Multitarget Neurostimulation of the Deep Brain: Clinical Opportunities, Challenges, and Emerging Technologies

Abstract

Recent computational, pre-clinical, and clinical studies have demonstrated the potential for using neuromodulation through simultaneous targeting of multiple deep brain regions. This approach has already been used by therapeutic and systems neuroscience applications. However, the broad clinical adoption of invasive distributed deep brain interfaces remains in its early stages. This review explores the barriers to implementation by addressing three key questions: Do the benefits of implanting multiple electrodes justify the associated risks for specific applications? What is the risk-benefit ratio, and what technological advancements will be necessary to encourage clinical adoption? We also examine next-generation technologies that could enable distributed brain interfaces, including system-on-chip micro-stimulators as well as nanoparticles. We highlight the role of novel machine learning algorithms in the optimization of stimulation parameters and for the guidance of device placement. Emerging hardware accelerators equipped with on-chip AI have demonstrated functionality that can be used to decode and to classify distributed neuronal data. This advance in hardware accelerators has also contributed to the potential for enhanced closed-loop stimulation control of devices. Despite these advances, significant technological and translational barriers persist, limiting the widespread clinical application of distributed brain interfaces. This review provides a critical analysis of recent prototypes and novel hardware for use in distributed systems. We will discuss both clinical and research applications. We will focus and highlight the utilization of multi-site technologies to meet the needs of neurological diseases. We conclude that there exists a critical need for further innovation and integration of multi-site technologies into clinical practice.

1. Introduction

Neurostimulation of deep brain structures that are designed to target one or two different brain regions has already grown in clinical interest. Brain signal recordings increasingly point to network-level, rather than isolated single site, dysfunction as the origin of most diseases. Multi-target technologies, an emerging class of therapeutic neurostimulator devices, allow for concurrent stimulation of multiple and distinct anatomical brain region targets. Multi-target and high-density stimulation and recording of the brain cortex have been achieved by use of microelectrode arrays, yet no comparable technology exists or alternatively has been used for humans on deep brain structures. The increase of deep brain neurostimulation locations aims to optimally interface with and modulate distributed circuits, with the goal of symptom reduction with fewer side effects. Therefore, we present a cost-benefit analysis that reviews early clinical works of multi-target deep brain neuromodulation to weight anticipated benefits with known risks. We review clinical applications of dual target (2 locations) and bilateral dual target (4 locations) neurostimulation. Additionally, we consider translational and clinical readiness of next-generation devices which allow stimulation of five or more locations.

In the following text, “multi-target” refers to neurostimulation which targets two or more anatomical brain region structures. Similar yet distinct terms, “multi-contact” and “distributed” are introduced to contextualize advances in multi-target technologies. We refer to “multi-contact” and “multifocal” stimulation as that which uses multiple electrode contacts to shape electric fields within one target region. “Bilateral stimulation” refers to stimulation of a structure in both hemispheres. The term “distributed stimulation” is used to indicate stimulation fields which are generated at the edge of an anatomical structure and spread to nearby regions.

One theoretical mechanism for neural response to single site deep brain neurostimulation posits that the stimulation acts as a virtual lesion of information, blocking pathological output signals from cells [1]. However, we now know that the mechanism of action of neuromodulation is complex and not just an informational lesion. There are important physiological, neurochemical, neurovascular and neurogenic changes as well as critical network level changes in neural oscillations [2–4]. For diseases with distributed rather than single site origins, clinicians have been compelled to explore emerging multi-target brain interfaces. This text reviews current research which suggests multi-target stimulation can impact brain networks and modulate pathological feedback oscillations, desynchrony, and network-level connections.

To fully develop and implement distributed brain therapeutics, it is essential to clarify neurobiological circuit dysfunction and mechanisms of action. While advances in research have begun to reveal systems-level insights, initiatives like the BRAIN Initiative continue to drive the understanding of neural coding, dynamics, and modulation—critical for linking network perturbations to symptom improvement [5,6]. To make such multi-target neurostimulation systems practical, several factors must be taken into consideration, not only from the hardware technical perspective, but also for patient safety. Furthermore, hardware barriers aside, clinical teams will require tools to determine how to optimally implement devices, including the precise locations and parameters for acute and chronic treatment. Machine learning, edge computing, and closed loop systems embedded on-chip may aid clinical decision making.

Clinicians who program neurostimulation devices have historically endorsed technologies that offer greater control over intracerebral electric fields (e.g. current steering electrodes [7,8]) and thus would likely support the adoption of multi-target brain stimulation interfaces if they provided a more targeted and adaptable treatment with fewer side effects and are equally or safer and more durable. Future research will be necessary to elucidate these possibilities. However, a recent review by Miller and colleagues acknowledges a “chicken-and-egg problem” where novel brain interface technologies necessitate justification for prototype development but appropriate support for such investments remains difficult in the absence of existing technology [9]. Additionally, several recent papers have discussed technical barriers to distributed neural interfaces [10–14]. One of the goals of this review is to examine emerging research on multi-target deep brain neurostimulation and to address this 'chicken-and-egg' problem by evaluating the clinical and technical literature in the context of disease-specific cost-benefit considerations. Section 2 explores network dysfunction theories and early multi-electrode DBS studies across neuropsychiatric and neurological diseases. Section 3 outlines clinical barriers that have shaped emerging technologies which differ from traditional DBS hardware (Section 4), and Section 5 discusses computational approaches needed to guide multi-target device implantation location and parameter optimization.

Methods: Clinical studies were collected from PubMed per disease application, with the requirement of distinct and multiple deep brain neurostimulation targets. Several multi-contact papers are indicated as such and included for context and differentiation between approaches. Device papers were collected from IEEE Databases and references, per category: system-on-chip, milli-meter scale, and magnetoelectric nanoparticle, and for three non-invasive modalities. Devices not designed for neural applications were excluded.

2. Disease Specific Applications of Multi-Site Neurostimulation

Nearly all the multi-target clinical works are small case studies, and we note a lack of powered and randomized controlled trials. Currently, multi-target experiments represent off-label use, often as a “rescue” therapy when existing stimulation lacks efficacy or for uncommon symptom profiles. While regulatory and legal aspects clearly play a role, the focus of this work is to review benefits of multi-target stimulation identified in case studies, which use modern macroelectrode systems; acknowledge and quantify the risks of the multi-target approach with existing hardware; and to review non-macroelectrode devices, and necessary hardware and software, to implement a truly multi-target system, directed at academic works with far-term translational readiness while acknowledging more mid- and near- term solutions.

