Intraoperative Research during Deep Brain Stimulation Surgery

16 Intraoperative Research during Deep Brain Stimulation Surgery

Shane Lee, Meghal Shah, Peter M. Lauro, Wael F. Asaad

Abstract

In this chapter, we discuss the process of conducting intraoperative research during deep brain stimulation surgery. Microelectrode recordings, which are routinely used for intraoperative mapping, present a unique opportunity to listen to and record from neurons in the brain. These recordings, with or without a behavioral task, offer a window into human neuronal circuit function with a granularity that is not otherwise available. This chapter will go over the types of research questions that are amenable to intraoperative neurophysiology research, patient selection, and the additional equipment needed. Considerations such as task development, data analysis, and related neuroimaging are covered. Finally, limitations and ethical considerations are discussed.

Keywords: deep brain stimulation, neurophysiology, microelectrode recordings, intraoperative research, methods, behavioral task, spiking activity

16.1 Introduction

Deep brain stimulation (DBS) surgery presents neurosurgeons with a rare opportunity to observe neural activity in the brain. DBS electrode targeting typically relies upon a combination of imaging and neurophysiology. Though magnetic resonance imaging (MRI) and computed tomography (CT) techniques are becoming more powerful, many gross structures and especially subregions within a therapeutic target of interest still remain difficult to visualize.1 In conjunction with preoperative—and increasingly intraoperative—imaging, microelectrode recordings (MERs) are often used for intraoperative mapping to delineate structural borders and identify subareas within a region of interest (ROI) that may lead overall to improved patient outcomes.^2,3,4,5,6

In combination with carefully designed behavioral assays, recording and analysis of intraoperative neuronal data can provide insight into the functions of these structures and how their activity relates to other areas of the brain, behaviors, or disease processes. This approach has been employed in a growing number of studies, helping to improve understanding of basic and pathological neural activity in essential tremor,^7,8 Parkinson’s disease,^8,9,10 Tourette syndrome,^11,12,13 obsessive-compulsive disorder,^14,15,16 and others. Other types of neural recordings are also being used increasingly in conjunction with MER, including electroencephalography (EEG) and electrocorticography (ECoG), adapting techniques largely pioneered within the context of epilepsy monitoring.^17,18,19

For Parkinson’s disease, the subthalamic nucleus (STN) and globus pallidus pars interna (GPi) are the most common therapeutic targets. Patients have performed tasks manipulating joysticks or haptic gloves while single neuron, multiunit, and local field potentials (LFPs) were recorded in STN or GPi.^5,20 Single neurons in these areas have demonstrated movement-related and direction-specific spike rate modulations, and STN neurons further showed oscillations at 3 to 5 Hz “tremor” frequencies or 15 to 30 Hz “beta” frequencies.^5,6,21 Other studies have engaged awake patients with tasks designed specifically to correlate neural activity with precise aspects of behavior. Both Zavala et al and Zaghloul et al demonstrated in distinct decision-making tasks that neuronal firing in the STN is correlated with conflict.^22,23 Using simultaneously recorded scalp EEG, Zavala et al showed that this STN activity was driven by activity in the frontal cortex.

In this chapter, practical considerations for conducting human intraoperative neurophysiology research are discussed.

16.2 Formulating Hypotheses

In developing a neurophysiology research study with human subjects, investigators must address the following questions while designing an experiment:

1. What cortical or subcortical structures are of interest?

2. What disease processes are of interest or would provide access to the structure in question?

3. Will this be an observational study or will there be behavioral task or measure?

4. What type(s) of neurophysiological recordings will be acquired?

The greatest potential limitation of intraoperative research is that by nature of the procedure, only patients with neurological disease will undergo DBS surgery. This may limit the interpretation of the data and may also limit the structures available for MER. Thus, the most common targets accessible in this fashion are the STN and GPi (patients with Parkinson’s, the latter also for primary dystonia)24 and the ventral intermediate nucleus (Vim) of the thalamus (patients with essential tremor).^8,25 MER can help localize specific substructures within these areas that are most desirable for electrode implantation.

In addition to these movement disorders that are now routinely treated with DBS, treatment of a number of psychiatric conditions has been explored with DBS therapy. For example, in extreme cases of obsessive-compulsive disorder and Tourette syndrome (a combination of both motor and psychiatric pathology) the ventral internal capsule/ventral striatum or cingulate cortex has been targeted for DBS with potentially impressive benefits in some patients.^14,15,16 For intractable obsessive-compulsive disorder, others have targeted the STN or the ventral anterior internal capsule/inferior thalamic peduncle.²⁶

Recordings can be made through nontarget structures that are encountered along the trajectory to the target structure, such as frontal cortex and striatum, and in some cases just beyond the target structure, such as the substantia nigra, if such regions are routinely mapped to define a target’s distal border.²⁷ In some cases, with the proper approvals, cortical recordings can be made with subdural electrodes, not typically required for DBS surgery, inserted through the standard burr holes.10

Often, a roadmap regarding what types of behaviors may be mediated by particular structures is available in the form of prior human or nonhuman primate functional MRI (fMRI) studies and in the wide body of literature describing electrophysiological correlates of behavior in animal studies. Adapting these behavioral paradigms to humans offers the opportunity to extend our knowledge of the neural correlates of behavior, especially those behaviors which may be elaborated in or are unique to humans.

