Closed-Loop Stimulation Methods: Current Practice and Future Promise

7 Closed-Loop Stimulation Methods: Current Practice and Future Promise

Vivek P. Buch, Andrew I. Yang, Timothy H. Lucas, H. Isaac Chen

Abstract

Deep brain stimulation and other neuromodulatory therapies have traditionally been performed in the absence of real-time feedback and with constant stimulation parameters (open-loop stimulation). Although this approach has been quite successful, the clinical outcomes are now approaching a plateau. In the development of the next generation of neuromodulatory therapies, an emerging concept showing significant promise is closed-loop stimulation or adaptive neuromodulation in which a real-time feedback signal triggers or modifies stimulation. Theoretically, closed-loop strategies are superior to conventional open-loop stimulation for several reasons, including a wider therapeutic window, improved efficacy, and prolonged battery life. According to control theory, three parts comprise closed-loop stimulation systems: a feedback signal, a module that extracts features from the signal and interprets them, and a stimulation paradigm. Myriad options exist for each of these components, which lead to both the complexity and exciting possibilities of closed-loop strategies. The objective of this chapter is to provide an overview of these options, as well as the commercially available closed-loop systems and limited clinical data that has been accrued thus far. Then follows a discussion of the scientific and technological challenges that need to be overcome for broad clinical adoption of closed-loop neuromodulation. An understanding of these principles will be crucial for practitioners of neuromodulation to partake in the process of designing closed-loop systems and realizing their full potential.

Keywords: Activa PC + S, adaptive DBS, beta oscillation, closed-loop stimulation, feedback signal, NeuroPace, responsive neuromodulation

7.1 Introduction

Current deep brain stimulation (DBS) technologies are built upon simplistic circuits and control algorithms that deliver stimulation constitutively and require an iterative process of adjusting parameters to optimize clinical efficacy.1,2 This “open-loop” stimulation paradigm lacks internal feedback. Recently, more intuitive “closed-loop” or “adaptive” paradigms have been developed which integrate real-time physiological feedback signals ( Fig. 7.1). These techniques promise to improve outcomes in patients with existing indications for neuromodulation, such as movement disorders,^3,4,5,6,7 and expand indications for additional disease states.⁸

There are several putative advantages of adaptive DBS (aDBS) over existing technologies. Closed-loop modulation of stimulation parameters could improve symptom control in conditions characterized by frequent symptom fluctuations, as in Parkinson’s disease (PD)⁹ or essential tremor (ET). In diseases characterized by episodic symptoms, such as seizure disorders, closed modulation would intervene at the point of seizure onset. In addition, aDBS would decrease the on-stimulation intervals by eliminating current delivery when symptoms are absent (e.g., during sleep), thereby extending battery life. Further, feedback control would modulate gain functions to reduce unwanted stimulation-induced side effects. Reducing unnecessary stimulation may reduce off-target effects that result from maladaptive plasticity.¹⁰ Finally, closed-loop modulation would reduce the number of visits to physician office and parameter setting changes, which in turn could lower healthcare delivery costs.

With these advantages come some disadvantages. The engineering trade-off of closed-loop systems is that they must be designed for very specific applications. The optimal control strategy for PD will not likely be optimal for dystonia, epilepsy, or ET. Thus, the number of independent systems will grow. Also, as the control mechanisms become increasingly sophisticated, the internal mechanisms and functionality will be less apparent to the surgeon end-user. Similar design evolution has been observed in car engines. While the standard engines developed in the 1970s and 1980s employed common design features that permitted the neighborhood mechanic access for routine repairs, the highly computerized and customized engines manufactured nowadays require specialized technicians and complex diagnostic equipment for basic repairs. It is thus incumbent on functional neurosurgeons and other clinicians involved in neuromodulation to understand the fundamentals of aDBS. This chapter reviews these principles in the context of emerging systems and their intended disease indications, and reviews future direction on our immediate horizon.

Fig. 7.1 Schematic representation of neuromodulation strategies. Color of arrows indicates signal type (black = input, green = activating signal, red = deactivating signal). Size of arrow indicates relative importance.

7.2 Approaches to Closed-Loop Neuromodulation

7.2.1 Considerations for Designing an Optimal System

The governing principles of closed-loop systems are derived from a field of engineering known as control theory that was first formally conceptualized in the mid 1800s by physicist James Clerk Maxwell 11 and was further developed over the next century.^12,13,14 Broadly speaking, feedback control theory dictates there should be three linked components: a controller, a system, and a sensor. The controller exerts an effect on the system known as the control action, which produces a measurable system output known as the process variable (PV). The PV, measured by the sensor, then informs the controller in the form of a feedback signal. The controller will then compare the PV to the programmed system set point (SP) and calculate an error measurement (PV − SP). Based on this error measurement, the controller will modulate its control action on the system ( Fig. 7.2). One of the most common examples of applied control theory is centralized heating. The controller is the heater unit, the system is the room temperature, and the sensor is the thermostat. The heater is turned on or off (control action) based on the thermostat reading (PV) compared to the desired temperature (SP).