Neuropsychiatric and neurological diseases have a large contribution from dysfunctional connectivity and synchronization among multiple brain regions (Figure 1, Table 1) [15–17]. However, these global trends may conceal circuit and symptom specific changes which may be associated with differing levels of hypo- or hyper-synchronization. Several theories of disease will be introduced which suggest maladaptive activity arises from global network dysfunction, distributed hubs, synchronization abnormalities, and feedback loops, all of which may effectively be targeted by distributed multi-target stimulation systems.

Figure 1. Brain Stimulation Targets for Several Neuropsychiatric Disorders and Neurological Diseases.

Deep brain structures targeted for neurostimulation are annotated. Disease applications are noted underneath each target. The following brain disorders are considered: addiction and use disorders (AUD), Alzheimer’s disease (AD), drug resistant epilepsy (DRE), dystonia (DYT), essential tremor (ET), obsessive compulsive disorder (OCD), Multiple Sclerosis Tremor (MST), Parkinson’s disease (PD), schizophrenia (SCZ), and Tourette’s syndrome (TS). United States Food and Drug Administration (FDA) approved Deep Brain Stimulation (DBS) targets are marked with an asterisk (*); targets awarded a Humanitarian Device Exemption (HDE) are marked with a double asterisk (**). Figure created with Figma and BioRender [289].

Table 1.

Diseases and Rationale for Multi-Site Stimulation

Disease	Rationale for Multi-Site Stimulation
Alzheimer’s Disease	Global changes across a wide neural network involving cholinergic dysfunction and other cells and pathways
Chronic Pain	Separate sensory, affective, and modulatory components
Depression	Distinct circuits for positive and negative affect
Dystonia	Distributed motor circuit dysfunction across multiple brain regions
Epilepsy	Network disorder with multifocal seizure zones
Essential Tremor	Subtypes with different circuit involvement
Multiple Sclerosis Tremor	Involvement of both cerebello-thalamo-cortical and pallidal circuits
Obsessive-Compulsive Disorder	Symptom-specific circuits for compulsion, depression, anxiety
Parkinson’s Disease	Distinct symptom circuits unaddressed by single target (e.g., gait, speech)
Schizophrenia	Network-wide disconnection affecting symptom dimensions
Substance Use Disorder	Distributed hedonic hotspots in reward circuits
Tourette’s Syndrome	CSTC loop and supplementary circuit dysfunction

Deep brain stimulation (DBS) and responsive neurostimulation (RNS) have received FDA approval for certain conditions, humanitarian device exemptions for others, and are under active investigation for additional indications. This review does not aim to cover all diseases considered for neurostimulation; however it will focus on those with well-established network-based models and documented exploration of multi-target approaches; thus, any omission should not be taken to imply incompatibility with multi-target based approaches.

One approach to targeting multiple sites for neurostimulation includes custom trajectories with electrode leads that lie in more than one structure, referred to as “multi-contact”, “one pass”, “blended”, or “combination” stimulation. The multi-contact method has been explored for dystonia [18], ET [19–21], tremor-dominant (TD) PD [22,23], and freezing of gait (FOG) PD [24–28]. Such methods were developed for TD and FOG PD subtypes, that often respond poorly to single-target DBS due to involvement of fiber tracts and regions not addressed by conventional approaches [29]. Compared to single-target stimulation, multi-contact DBS has shown preliminary evidence of clinical benefits and improved EEG activity patterns [28]. The use of a single lead for multiple targets retains parameter flexibility through methods such as interleaved deep brain stimulation (iDBS), where impulses are alternately delivered on two different electrode contacts [24,26,28]; and multiple independent current control (MICC), which enables distinct stimulation parameters across contacts on the same lead [27]. Alternatively, advances in IPG technology and surgical planning software now allow multi-target stimulation with separate electrodes for distinct targets, the focus of the following section (Tables 2–5). An example of the multi-target approach is presented in Figure 2, a reconstruction of dual-target DBS leads overlayed on a magnetic resonance image (MRI) [30].

Table 2.

Parkinson’s Disease Clinical Studies Employing Two (or More) Targets.