16.3 Patient Selection and IRB Approval

Approval of the research protocol by an institutional review board (IRB) is mandatory even for observational studies. Patients who are considered for DBS are ideally first evaluated by a multidisciplinary team of clinicians. Appropriateness and fitness for surgery is determined by the surgeon, neurologist, anesthesiologist, and any other clinicians who are involved in the patient’s care. Once patients are considered appropriate for surgery, they can be approached by a member of the clinical or research team per their specific IRB protocol to obtain voluntary consent after explaining the potential risks of the research-specific procedures (outlined in greater detail below). Because patients typically desire to please their physicians, especially in situations where they think this could improve the care or attention they receive, one should explain clearly that the quality of care provided will not depend on their participation. In addition, respecting a patient’s decision-making autonomy extends throughout the process such that they should be allowed to withdraw from participation at any time, including during the procedure.28

Potential risks of intraoperative research include those related to additional time incurred during surgery to carry out the experimental procedures, such as behavioral tasks, the placement of additional electrodes (e.g., subdural electrodes) that are not typically required for the clinical procedure, and discomfort of the patient or anxiety related to task performance. Particular research protocols may incur other risks. In general, risks accrue as a result of any nonstandard surgical maneuvers or deviations from the clinical procedure. For example, the placement of subdural ECoG electrodes is not required for routine DBS procedure. While placement has been reported as generally safe, there is nonetheless a nonzero risk associated with any additional maneuver, and there may be risks not immediately considered (e.g., the additional time required to insert an ECoG electrode may result in increased pneumocephalus which could affect the accuracy of final DBS electrode placement).²⁹ While in some studies ECoG electrodes are used in hopes of improving the future efficacy of neuromodulation (such as a source of control signals for closed-loop DBS), in other cases the goal may be basic science. It may be easier, therefore, to justify the additional maneuver in the former case than in the latter, so careful deliberation over these issues is mandatory. Simply because an IRB may be convinced that a particular protocol is reasonable does not mean that the protocol is necessarily in a patient’s best interest.

16.4 Equipment and Setup

In most of human acute recordings, the operating theater also serves as the laboratory. In this unique arrangement, some of the equipment serve a principally clinical purpose but may also serve research goals with no or minimal modification.

MER in DBS allows an assessment of somatotopic responses, in which high bandpass filtered neural recordings at multiple sites are monitored over audio speakers while a clinician elicits various neural responses by manipulating the face/jaw and limbs. Typically, 1 to 5 microelectrodes arranged in a Ben-Gun array are advanced toward a predefined target structure while somatotopic assessments performed at various locations along the trajectories.

In general, neural data recording requires electrodes, signal amplifiers, and an acquisition system. Depending on the research questions, additional systems may be necessary to measure movement or administer tasks to awake patients while recording. An example of multichannel neural and behavioral recording is shown in Fig. 16.1.

Typical sharp tungsten or platinum-iridium electrodes with impedances around 300 to 1000 kΩ are typically used to record single- and multiunit spiking activity. Online during a case, signals measured with these electrodes are typically bandpass filtered from approximately 300 Hz to around 10 kHz, appropriate for isolating action potentials from neurons surrounding the recording tip.

The Nyquist sampling theorem sets a lower bound on the appropriate sampling rate of the digital acquisition system. Nyquist states that the sampling rate must be two times greater than the maximum frequency of the activity of interest. For example, if one wants to sample single-unit activity at 10 kHz, then the minimum sampling rate according to the Nyquist theorem would be 20 kHz. In practice, due to the noisy nature of these data, it is generally advised to allocate spectral “overhead” to this calculation which helps to guarantee that the signal of interest will be recorded faithfully; though higher sampling rates require greater data storage and an analog-to-digital interface capable of handling these rates. Data storage is relatively inexpensive, and acquisition systems’ capabilities are growing, so sampling rates of 30 to 50 kHz are commonly employed.

If one wishes to test hypotheses about single- or multiunit spiking activity, then the typical 300 Hz to 10 kHz band will be appropriate. If one is testing hypotheses involving lower frequency LFPs (approximately 0.5–600 Hz), the neural recordings must be filtered appropriately with a very low high pass band stop (~ 0.1 Hz) and a low pass band stop of at least 1200 Hz to capture the highest frequency signal (2 × 600 Hz). Alternatively, if filtered data is not needed “online” as it is being acquired, data can be saved in its “raw” form, with the bandpass filter characteristics set to be the most permissive, for offline filtering as needed.

