulation method for epilepsy treatment requires a large amount of time and resources to find effective stimulation parameters and often fails due to inter-subject variability in stimulation effect. As an alternative, we present a novel data-driven surrogate approach which can optimize the neuromodulation efficiently by investigating the stimulation effect on surrogate neural states. APPROACH: Medial septum (MS) optogenetic stimulation was applied for modulating electrophysiological activities of the hippocampus in a rat temporal lobe epilepsy model. For the new approach, we implemented machine learning techniques to describe the pathological neural states and to optimize the stimulation parameters. Specifically, first, we found neural state surrogates to estimate a seizure susceptibility based on hippocampal local field potentials. Second, we modulated the neural state surrogates in a desired way with the subject-specific optimal stimulation parameters found by in vivo Bayesian optimization. Finally, we tested whether modulating the neural state surrogates affected seizure frequency. MAIN RESULTS: We found two neural state surrogates: The first was hippocampal theta power by considering its well-known relationship with epilepsy, and the second was the output of pre-ictal state model (PriSM) which was built by characterizing the hippocampal activity during the pre-ictal period. The optimal stimulation parameters found by Bayesian optimization outperformed the other parameters in terms of modulating the surrogates toward anti-seizure neural state. When treatment efficacy was tested, the subject-specific optimal parameters for increasing theta power were more effective to suppress seizures than fixed stimulation parameter (7 Hz). However, modulation of the other neural state surrogate, PriSM, did not suppress seizures. SIGNIFICANCE: The surrogate approach can save enormous time and resources to find subject-specific optimal stimulation parameters which can effectively modulate neural states and further improve therapeutic effectiveness. This approach can also be used for improving neuromodulation treatment of other neurological or psychiatric diseases.

A Machine Learning Approach to Characterize the Modulation of the Hippocampal Rhythms Via Optogenetic Stimulation of the Medial Septum.

The medial septum (MS) is a potential target for modulating hippocampal activity. However, given the multiple cell types involved, the changes in hippocampal neural activity induced by MS stimulation have not yet been fully characterized. We combined MS optogenetic stimulation with local field potential (LFP) recordings from the hippocampus and leveraged machine learning techniques to explore how activating or inhibiting multiple MS neuronal subpopulations using different optical stimulation parameters affects hippocampal LFP biomarkers. First, of the seven different optogenetic viral vectors used for modulating different neuronal subpopulations, only two induced a substantial change in hippocampal LFP. Second, we found hippocampal low-gamma band to be most effectively modulated by the stimulation. Third, the hippocampal biomarkers were sensitive to the optogenetic virus type and the stimulation frequency, establishing those parameters as the critical ones for the regulation of hippocampal biomarker activity. Last, we built a Gaussian process regression model to describe the relationship between stimulation parameters and activity of the biomarker as well as to identify the optimal parameters for biomarker modulation. This new machine learning approach can further our understanding of the effects of neural stimulation and guide the selection of optimal parameters for neural control.

A characterization of epileptogenesis presented in hippocampal neural activity in a rat tetanus toxin model.

We built a regression model to describe the progress of epileptogenesis in a rat intrahippocampal tetanus toxin (TeNT) epilepsy model by identifying informative neural features from hippocampal local field potentials (LFPs). The LFPs were recorded from the awake and freely behaving animals during the latent period and the active-seizure period. Frequency domain neural features including power spectral density, coherence and phase coherence were calculated from the hippocampal LFPs. A least angle regression with elastic net regularization (LARS-ENR) model successfully predicted a relative day from the first seizure in multiple rats (R2test = 0.724±0.025). By leveraging a characteristic of LARS-ENR which reduces unnecessary features, we identified the neural features related to epileptogenesis in a TeNT model.

Learning State-Dependent Neural Modulation Policies with Bayesian Optimization.

Neural modulation is becoming a fundamental tool for understanding and treating neurological diseases and their implicated neural circuits. Given that neural modulation interventions have high dimensional parameter spaces, one of the challenges is selecting the stimulation parameters that induce the desired effect. Moreover, the effect of a given set of stimulation parameters may change depending on the underlying neural state. In this study, we investigate and address the state-dependent effect of medial septum optogenetic stimulation on the hippocampus. We found that pre-stimulation hippocampal gamma (33-50Hz) power influences the effect of medial septum optogenetic stimulation on during-stimulation hippocampal gamma power. We then construct a simulation platform that models this phenomenon for testing optimization approaches. We then compare the performance of a standard implementation of Bayesian optimization, along with an extension to the algorithm that incorporates pre-stimulation state to learn a state-dependent policy. The state-dependent algorithm outperformed the standard approach, suggesting that incorporating pre-stimulation can improve neural modulation interventions.