Targets		Summary of Results	# of Subjects	Leads	IPG	Frequency (Hz), (Min, Max)	Pulse Width (μSec), (Min, Max)	Voltage (V), (Min, Max)	Current (mA), (Min, Max)	Publication Identifier
GPi	VIM	Tremor score reduction 90.6% (GPI+VIM) vs 21.8% (GPI only)	13							[42]
GPi	STN	UPDRS-IV score 6 (pre-surgery) vs 2 (24 post surgery)	1	PINS G102R		90, 170	70, 90	2, 3.5		[289]
GPi	STN	UPDRS-III score 33 (GPi+STN)	2	Medtronic, Model 3387		130, 180	60, 90	2.8, 4.4		[32]
GPi	STN	Larger reduction in beta coherence in dual (GPi+STN) vs single target	1	Medtronic, Model 3387, B3301542	Medtronic Percept PC	125	60		2.4 (STN), 2.2 (GP)	[34]
GPi	STN	UPDRS-III score 25 (STN) vs 10 (STN+GPi)	1			130	120 (L STN), 90 (R STN), 60 (L&R Gpi)	4.0 (L STN), 4.3 (R STN), 1.5 (L Gpi), 2.3 (R Gpi)		[290]
GPi	STN	UPDRS-III score lower with dual (SPi+STN) vs single (GPi only) p<0.05	13	Medtronic, Model 3387, 3389	Medtronic Kinetra, Soltera 7426, 7498 Dual connector					[291]
GPi	CM/Pf	UPDRS-III mean score 31.2 (CM/Pf+GPi) vs 40.2 (CM/Pf only) vs 36.4 (GPi only)	6	Medtronic, Model 3387, 3389	Medtronic Kinetra					[36]
GPi	STN	Total UPDRS-III mean score 31.8 (STN+GPi) vs 41 (GPi only) vs 37.2 (STN only)	3	Medtronic, Model 3387, 3389	Medtronic Summit RC + S	125	60, 90		1.1, 3.1	[43]
STN	VIM	UPDRS-III score 31 (VIM only) vs 28 (VIM+STN)	1	Medtronic, Model 3387, 3389	Medtronic Activa PC	200	60, 90	1.6, 2.6		[292]
GPi	STN	UPDRS-III dual 6.8 point improvement vs GP only (p<0.001), dual 3.1 point improvement vs STN only (p=0.021)	6		Medtronic Summit RC + S, Medtronic Y-lead extension					[31]
PPN	STN	UPDRS-III mean reduction 33% (PPN only) vs 54% (STN only) vs 56% (STN+PPN)	6	Medtronic, 3389	Medtronic Kinetra	10, 80 (PPN), 130–185 (STN)	60 (PPN), 90 (STN)	1.5–2 (PPN), 1.5–2.4 (STN)		[37]
STN	CM/Pf	UPDRS-III mean score 38.5 (CM/Pf+GPi) vs 41.4 (GPi only) vs 46.2 (CM/Pf only)	2	Medtronic, 3389	Medtronic Kinetra	185 (Gpi, CM, Pf), 135 (STN)	210 (GPI), 60–90 (STN), 90 (CM, Pf)	3–4 (GPI), 1.8–3.2 (STN), 1.5–2.5 (CM, Pf)		[35]
GPi	CM/Pf		6
VO	VIM	UPDRS-III items 20/21/22 score 4/3/8 (baseline) vs 1/1/3 (multi-contact)	1	Medtronic, 3389	Medtronic Soltera 7426	160	90	2.1		[293]
GPi	STN	UPDRS-III reduction from baseline 43% (GPi+STN)	8	Medtronic, Model 3387, 3389; PINS L302, 302						[33]
PPN	cZi	UPDRS-III reduction from baseline 18.8% (PPN only) vs 46.8% (cZi only) vs 47.3% (PPn+cZi)	7	Medtronic, 3389	Medtronic Kinetra	60 (PPN), 60, 120 (cZi)	60	2.4 (PPN), 3.2 (cZi)		[41]
PPN	cZi	UPDRS-III reduction from baseline 21.6% (PPN only) vs 36.9% (cZi only) vs 50.5% (PPN+cZi)	4	Medtronic, 3389		25, 60 (PPN), 60 (cZi)	60			[40]
PPN	GPi	MTT45 7.2 (PPN only) vs 8.1 (GPi only) vs 1.7 (PPN+GPi)	1	Medtronic, Model 3387, 3389		25 (PPN), 130 (GPi)	60 (PPN), 210 (GPi)	2.5 (PPN), 3 (GPi)		[39]
VIM	STN	UPDRS-III reduction from baseline 53% (VIM+STN multi-contact)	5	Medtronic, Model 3387; PINS L302		130, 190	60, 90	1.6, 2.4		[23]
SN	STN	SPT Variability reduced in dual (STN+SN) vs STN only (p=0.031)	12	Medtronic, 3389; Boston Scientific, 2201, 2202		125, 130	60	1, 2 (SN), 1.5–4 (STN)		[28]
PPN	STN	UPDRS-III median score 19 (STN only) vs 18.5 (STN+PPN)	6	Medtronic, 3389		15, 25 (PPN)	60, 90 (PPN)	1.2, 3.8 (PPN)		[294]
SN	STN	HFPS Cycles mean 0.71 (STN only) vs 0.79 (STN+SNr) (p=0.038)	10	Boston Scientific, octopolar electrode	Boston Scientifi, Vercise	130	60			[27]
SN	STN	UPDRS-III score 36 (STN only) vs 29 (STN+SNr)	1	Medtronic, 3389	Medtronic Activa PC	125				[26]
SN	STN	UPDRS-II + UPDRS-III mean combined score 14.25 (STN only) vs 13.42 (STN+SNr) (p=0.470)	12		Medtronic Kinetra, Activa PC	125	60			[25]
SN	STN	Time to Walk 4m improved in interleaved multi-contact STN+SN vs STN (p<0.001)	1			15, 125	80	5.7		[24]
VIM	STN	UPDRS-III reduction from baseline 43.2% (p=0.07) VIM+STN multi-contact	5	Medtronic, 3389; Boston Scientific, octopolar electrode		85, 180	50, 100	1, 5.6	2.5, 7	[22]
STN	SN	No additional benefits on dysphagia	23			125, 130	60	1.5, 4.3		[295]

Caudal Zona Incerta (cZi); Centromedian-Parafascicular complex (CM/Pf); Globus Pallidus pars Interna (GPi); Pedunculopontine Nucleus (PPN); Subthalamic Nucleus (STN); Substantia Nigra (SN); Ventralis Intermedius (VIM); Venralis Oralis (VO)

Heel Strike, Foot Contact, Push Off and Swing Cycles (HFPS Cycles); Minimal Time for Turning 45 degrees (MTT45); Step-to-Step Time (SPT); Unified Parkinson's Disease Rating Scale (UPDRS): Activities of Daily Living (UPDRS-II), Motor Examination (UPDRS-III), Complications of Therapy (UPDRS-IV)

Table 5.

Neuropsychiatric Clinical Studies Employing Two (or More) Targets.