There are sometimes options available for online spike detection. Though they may be useful for rapid online analysis, in principle there are no benefits to online-only spike detection if not immediately required for closed-loop control or feedback. Saving raw data and performing offline spike sorting is preferable, because spike sorting can be performed in a more systematic manner without the limitations of the often-busy surgical environment.

Fig. 16.1 Synchronization of behavioral and neural data. Intraoperative data recorded from ventral intermediate nucleus of the thalamus (LFPs) and somatomotor cortices (ECoG). Examples of 3.5 seconds of composite three-axis accelerometer and three-axis gyroscope (green), task joystick (black), 8 channels of ECoG (blue, from anterior to posterior), and 3 channels of depth electrodes (orange/red, from anterior to posterior). Macro tip is 3 mm above the micro. Electrodes are 2 mm from each other. LFP on both, also single unit responses on micro. Movement starts at the black line.

Signal amplifiers and acquisition systems should be selected to suit current and anticipated future requirements. The role of the signal amplifier is to faithfully capture very small, noisy neural signals with high fidelity, while the acquisition system must be able to write multiple channels of data rapidly with no loss. The number of channels also depends on the specific clinical and research aims. For a minimal system, the number of channels might equal the number of microelectrodes implanted for recording, but it is more likely that a research system requires additional analog and digital inputs. Having a primary, clinical system that can serve as a data hub for other recording streams is convenient, as this will implicitly synchronize any data streams for which it is responsible (see later). These additional channels often come in a wide range of connector types and can be sampled at widely varying sampling rates, often with both an upper limit on the maximal signal amplitude or a bound on the amplitude resolution.

As an example, accelerometers can be placed on patient limbs to assess movements, and these signals can be sent to the amplifier and acquisition systems. But because relevant limb movements are biomechanically limited in speed and frequency, 10 kHz or greater sampling is potentially superfluous. Therefore, these inputs should be software-limited to an appropriate sampling rate that accounts for the trade-offs with storage mentioned previously. At our site, we routinely record accelerometer activity at 1000 to 3000 Hz, which results in manageably small data sizes but faithfully captures the fine details of movement.30

Routinely, the data recorded for three full-bandwidth channels of microelectrode data, three lower rate field recordings, and eight analog channels at lower sampling rates results in approximately 10 GB of data for 2 to 3 hours of recording. Including additional high-bandwidth channels, such as ECoG, can triple these data sizes for each case. Neurophysiology systems capable of recording high-bandwidth data should be capable of rapid transfer of these data to an external device for offline analysis.

Another important consideration in selecting a neurophysiology monitoring system is the software. Though hardware specifications may seem to be appropriate, it is the software that provides the interface that will be critical for both providing quality patient care as well as efficiency in processing the recorded data. Equipment and interfaces approved by a country’s health and safety regulatory commissions may impose restrictions on how frequently software is updated, despite a company’s best intentions, so the shipping product must be free of major issues. Good commercial vendors who appreciate the importance of intraoperative research and are committed to supporting it will work to mitigate issues with their hardware and software as they are identified. Wherever possible, open-source and cross-platform data formats and software tools are preferable to closed formats, as this will ensure longevity in data archival and future access.

16.5 Behavioral Task Control

In most cases, research involving human intracranial recordings requires quantitatively rigorous behavioral and precisely timestamped metrics for correlation with neural signals. A simple accelerometer attached to a patient’s wrist may be sufficient for some questions about the relationship of movement to neural activity, but for other behavioral activity, more interesting questions addressing complex motor behaviors and cognition will likely require a dedicated behavioral task control system. For example, our system uses a portable case with rack mounted hardware to house a standard desktop computer, a digital acquisition system used for behavior that is different from the neurophysiological system, and a multi-monitor mount. This system includes a monitor that can be positioned in front of the patient, as well as a joystick that controls the tasks. We present visual tasks to the patients while they manipulate a joystick or button box to provide behavioral responses. For other types of tasks, haptic gloves or other unique manipulanda might be employed for patient interaction. Irrespective of the input device selected for the tracking of behavioral data, patient comfort and reproducible placement are crucial to capturing performance accurately and reliably.

In our laboratory, tasks are programmed in MonkeyLogic, a free, MATLAB-based software toolbox 31,32,33 that enables millisecond precision in our psychophysical experiments (Monkey-Logic is currently supported and maintained at the NIH: https://www.nimh.nih.gov/labs-at-nimh/research-areas/clinics-and-labs/ln/shn/monkeylogic). Importantly, this software also sends precisely timed digital event codes to the neurophysiology acquisition system, enabling synchronization between the two systems. The goal of behavioral–neural synchronization in neurophysiology is to be accurate to ~1 millisecond timescale; in contrast, synchronization between behavior and slower modalities, such as fMRI, is often performed manually (the experimenter simultaneously initiates both systems by striking a key on each system, one with a finger of each hand).