Closed-loop neuromodulation applies control theory to translational devices. In control theory parlance, systems stimulate the nervous system (control action) in response to physiological signals (PV) relative to a desired physiological state (SP). Understanding control theory as it is applied to neuromodulation devices affords neurosurgeons and engineers a common language that facilitates communication and accelerates the pace of translation.

Several fundamental considerations influence the conversation ( Table 7.1). First, each disease condition requires a specific design solution. For example, intractable epilepsy requires a device with sensitive, customized algorithms for detecting seizure onsets within a narrow temporal window. Rapid seizure detection, in turn, must trigger recurrent stimulation to effectively suppress seizure propagation before seizure spread. By contrast, the disease state of ET is characterized by much slower time-scale oscillations and may consequently only require basic phase-amplitude algorithms averaged over long time sweeps to trigger suppressive stimulation. Thus, the device functionality must be specific to the disease state. Beyond disease specificity, the device must be tunable within a dynamic range to match patient-specific disease characteristics. This is because there is variability in physiological control signals within disease states. Returning to the epilepsy example, patients may have multiple seizure types and seizure onset detection may evolve over time. Accordingly, devices must have the capacity to be both disease-and patient-specific.

Second, there are engineering trade-offs between size, weight, power, and cost. Termed SWAP-C in aerospace engineering, the design trade-offs are equally applicable to medical device development. Increasingly, devices are designed for low-power architecture, compact form factor, and minimal weight.^15,16,17 When operationalized, these trade-offs mean that systems often sacrifice circuit complexity—and hence algorithm flexibility—for lower power consumption requirements and reduced device size. Reduced complexity leads to lower unit production costs and lower barriers to entry into the competitive device market. Of course, lower complexity means less programming flexibility. Neurosurgeons and engineers must work together to maximize idealized device features within design constraints.

Table 7.1 Criteria for designing closed-loop systems

Criteria	Design parameter considerations
Indication	Disease-specific physiology dictates ideal control algorithm and stimulation paradigm
Complexity	Increasingly complex parameters lead to increased sophistication but decreased feasibility
Clinical acceptability	Characteristics of electrodes, battery, and interface affect ergonomics of daily use and longevity
Stage of technology	Proven effectiveness of existing technology must be weighed against potential benefits of new technology

Fig. 7.2 Box diagram of general feedback control theory.

Third, system design elements must be acceptable to patients who have limited capacity to interface with the system directly. Electrodes, chassis, connector cables, and external peripherals must operate with little or no disruption in normal life activities of the patients. Increasing invasiveness of a device can become a barrier to patient adoption. For instance, in our experience, many patients refuse to undergo placement of Neuro-Pace® systems because of the requirement of a craniectomy and the limitation on future brain MRI. Devices which require battery replacement obligate patients to undergo multiple surgeries, a further barrier to wide adoption. Vagal nerve stimulators can produce painful cervical sensations or repeated urges to swallow that patients find dissatisfying. These and other factors diminish device adoption and must be considered in the early phase of development.

To provide further insight into possible approaches for closed-loop neuromodulation, the subsequent sections will summarize sources of feedback, stimulation strategies and control algorithms, and the small but growing literature on clinical outcomes with aDBS.

7.2.2 Sources of Feedback Signals

The main feature of aDBS is real-time modulation driven by physiological signals. Many signal sources have been considered ( Fig. 7.3). Current versions of aDBS rely upon a single feedback signal, but future devices will likely utilize multiple signals in parallel.

Single- and multi-unit activity

Extracellular action potentials are a robust source of discrete physiological data. Because individual neurons may be tuned to specific behavior features, such as the direction of extremity motion¹⁸ and the orientation of visual stimuli,¹⁹ devices may be highly precise in signal input. This type of information drives a variety of human brain–computer interfaces (BCIs), including systems for controlling neuroprosthetics,²⁰ graphical user interfaces,²¹ and functional electrical stimulation.^22,23

Extracellular action potential recordings require high electrical impedance. Therefore, multielectrode arrays, such as the Utah array (Blackrock Microsystems, LLC, Salt Lake City, UT) that consists of a 10 × 10 array of electrodes, are used. Besides its use in BCIs, the Utah array has also been used to study the spread of the ictal wavefront in seizures.²⁴ Michigan probes and microwire constructions are other methods of chronically recording neurons. Future arrays, such as those being developed by MIT’s Media Lab, promise thousands of contacts for chronic recording. As the number of input channels increases, so too must the complexity of the circuitry and computational processing to analyze this data.