Disease	Summary of Results	Strategy	Target		# of Subjects	Leads	IPG	Frequency (Hz), (Min, Max)	Pulse Width (μSec), (Min, Max)	Voltage (V), (Min, Max)	Current (mA), (Min, Max)	Publication Identifier
substance use disorder	VASC score 7.6 (baseline) vs 0.8 (NAc+ALIC)	Multi-Contact	NAc	ALIC	8	Custom leads	Suzhou Screneray IPG	145, 185	120, 300	1.5, 7		[304]
obsessive compulsive disorder	Y-BOCS mean score 14.17 (VC/VS+STN) vs 17 (VC/VS) vs 19.83 (STN)	Multi-Target	VC/VS	STN	6	Medtronic, Model 3387, 3389	Medtronic Activa PC	130	60	0.7, 2.85 (amSTN) 5.3, 7 (VC/VS)		[84]
	Y-BOCS improvement from baseline 50% (STN+ALIC) vs 41.2% (STN only) vs 31% (ALIC only)	Multi-Target	ALIC	STN	6	Medtronic, Model 3387, 3389						[85]
	Y-BOCS improvement from baseline 13.34% (VC/VS) vs 16.38% (VC/VS+SMA)	Multi-Target	VC/VS	SMA	1	Medtronic, Model 3386, 3391	Medtronic Activa PC + S	15, 135 (VC/VS), 10, 130 (SMA)	90, 150 (VC/VS), 90 (SMA)	2.5, 4.5 (VC/VS), 2, 5.1 (SMA)		[91]
	Y-BOCS score max improvement from baseline 72.5%	Multi-Target	VC/VS	BNST	3	Medtornic, Model 3387 SenSight Directional	Medtronic Summit RC+S, Activa RC, Activa PC, Percept	125, 150	90		2, 5	[92]
epilepsy and obsessive compulsive disorder	Y-BOCS improvement from baseline 37.5%	Multi-Target	STG	NAc	1	Neuropace depth leads + strips	Neuropace RNS programmable neurostimulator	50, 200	40, 120		0.5, 7	[296]
Gilles de la Tourette syndrome and obsessive compulsive disorder	YGTSS score max improvement from baseline 80%; Y-BOCS score max improvement from baseline 56%	Multi-Target	VC/VS	GPi	2		Medtronic Percept PC					[94]
	YGTSS score max improvement from baseline 60%; Y-BOCS score max improvement from baseline 45%	Multi-Target	VC/VS	ALIC / NAc	1	Medtronic, Model 3387	Medtronic Activa PC	130	90	2.5, 5		[93]
depression	Improvement of mood and anxiety	Multi-Target	SCC	VC/VS	1	Boston Scientific, Cartesia		6, 130	50, 180		2, 5	[106]
	HAMD reduction	Multi-Target	SCC	MFB	2	Medtronic, Model 3387		140, 150	60, 90	3.2, 4 (SSC), 3.5 (MFB)		[105]
	HAMD-17 improvement from baseline 57%	Multi-Contact	BNST	NAc	23			130	90	2, 6		[107]
chronic pain	VASP min score 0 (NAC+PVG) vs 3 (NAC only) vs 3 (PVG only)	Multi-Target	PVG	NAc	1	Medtronic, Model 3387		130	300 (PVG), 450 (NAC)	3.5 (PVG), 1 (NAC)		[109]
	NRSP mean improvement from baseline 56% (PAG only) vs 67% (CmPf only) vs 73% (PAG+CmPf)	Multi-Target	PAG	CmPf	3	Medtronic, Model 3387		5, 10 (PAG), 70, 150 (CMPf)	90, 150 (PAG), 60, 90 (CMPf)	1, 5 (PAG), 2, 2.5 (CMPf)		[110]
	Medication reduction achieved	Multi-Contact	CmPf	PAG / PVG	3	Boston Scientific, Vercise, Model DB1110 Generator	Boston Scientific, model DB2201	10 (PAG/PVG), 132 (CMPf)	60, 110 (PAG/PVG), 60, 90 (PAG/PVG)		3.5, 4.5 (PAG/PVG), 4, 4.5 (CMPf)	[108]

Anterior Limb of the Internal Capsule (ALIC); Bed Nucleus of the Stria Terminalis (BNST); Centromedian Intra-Laminar Parafascicular Complex (CMPf); Globus Pallidus Pars Interna (GPi); Medial Forebrain Bundle (MFB); Nucleus Accumbens (NAc); Periaqueductal Grey (PAG); Periventricular Gray Region (PVG); Subcallosal Cingulate (SCC); Supplementary Motor Area (SMA); Superior Temporal Gyrus (STG); Subthalamic Nucleus (STN); Ventral Capsule/Ventral Striatum (VC/VS)

Hamilton Depression Rating Scale (HAMD-17); Numeric Rating Scale for Pain (NRSP); Visual Analogue Scale for Craving (VASC); Visual Analog Scale for Pain (VASP); Yale-Brown Obsessive Compulsive Scale (Y-BOCS); Yale Global Tic Severity Scale (YGTSS)

Figure 2.

Example of Multi-Target Neurostimulation. Dual-lead Deep Brain Stimulation (DBS), with hardware targeting separate anatomical structures, illustrates the multitarget approach. Two DBS lead reconstructions (orange) and targets (purple; the left subcallosal cingulate cortex, LSCC, and left ventral striatum/ ventral capsule, LVC/VS) are over lay on a magnetic resonance neuroimage, sagittal view. Reprinted from [30] based on the IOP Published Content Guidelines.

2.1. Neurological Disorders

Neurostimulation has been well researched and clinically deployed in the context of neurological disorders. However, some diseases display alternative symptom profiles or less common phenotypes which are not adequately addressed with current technologies. Network models of neurological diseases suggest these pathologies arise from dysfunctional connections and hyper- or hypo- synchronization. We begin with a review of movement disorder clinical studies which implement multi-target (Table 2, 3, 4), highlighting a recent study of combined subthalamic nucleus (STN) and globus pallidus (GP) stimulation for Parkinson’s Disease (PD) patients [31].

Table 3.

Epilepsy Clinical Studies Employing Two (or More) Targets.

Targets		Summary of Results	# of Subjects	Leads	IPG	Frequency (Hz), (Min, Max)	Pulse Width (μSec), (Min, Max)	Voltage (V), (Min, Max)	Current (mA), (Min, Max)	Charge Density (μC/cm2)	Publication Identifier
ANT + CM, ANT + MD + CM, ANT + CM + Pulvinar, ANT + MD, ANT + Pulvinar, CM + STN, or ANT + STN		Seizure frequency improvement 45.5% (ANT+CM), 37.5% (ANT+MD)	8	Medtronic, 3389; Boston Scientific, Vercise Cartesia	Medtronic, Percept PC; Boston Scientific, Vercise Genus 32	145	90		0.5, 1		[60]
CM	ANT	Seizure frequency improvement 60% (CM+ANT) vs 56% (CM only) (p=0.583)	11	Medtronic 3387, 3389	Activa, Inbtellis, Restore	2, 100 (CM) 5, 100 (ANT)	60, 150 (CM) 90, 120 (ANT)	1, 6.3 (CM) 2.1, 6.3 (ANT)			[59]
NAC	ANT	No additional effects to seizure frequency or severity	4	Medtronic 3387	Activa-PC	125	90	5			[58]
Hippocampus	Parietal PVNH (1), Temporo-occipital PVNH (2), or Occipital PVNH (1)	Seizures involved both PVNH and non-PVNH locations in 81.6% of cases	4	Neuropace depth leads + strips	Neuropace RNS programmable neurostimulator	125, 200	80, 160		1, 6.5	0.6, 4.1	[57]
Hippocampus	ANT	Seizure frequency improvement 47% (Hippocampus +ANT) vs 50% (Hippocampus only) (p=0.59)	4	Neuropace depth leads + strips	Neuropace RNS programmable neurostimulator						[52]
NAC	STG	Seizure frequency improvement 60%	1	Neuropace depth leads + strips	Neuropace RNS programmable neurostimulator	50, 200	40, 120		0.5, 7		[296]