16.6 Data Analysis

Creating a robust data processing and analysis pipeline is critical for an efficient and reliable research workflow. Even though most analyses can be performed post hoc and not online in the operating room, the acquisition system’s hardware and data format serve as the starting point. When dealing with separate systems that are synchronized, custom software is often necessary to align the data according to the synchronizing signal.

Modern neuroscientific data analysis generally falls into two categories: continuous and point process. Continuous data consists of any time series, such as neurophysiological field potentials or accelerometer output. Point process data consists of discrete events, such as spiking activity or activity counts (e.g., number of choices A versus B). Specific methods exist for each class of data, though it is often necessary or desirable to convert between the two data types. Several neural data-specific guides are available that are balanced in presenting both theory and practical implementation.34,35

One of the most common and critical preprocessing steps in neurophysiology is spike sorting, which takes a continuous time series recording as input and converts it to a set of events that are labeled into one or many single “units.” In general, spike sorting is a procedure to isolate an individual neuron’s spikes from other neurons’ spikes in an MER. Typically, MERs sampled at a high rate³ (30 kHz) are bandpassed (approximately 0.3–10 kHz), resulting in a zero-mean noise baseline. A threshold is calculated based on the noise distribution and spike waveforms are isolated as threshold crossings. These waveforms are then analyzed using automated or semiautomated methods, such as principle components analysis and clustering algorithms (e.g., k-means algorithm). Manual methods that categorize waveforms on the basis of waveform features, fully automated methods, or a hybrid of approaches are commonly used, but knowing the ground truth is difficult, so accurate spike sorting remains an active area of research.³⁶ Both open-source and commercial solutions exist to perform spike sorting.

Even within a data type, different techniques that are commonly employed can lead to different qualitative and quantitative interpretations of neural activity. In Fig. 16.2, 2 seconds of ECoG data recorded from human somatomotor cortex of a patient with essential tremor are shown with different spectral techniques demonstrating different results. The choice of analysis can make a substantial difference in the interpretation of the results. Fig. 16.2a shows the time series, referenced and zscored. There was clear oscillatory activity that occurred at different times in the epoch, but the precise frequency characteristics need to be quantified. Fig. 16.2b shows a discrete Fourier transformed (DFT) power spectrum in which two distinct peaks were seen at 1.5 and 22 Hz. The DFT analysis assumed that the data were constant within the analysis window—a property called stationarity—which may be a poor assumption here, considering that different oscillatory activity was variable within this epoch.

Multiple methods are available for investigating time-varying spectral features. The most common is the short-time Fourier transform (STFT), which is also commonly referred to as a spectrogram ( Fig. 16.2c). In the STFT, small segments of time (in this case 0.5 s) were analyzed, and the window was slid across at short intervals (0.025 s) to provide a time-varying estimate of the activity. As the frequency interval is inversely proportional to the amount of time in the analyzed window, shorter time windows result in a larger frequency interval or poorer frequency resolution. The trade-off of frequency resolution, amount of data, and stationarity of data should be considered when using the STFT. Here, both low- and higher-frequency activities were seen, but the activity around 25 Hz was mostly limited near 1.2 s. Furthermore, the first estimate of data is centered around 0.25 s, and the last sample was centered around 0.75 s, and no estimates were available outside of those, meaning the data of interest needed to be within the boundaries set by the temporal window parameters. The timing information seen in the STFT was lost with a power spectrum ( Fig. 16.2b).

Wavelet-based time-varying spectral analyses are also commonly employed. These can provide estimates for an entire short window but also have their own shortcomings. In Fig. 16.2d, the power was calculated from a family of Morlet wavelets convolved with the time series. This method showed a consistent result with the STFT for the higher-frequency activity and its timing, but the lower- frequency activity was not clearly captured.

Finally, in Fig. 16.2e, a Hilbert transform spectral method was applied in a similar manner to the Morlet wavelets. This method captured the activity around 22 Hz well along with the lower-frequency activity which appeared to be primarily isolated to the first 0.5 s of the data. The power spectrum ( Fig. 16.2b) picked up this activity but not the timing, while the STFT ( Fig. 16.2c) picked up this activity in its first estimate, though it was difficult to see represented in the figure, and the Morlet ( Fig. 16.2d) had these spectral features washed out by the much higher powered activity around 25 Hz. In general, the choice of spectral technique comes down to empirical questions about one’s hypothesis and represents a trade-off between temporal precision and frequency precision. This example illustrates the necessity for these analyses to be selected on the basis of the hypothesis prior to analysis of one’s data.