Despite the advantage of highly discrete input signals afforded with single- or multi-unit activity, certain drawbacks limit their use as feedback signals. The technology available for recording this activity in patients allows only small areas of the brain to be sampled. Also, these arrays are designed for the cortical surface. There are few solutions for subcortical targets. Single-unit recording quality is also finite, owing to the inherent mismatch in material properties between the brain and electrodes.²⁵ Small shifts in electrode position, build-up of impedance over time, or loss of individual neurons due to damage results in unstable source signals. Thus, chronic applications require simplifying decoding algorithms²⁶ or frequent recalibration²¹ to maintain performance over time. These challenges limit long-term durability of penetrating arrays and may prevent their widespread adoption for translational indications.

Local field potentials

Local field potentials (LFPs) represent the integrated analog synaptic inputs into cortical layers rather than the discrete spike activity of those neurons (i.e., postsynaptic activity).²⁷ This oscillatory activity is classified by dominant frequency ranges ( Table 7.2) and it plays a role in brain functions like memory formation²⁸ and temporal binding of neural activity.²⁹ Interactions exist between LFPs across frequency bands (e.g., theta-gamma coupling in the coding of multi-item messages³⁰) and between LFPs and the firing of individual neurons that may contribute to disease states such as PD.

With existing technology, LFPs are thought to be a reliable source for signal detection.³¹ LFP recordings benefit from the fact that they can be recorded with low-impedance electrodes with larger contact surface areas. These electrodes cause less local tissue damage. Accordingly, LFPs are less sensitive to changes in electrode impedance caused by tissue destruction and gliosis. The relaxation of the electrode size constraint enables large sheets of electrode arrays to be deployed over the brain surface. Moreover, LFPs can be reliably recorded from concentric ring electrodes, as in the case of depth electrode contacts. Thus, LFPs can be recorded from deep structures safely for long durations. This relaxed electrode design feature makes it possible to record and stimulate from the same lead,³² which is a major advantage for neuromodulatory devices. Highly flexible graphene arrays further improve tissue–electrode compliance matching and demonstrate superior performance over millions of duty cycles of stimulation and recording.³³ Penetrating, high-impedance electrodes use sharp tips that erode and breakdown in the presence of electrolytic charge buildup during stimulation, and consequently have limited utility in chronic stimulation paradigms.

Fig. 7.3 Types of currently used feedback signals and corresponding systems.

Table 7.2 Types of local field potentials

Oscillation	Frequency (Hz)
Delta	1–4
Theta	4–10
Alpha	10–14
Beta	14–36
Low gamma	36–70
High gamma	> 70

The accessibility of LFP data in humans has elucidated a number of biomarkers for closed-loop neuromodulation. Beta oscillations in PD are a prominent example. Though the causal significance of beta oscillations in PD pathology remains unclear, their strong association with PD symptoms makes them a potentially useful feedback signal for closed-loops systems. While neural activity within the globus pallidus of healthy rhesus monkeys is not synchronized, treatment with 1-methyl-4-phenyl-1,2,3,6-tetrahydropyridine (MPTP) induces periodic oscillations in a significant fraction of neurons.34 Similarly, oscillations in the 15 to 30 Hz band are a prominent feature of subthalamic nucleus (STN) recordings in PD patients.^35,36,37 The power of beta oscillations in the basal ganglia is related to symptom severity in PD, as it is increased when dopamine is depleted^38,39,40 and reduced during voluntary movement^39,41,42 and by DBS.^43,44,45 Moreover, low-frequency stimulation (20 Hz) of the STN in patients modestly slows motor activity.⁴⁶ The origin of these beta oscillations is not completely clear, although there is evidence to suggest roles of both the motor cortex^47,48,49 and the STN–globus pallidus pars externa circuit.^50,51

Other LFP relationships are also being explored. Recently, it was demonstrated that DBS reduces coupling between the phase of beta oscillations and broadband amplitude in the motor cortex.⁵² In addition to beta oscillations, dopamine influences the power of low-frequency bands^40,53 and non-beta frequencies may be better correlated with PD symptoms.⁵⁴

LFP signals may be used in novel ways, such as in temporary applications for brain rehabilitation following injury. This concept is made possible by a number of electrode arrays which dissolve over time, as their use is no longer needed. Examples include resorbable silicon electrodes that dissolve after preprogrammed durations and do not require removal.⁵⁵