Anterior Nucleus of the Thalamus (ANT); Central Medial (CM) nucleus; Mediodorsal (MD) nucleus; Nucleus Accumbens (NAC); Periventricular Nodular Heterotopia (PVNH); Superior Temporal Gyrus (STG);Subthalamic Nucleus (STN)

Table 4.

Non-Parkinsonian Movement Disorders Clinical Studies Employing Two (or More) Targets.

Disease	Summary of Results	Strategy	Targets		# of Subjects	Leads	IPG	Frequency (Hz), (Min, Max)	Pulse Width (μSec), (Min, Max)	Voltage (V), (Min, Max)	Current (mA), (Min, Max)	Publication Identifier
dystonia	BFMMS max improvement from baseline 52% (GPi only) vs 92% (VIM/VO only) vs 92% (VIM/VO+GPi)	Multi-Target	VIM / VO	GPi	3	Boston Scientific, Vercise Cartesia	Boston Scientific, Vercise Gevia	89, 130	60		1.1–2	[64]
	BFMDRS improvement from baseline 59% (GPi+VO) vs 55% (GPI only) vs 56% (VO only) (p<0.01)	Multi-Target	GPi	VO	5	Medtronic, Model 3387		130	90	4		[66]
	BFMDRS score 12 (GPi only) vs 0 (GPi+VIM/VO)	Multi-Target	VIM / VO	GPi	1	Medtronic, 3389	Unnamed rechargable IPG	185	160	3.5		[62]
	BFMDRS improvement from baseline 68.8%	Multi-Target	GPi	VIM	9		Boston Scientific; Medtornic, Activa RC	119, 125	40, 210	1–2.6	1–4.7	[63]
	BFMDRS improvement from baseline 0% (GPi only) vs 0% (VIM/VO only) vs 25% (GPi+VIM/VO)	Multi-Target & Multi-Contact	GPi	VIM / VO	1	Medtronic, Model 3387, 3389	Medtronic, Activa PC	185	60, 90	4		[67]
	No additional benefit	Multi-Contact	GPi	VIM	1	Medtronic, Model 3387	Medtronic, Kinetra	150	210	3		[297]
dystonia tremor	BFMDRS score (patient 1, patient 2) before 51,74 vs after 21,23	Multi-Target	GPi	VIM	2							[68]
	FTMTR max improvement from baseline 43.1%	Multi-Contact	VIM	cZI	18	Boston Scientific Vercise™	Boston Scientific, Vercise IPG / Gevia IPG	170, 185	40, 100		1, 4.5	[74]
myoclonus-dystonia syndrome	UMRS improvement from baseline 54 (VIM only) vs 53.3 (GPi only) vs 41 (GPi+VIM), (GPI+VIM vs VIM only, p=0.043)	Multi-Target	GPi	VIM	7	Medtronic, Model 3387, 3389	Medtronic, Kinetra	130, 250 (VIM), 130, 180 (GPi)	60, 210	0.3, 4.5 (VIM), 1.5, 4.7 (GPi)		[69]
	STN+VLa induced dyskinesia, GPi+VLa induced dysarthria	Multi-Target & Multi-Contact	GPi	VLa / STN	1	Medtronic, Model 3387		135 (VO)	120, 210 (VO)	3.6 (VO)		[70]
	UMRS myoclonus w/ action score 3 (GPi+VIM/VO) vs 9 (VIM/VO only) vs 6 (GPi only)	Multi-Target & Multi-Contact	GPi	VIM / VO	1	Medtronic, Model 3387	Medtronic, Kinetra	140	120 (GPi), 90 (VIM/VOA)	2.3 (GPi), 2.1 (VIM/VOA)		[71]
spinocerebellar ataxia type 3 (SCA3)	SARA improvement from baseline 4% (GPi only) vs 38% (DN only) vs 42% (GPi+DN)	Multi-Target	GPi	DN	1	PINS L301		60, 185	90, 200	2, 3		[298]
essential tremor	Tremor control mean improvement 79.7%	Multi-Contact	VIM	PSA	33	Medtronic, Model 3387		125, 200	60, 110	1–4.2		[19]
	FTMTR improvement from baseline 86%	Multi-Contact	VIM	PSA	1	Boston Scientific		130	50		1.5–2.6	[20]
	Report on surgical techniques	Multi-Contact	VIM	PSA	6	Medtronic, Model 3387		135 (VIM), 160 (PSA)	60 (VIM), 90 (VIM)	2.1 (VIM), 1.7 (PSA)		[21]
	TRS dual vs VIM only p=0.36, dual vs non-VIM lead only p=0.82	Multi-Target	VIM	VO anterior	2			130, 190	60, 90 (VIM), 90 (VOA)	2.2, 3.5 (VIM), 2.8, 3.6 (VOA)		[75]
			VIM	Raprl / PSA	1			135	90	4 (VIM), 3.9 (Raprl)
complex essential tremor	FTMTR max improvement from baseline 58.2%	Multi-Contact	VIM	cZI	18	Boston Scientific Vercise™	Boston Scientific, Vercise IPG / Gevia IPG	170, 185	40, 100		1, 4.5	[74]
Holmes’ tremor (HT)	TRS mean score 4 (VIM/VO only) vs 3.7 (Raprl/cZI) vs 0.7 (VIM/VO+Raprl/cZI)	Multi-Target	VIM / VO posterior	Raprl / cZI	4	Medtronic, Model 3387		135	210			[299]
	TRS score 16 (VIM+VO) vs 20 (VIM only) vs 19 (VO only)	Multi-Target	VIM	VO	3	Medtronic, Model 3387		135, 185	60, 120	2.9, 4.1		[300]
multiple sclerosis tremor	TRS mean score 31 (VIM only) vs 30.88 (VO only) vs 27.61 (VIM+VO)	Multi-Target	VIM	VO	1	Medtronic, Model 3387		135	120	4		[77]
	TRS dual vs VIM only p=0.36, dual vs non-VIM lead only p=0.82	Multi-Target	VIM	VO anterior	1			130	60 (VIM), 90 (VOA)	1.8 (VIM), 2.8 (VOA)		[75]
			VIM	Raprl / PSA	1			130	90	4.1 (VIM), 2.5 (Raprl)
	TRS reduction from baseline 29.6% (VIM+VO) vs 20% (VO only) vs 21.1% (VIM only)	Multi-Target	VIM	VO	12	Medtronic, Model 3387		135, 185	60, 150 (VIM), 90, 120 (VO)	2, 4.2 (VIM), 1.5, 3.5 (VO)		[76]
fragile X-associated tremor/ataxia syndrome (FXTAS)	Tremor resolution	Multi-Contact	VO	VIM	1			130		2.5		[301]
spasmodic dysphonia (SD)	USDRS VO vs VIM+VO p<0.001, VIM vs VIM+VO p=0.20	Multi-Contact	VO	VIM	1	Medtronic, Model 3387		185	60, 90	2,3		[302]
neuroacanthocytosis	UHDRS score (patient 1, patient 2) GPi only 42, 36; VO only 35, 30; GPI+VO 31, 26	Multi-Target	GPi	VO	2	Medtronic, Model 3387		160	90 (GPi), 60 (VO)	3.5 (GPi), 3 (VO)		[303]