Predicting the stimulation effectiveness using pre-stimulation neural states via optogenetic activation of the medial septum glutamatergic neurons modulating the hippocampal neural activity.

In this study, we explored the role of pre-stimulation neural states on the effectiveness of optogenetic stimulation. Optogenetic stimulation was applied to the medial septum glutamatergic neurons to modulate the hippocampal neural activity in a rat tetanus toxin seizure model. The hippocampal local field potential was recorded using a multi electrode array in an awake and behaving rat. Optical stimulation with a 465nm light source was applied at 35Hz in a 20 seconds off / 20 seconds on pattern with simultaneous recording from the hippocampus. Both the baseline and the stimulation period recordings were divided into 2 second segments and used for the further analysis. In the first experiment, a support vector machine (SVM) model classified the neural states by using spectral features between 0 and 50Hz. 447 out of 545 segments (82.02%) were correctly labeled as `Baseline' while only 326 out of 544 (59.93%) segments from the stimulation period were correctly labeled as `Stimulation.' As the ratio of mislabels is significantly higher for the stimulation period (chi-squared, p<;0.01), we concluded that the stimulation was not always effective. In the second experiment, an SVM model predicted the stimulation effectiveness using the spectral features of the pre-stimulation segments. The classification result shows that 63.7% of the pre-stimulation segments correctly predicted the stimulation effectiveness. These findings suggest that the prediction of the stimulation effectiveness may improve the stimulation efficacy by implementing a state-based stimulation protocol.

A machine learning approach to characterizing the effect of asynchronous distributed electrical stimulation on hippocampal neural dynamics in vivo.

Asynchronous distributed microelectrode theta stimulation (ADMETS) of the hippocampus has been shown to reduce seizure frequency in the tetanus toxin rat model of mesial temporal lobe epilepsy suggesting a hypothesis that ADMETS induces a seizure resistant state. Here we present a machine learning approach to characterize the nature of neural state changes induced by distributed stimulation. We applied the stimulation to two animals under sham and ADMETS conditions and used a combination of machine learning techniques on intra-hippocampal recordings of Local Field Potentials (LFPs) to characterize the difference in the neural state between sham and ADMETS. By iteratively fitting a logistic regression with data from the inter-stimulation interval under sham and ADMETS condition we found that the classification performance improves for both animals until 90s post stimulation before leveling out at AUC of 0.64 ± 0.2 and 0.67 ± 0.02 when all inter-stimulation data is included. The models for each animal were re-fit using elastic net regularization to force many of the model coefficients to 0, identifying those that do not optimally contribute to the classifier performance. We found that there is significant variation in the non-zero coefficients between animals (p <; 0.01), suggesting that the ADMETS induced state is represented differently between subject. These findings lay the foundation for using machine learning to robustly and quantitatively characterize neural state.

The influence of the pre-stimulation neural state on the post-stimulation neural dynamics via distributed microstimulation of the hippocampus.

In this study we investigated how the neural state influences how the brain responds to electrical stimulation using a 16-channel microelectrode array with 8 stimulation and recording channels implanted in the rat hippocampus. In two experiments we identified the stimulation threshold at which the brain changes to an afterdischarge state. In one experiment a range of suprathreshold stimulations were applied, and in another the stimulation was not changed. The neural state was measured by the power spectral density prior to stimulation. In the first experiment, these measures and the stimulation parameters were used as features, either together or separately, for training a Support Vector Machine (SVM) classifier to predict whether the stimulation would produce an afterdischarge. In the second experiment, recursive feature elimination was used to iteratively remove the neural state features from the recording channels that had the least impact on the overall accuracy. In the first experiment 43 stimulations elicited 26 afterdischarges. In predicting the post-stimulation state-change (afterdischarge vs. no afterdischarge) the feature space of only neural state had a higher accuracy (67.4%) than when combined with the stimulation parameters (65.1%) or the stimulation parameters alone (58.1%). The overall classification results from both feature spaces containing the neural state were non-independent (chi-squared p <; 0.01). In the second experiment, the channels that were the least predictive were those on the more distal ends of the recording electrode, and the most predictive were clustered in the center of the electrode. Additionally, the accuracy increased when 4 channels were removed. The findings from these experiments suggest that both the pre-stimulation state and the spatial properties from where it is measured can play a role in how neural stimulation can induce functional changes in the hippocampal networks.