Peripheral electromyography and inertial recordings

Clinical phenotypes are important manifestations of neurological disease. For movement disorders, real-time information concerning extremity function is a natural feedback source for closed-loop neuromodulation. Surface electromyography^56,57 and multi-axis accelerometers^58,59,60 can monitor tremor amplitude. Such functionality may even be built into smart watches.⁵⁹

As these systems mature to clinical reality, wireless data fidelity and security will become driving factors in design. Environmental electromagnetic noise that is present in everyday life—will influence the utility of these devices and corresponding control algorithms. Cell phones, smart watches, activity monitors, and other sources of wireless signal may confound wireless medical device communication. In the modern era, engineers must design communication systems with the understanding that systems may be hacked (intentionally or unintentionally). Thus, secure, high-fidelity communication protocols must be developed.

Other signals

Although neural activity is most often measured electrically, there are other sources of physiological signals that may be tapped. In PD, symptoms arise due the loss of striatal dopamine from substantia nigra inputs. DBS may increase striatal dopamine release.⁶¹ Thus, Dopamine metabolite concentration could serve as a biomarker of stimulation efficacy. Indeed, microdialysis has been used to assess extracellular neurotransmitter levels during the placement of DBS leads.⁶² Similar to insulin pumps that use blood glucose measures to adjust insulin release, it is not difficult to imagine a system that monitors neurotransmitter levels to modulate stimulation. Such a system could be used in movement disorders or neuropsychiatric conditions, such as refractory depression. Alternatively, carbon fiber electrodes can be used to detect electroactive molecules, such as dopamine, adenosine, and oxygen, in real-time using fast-scanning cyclic voltammetry or amperometry techniques.⁶³ These methods are primarily being tested in animal models.^63,64 Feasibility in a human subject has been demonstrated,⁶⁵ though the long-term viability of this technology remains to be studied.

As reliable biomarkers are identified in other diseases or cognitive states, neuroengineering design principles may be applied. An emerging example is the use of prefrontal cortex signals of volition as control signals in psychiatric disorders.⁶⁶ In obsessive compulsive disorder (OCD), for instance, the volition to induce stimulation could enable a patient to intentionally “will” the system to trigger stimulations when obsessive thoughts are particularly intrusive.⁶⁷

7.2.3 Control Systems and Stimulation Paradigms

Control systems depend on the disease- or state-specific bio-markers. In epilepsy, control systems are based on seizure detection algorithms.68,69,70,71 In PD, beta oscillatory features are the candidate biomarkers. Specifically, amplitude and phase appear to be salient features. Amplitude-modulated approaches trigger stimulation when the amplitude of beta oscillations reaches a certain threshold. In contrast, phase-modulated approaches trigger stimulations at a particular phase of the oscillation that can either attenuate or potentiate the oscillation.⁷²

Table 7.3 Stimulation paradigms in closed-loop systems

Paradigm	Closed-loop triggered stimulation
Binary	Predetermined stimulation parameters “on” or “off”
Graded	Changes in shape, intensity, frequency, pulse width, or location
Coordinated reset	Short bursts of high-frequency pulse trains
Hybrid	Coincident detection, phase-amplitude coupling, spike-phase coupling, central-peripheral biomarker pairing

Once a suitable feature has been identified, a number of stimulation paradigms are possible ( Table 7.3). The simplest strategy is binary modulation, in which preset stimulation parameters are triggered when feature criteria are satisfied. Cardiac sensitivity with Model 106 Aspire VNS (LivaNova PLC, London, United Kingdom previously Cyberonics, Inc., Houston, TX) is an example. When a rapid increase in heart rate is detected, the VNS runs a stimulation routine. Step-wise or graded responses are an alternative strategy. In this scenario, one or more stimulation parameters are modulated once feature criteria are met.³¹ Paradigms can implement numerous steps to create near-continuous changes in stimulation intensity as driven by the feedback signal.⁶ The specific method for defining the relationship between feedback and stimulation will likely rely upon different aspects of control theory.^5,6,7 Stimulation variables in this paradigm could include stimulation amplitude steps, pulse train, frequency, pulse width, electrode channel, and more complex combinations of parameters. Another paradigm is the coordinated reset. This method desynchronizes network activity and inhibits negative plasticity effects through short bursts of high-frequency pulse trains.^73,74,75 Finally, hybrid paradigms use complex feedback signals, such as coincident detectors, spike-phase, and central-peripheral detectors. These use input schemes that integrate information from multiple sources simultaneously. A hypothetical example of such a system could include one that detects beta oscillations in motor cortex and peripheral tremor oscillations from a wearable accelerometer to modulate PD tremor.