caudal Zona Incerta (cZI); Dentate Nucleus (DN); Globus Pallidus pars Interna (GPi); Posterior Subthalamic Area (PSA); Prelemniscal Radiations (Raprl); Ventralis Intermedius (VIM); Ventral Lateral Anterior (VLa); Venralis Oralis (VO)

Burke-Fahn-Marsden Dystonia Rating Scale (BFMDRS); Burke-Fahn-Marsden Movement Scale (BFMMS); Fahn-Tolosa-Marin Tremor Rating (FTMTR); Scale for the Assessment and Rating of Ataxia (SARA); Tremor Rating Scale (TRS); Unified Huntington’s Disease Rating Scale (UHDRS); Unified Myoclonus Rating Scale (UMRS); Unified Spasmodic Dysphonia Rating Scale (USDRS)

2.1.1. Parkinson’s Disease

Single-target DBS for patients with Parkinson’s disease (PD) has been highly successful for on-off fluctuations, tremor, and dyskinesia. However, with disease progression, the walking, balance, speech, and cognitive symptoms remain unmet needs. Clinical studies have explored multi-target stimulation for PD (Table 2) to improve alternative symptom profiles.

In a recent clinical trial, Schmidt and colleagues combined STN and GPi stimulation for six PD patients, which resulted in greater improvements in UPDRS III scores—both on and off medication—compared to STN or GPi stimulation alone, in both clinic and home settings [31]. The authors pursue a multi-target approach since (i) meta-analysis finds STN superior for motor symptoms but GPi outperforms in dyskinesia reduction, (ii) case studies suggest combined stimulation can achieve both priorities, and (iii) to provide more options of neural signals for their adaptive DBS algorithm. Prior works include small case studies in which an additional electrode lead was implanted to “rescue” DBS; for example, additional GP targets in instances of refractory dyskinesia following STN single-target stimulation [32,33]. With STN+GPi stimulation, benefits may arise from suppressed STN-GP coherence [34]. Additionally, multi-target strategies have been employed for non-traditional symptom profiles such as freezing of gait PD [35–41], tremor dominant PD [42], and to improve PD rigidity [43] and bradykinesia [31]. Schmidt and colleagues suggest more typical phenotypes may also benefit from dual target. However, large parameter spaces lengthen programming time: the authors optimized each single target, before optimizing dual target, a process which took four to twelve months of multiple 90 min programming sessions [31]. In summary, further studies are needed to clarify the efficacy and mechanisms of dual-target DBS across both typical and atypical symptom profiles, while adaptive programming strategies will be essential for broader clinical implementation.

2.1.2. Multi-Focal Epilepsy

Patients with diffuse or multifocal seizure onset zones are challenging to treat [44,45], at times requiring surgical resection in addition to neurostimulation [46]. We now appreciate that epilepsy affects the normal connectivity of the brain, even cases of focal epilepsy, therefore targeting only one or two areas may, in some cases, not be enough modulation to control the disease [47]. Much of neuromodulation’s efficacy may be due to long-term modulation of the underlying brain networks [47–49], rather than immediate interruption of the electrical signal(s) underpinning a seizure [50]. This view of epilepsy as a network disorder enables more personalized candidacy and targeting strategies. Stimulation targets are individualized for each patient, allowing for the possibility of multi-target approaches. We propose that the network perspective also supports the use of multi-target stimulation—enabled by DBS and RNS— to more effectively modulate brain networks. Several case series have explored the use of multi-target RNS and DBS for drug resistant epilepsy (DRE) (Table 3), yet large controlled studies are lacking.

Clinical RNS systems already include multi-target interfaces for epilepsy treatment. The NeuroPace RNS system qualifies as a multi-target device, having received FDA approval in 2013 for the stimulation of one or two epileptogenic foci using 4-contact depth electrodes or cortical strip electrodes [51]. Network-based modulation may be more effective with coordinated stimulation using a distributed system, such as Neuropace’s simultaneous cortical and depth leads [52]. Compared to studies of RNS targeting a deep brain target plus the cortex [53–56], fewer have considered two or more deep structures. Two deep targets, the hippocampus and periventricular nodular heterotopia (PVNH, which arise from neurons that fail to migrate and cluster near the ventricles), were considered for four patients, who achieved seizure reduction rates of 80.2 to 88.1% [57]. Targets were chosen based on ECoG, and device recordings confirmed that both regions were involved in seizure activity. In another work, authors conclude that for four patients with mesial temporal lobe epilepsy, their clinical response to thalamus and hippocampus stimulation was dependent on contact distance to the foci, and that additional electrodes only increase the probability of interfacing with the foci [52]. Therefore, these case series suggest that multi-target RNS may be necessary when the spatial distribution of seizure foci limits the efficacy of single-target approaches.

The effectiveness of multi-target DBS for DRE remains uncertain, with mixed evidence supporting its clinical utility. Nucleus accumbens (NAC) stimulation was hypothesized to complement anterior nucleus of the thalamus (ANT) stimulation in the inhibition of frontal and temporal lobe seizures; however, technical limitations required identical stimulation frequencies at both targets and no benefits were observed [58]. Centromedian nucleus (CM) + ANT was expected to desynchronize cortical and thalamocortical seizure circuits, but when compared to CM only stimulation, no significant difference was found in seizure frequency reduction [59]. In contrast, one series found combinations of primary and secondary targets benefited two epilepsy patients with multiple seizure onset zones [60]. Primary targets were chosen for proven effectiveness from randomized controlled trials [61,62], and compatibility with secondary targets that required no patient repositioning. Secondary targets were chosen according to seizure characteristics. For example, the mediodorsal nucleus (MD) was added for seizures with severe motor features. One patient had seizure reduction with MD only stimulation increased from 16.7% to 37.5% with added ANT; the other failed to respond to MD or ANT alone but achieved a 45.5% seizure reduction with combined [60]. The authors note the difficulty and length of manual multi-target DBS programming. Single target stimulation was optimized and exhausted before proceeding to additional targets such that multi-target stimulation was not attempted in the other three patients [60]. Further studies are needed to determine whether mixed outcomes reflect true biological limitations or suboptimal selection of targets and stimulation parameters.

In summary, for patients with multi-focal epilepsy, single-site stimulation unresponsiveness and multi-target efficacy may depend on patient specific foci patterns, yet these are difficult to predict. Case studies present mixed findings due to this foci pattern heterogeneity and large parameter spaces. Additionally, our current classification of epilepsy phenotypes is based on clinical assessment and may not capture brain network involvement adequately. Thus, our current understanding of multi-target results may not represent optimal applications. Retrospective studies should continue to document the benefits of single versus multi-target stimulation in stimulation refractory cases where DBS or RNS targets have been added on. Such studies can help determine foci characteristics which may preferably respond to a multi-target approach, and whether additional sites should be targeted in initial surgeries or added later once single site efforts have been exhausted. Lastly, RNS provides on-demand stimulation while traditional DBS is continuous, and it remains to be seen which strategy is amendable to the multi-target approach.

2.1.3. Other Movement Disorders

Several other movement disorders have been considered for dual target neurostimulation in attempts to target specific symptoms and to achieve increased clinical benefits. Several diseases are discussed here while a more comprehensive summary can be found in Table 4.

The circuitry of dystonia includes multiple locations [63], including thalamic connections to the cerebellum in the ventralis intermedius (VIM) and ventralis oralis (VO) nuclei [64–67] and cortical connections in the GPi [65,66]. Multi-target stimulation of these regions has been explored in cases of dystonia [64–66,68,69], dystonia tremor [67], and myoclonus-dystonia syndrome [70–73]. Mixed findings suggest greater symptom reduction may arise from dual stimulation of cerebello-thalamo-cortical and striato-pallido-thalamo-cortical networks [64–66,68–70,72,73]. There are many causes of dystonia which can be classified based on the location of symptoms or by possible pathologies or genes [63,74]. Dystonia, with different symptom subtypes arising from specific circuits throughout the brain, may be amendable to multi-target neurostimulation.

Similarly, tremors related to multiple sclerosis (MS) and essential tremor (ET) represent potential applications for distributed stimulation, given emerging hypotheses that tremor pathophysiology involves dysfunction not only within the cerebello-thalamo-cortical (CTC) loop but also within pallidal circuits. Multi-target DBS may be required in some cases to achieve volume of tissue activated over relevant fiber tracks necessary for tremor suppression [75]. The VIM has served as a primary target while the VO nuclei or prelemniscal radiation (RAPRL) serve as secondary targets for refractory cases. Multi-contact approaches were prospectively implemented to provide alternative stimulation options in the event of VIM-refractory tremor [19–21,76]. Multi-target approaches were initiated after VIM stimulation showed diminishing efficacy: a retrospective study found dual stimulation of VIM with either VO or RAPRL provided greater tremor control than VIM alone in two of three patients with essential tremor, and for two patients with MS related tremor [77]. Multi-target case series for multiple sclerosis patients suggests benefits of dual over single target stimulation [78,79]. With regards to ET, not all sub-types respond ideally to DBS, potentially due to varying cellular pathologies [80]. Experts now recognize that ET may not be one disease, with some cases progressing more quickly or having more axial symptoms. ET and other disorders which display dysfunction in non-overlapping circuits and locations thus may benefit from distributed neurostimulation.

2.2. Neuropsychiatric Disorders

Neurostimulation of deep brain structures has been posed as a method to correct dysfunctional synchronization in neuropsychiatric disorders [81–84], most notably in obsessive-compulsive disorder (OCD). However, the benefits of multi-site stimulation remain largely unknown as animal models cannot fully capture the human mental experience. Technical and ethical barriers limit human experimentation. Nevertheless, several small case studies and early clinical trials (Table 5) suggest that distributed neurostimulation technologies may be well positioned to influence system level disturbances and provide desired clinical symptom relief.

2.2.1. Cortico-Striatal-Thalamic-Cortical Loop Circuit Disorders

Research enabled by the 2009 FDA Humanitarian Device Exemption (HDE) for OCD DBS [85] has revealed several possible brain targets [86,87] yet interpatient differences and variable symptoms remain unaddressed. There remains a need for therapies which impart lasting and replicable treatment benefits. OCD is increasingly seen as a network disorder, a perspective which may permit new treatment opportunities, such as for multi-target neurostimulation. For example, a “sweet spot mapping” and “fiber filtering” paradigm estimated that there were optimal and separate stimulation sites for compulsion, depression, and anxiety symptoms in patients with OCD [88,89]. Combinations of symptom motifs may arise from dysfunctions in several overlapping brain circuits, with contributions from frontal, parietal, and limbic areas [90–92]. Traditionally, dysfunction was attributed to hypersynchrony in the cortico-striatal-thalamic-cortical (CSTC) loop circuit; more recent and higher resolution recordings suggest hypo-synchrony may contribute as well [93].

Inspired by DBS protocol for movement disorders, one study used an awake intraoperative optimization approach, where a neuropsychologist and psychiatrist adjusted stimulation parameters based on patient feedback to determine whether one of two anterior commissure targets—the VC/VS or the BNST— or combined provided optimal benefit [94]. Other CSTC structures have been targeted with benefits reflected in the Yale-Brown Obsessive-Compulsive Scale (Y-BOCS) score and connectivity profiles [86,87,93]. Stimulation of multiple nodes along neural oscillator and feedback circuits of the CSTC fiber tracts may enable regulation of symptoms caused by underlying hyper and hypo synchronization.

Patients with OCD may have comorbid Tourette’s syndrome (TS) and a few studies have considered multitarget DBS to manage both diseases simultaneously [95,96]. Both target the ventral capsule/ventral striatum (VC/VS) for OCD with additional movement disorder targets (GPi or CM) for Tourette’s. Benefits are achieved for both diseases with reduced tics, vocalizations, and visual disturbances [95,96]. CSTC stimulation may mitigate not only OCD symptoms but also those of TS: CSTC dysfunction may contribute to the impairment of action/object representations [97] and the misprocessing of noise [98] in TS. Other theories— the suppressing of the motor pattern generator hypothesis and the striatum-basal ganglia-thalamus network theory [99]— propose additional pathways [97,100] most likely to also contribute to TS. Multi-target DBS may provide a strategy to treat diseases of the anatomically distributed CSTC loop circuit.

2.2.2. Disorders of Mental States

Distinct and distributed brain structures and circuits are associated with various emotional states, as seen in depression [101,102]; “multidimensional” facets of chronic pain—sensory, affective, and modulatory [103–105]; and opposing symptom dimensions, such as the positive and negative symptoms of schizophrenia [106].

Stimulation of distinct structures has been found to elicit different mental states, relevant for the treatment of depression and chronic pain. Treatment-resistant depression is characterized by a diminished capacity for reward and heightened sensitivity to distress. Connectivity analyses have identified distinct brain networks underlying opposing positive and negative emotional states [101,102]. Small case studies of dual-target DBS aim to modulate structures associated with emotional valence [107–109]. Specifically, one case report leveraged subcallosal cingulate (SCC) DBS to mitigate distress and ventral capsule/ventral striatum (VC/VS) DBS to augment positive valance [108]. Similarly, the multidimensional theory for chronic pain has inspired small studies of blended target [110] and multi-target [111,112] DBS to influence separate factors of pain. Multi-target stimulation may allow for enhanced efficacy and sensitivity of treatment for diseases in which stimulation of distinct structures augments or diminishes aspects of mental states.

Friston and Frith’s “Disconnection Theory” of schizophrenia posit malfunctions arise in network-wide computations which underly the brain’s prediction capabilities [113,114], although the exact mechanisms of the disease remain undetermined. Decreased connectivity may prevent deep brain structures from engaging in cortical feedback, thereby contributing to the positive and negative symptom profile associated with the disorder. Small case studies of patients with schizophrenia, exploring off-label single-target [115] and bilateral [116] DBS, have demonstrated safety yet remain inconclusive in efficacy. It remains an open question if restoring network periodicity would provide clinical relief to patients. Future works should determine if this aim can be achieved with single site DBS or if greater network coverage and multiple targets are required.

2.2.3. Substance Use Disorders

A ”hedonic hotspot” theory suggests that brain targets responsible for substance use disorder are inherently distributed. This Incentive Sensitization Theory posits that addiction arises from hyperactivation of cubic-millimeter size nodes throughout the brain and the meso-cortico-limbic circuit which produce pleasure and drive incentive seeking [117–119]. Proponents of the theory have uncovered distributed hotspots in animal models—in the nucleus accumbens, ventral pallidum, hypothalamus, and amygdala—with the aid of chemical microinjections, optogenetics (with Fos measurements), and behavioral experiments. However, clinical trials are ultimately necessary to ascertain the model’s validity in humans and potential impact on clinical treatment. Small, in-human studies have suggested the possible effectiveness of single-site DBS for alcohol [120,121] and opioid use disorders [122]. If the hedonic hotspot and other network-wide models of addiction are validated in humans, multi-target neurostimulation may serve as a therapeutic tool to simultaneously modulate the locations responsible for craving and misuse.

2.2.4. Alzheimer’s Disease

Human clinical trials of single-site DBS for Alzheimer’s disease (AD) have been small and clinically disappointing. The ADvance Phase II Clinical Trial was a randomized, double-blind study of fornix region single-site DBS versus sham which found no significant differences between the study groups at one [123] or two [124] years. The lack of single-site electrophysiology or symptomatic biomarkers contributed to difficulties in target choice and parameter optimization for Alzheimer’s disease [125]. Despite ongoing efforts to identify targets—particularly within the Papez circuit—network-wide cholinergic dysfunction suggests that moving beyond single-target DBS and exploring multi-target DBS may accelerate progress [126,127].

Challenges in Alzheimer’s disease (AD) DBS research provides insights relevant to other neuropsychiatric disorders applications. In contrast to neuropsychiatric disorders, Parkinson’s disease and epilepsy have begun to establish biological markers to determine the efficacy of treatment. Benefits of neurostimulation for neuropsychiatric disorders may manifest over longer time scales than movement disorders and real-time assessments of treatment efficacy are critically needed. For example, a recent in-human study characterized electrophysiology biomarkers for chronic pain [128]. Their biomarker analysis suggests affective and sensory components arise from separate brain circuits. Such biomarkers could assist in performing real-time assessments of both current single-target and future distributed brain stimulation systems. That said, for all neuropsychiatric applications, a careful ethics board review as well as informed consent will be required.

Many studies in Tables 2–5 implement multi-target neurostimulation with dual-lead DBS for cases with irregular symptom profiles, known gaps in treatment efficacy, or as a rescue when a first target lacks desired outcomes. Network models of disease propose multi-target stimulation can engage multiple circuits or better coverage of dysfunctional brain structures. However, more randomized controlled trials will be needed to directly compare single- and multi-site neurostimulation. Further advances remain necessary to safely increase the number of electrodes, the topic of the subsequent sections.

3. Surgical Barriers to Increases in Neurostimulation Target Number

Patient safety remains central to the translation of multi-target neurostimulation. Risks of multi-contact stimulation mirror barriers to multi-target systems: Hollingworth and colleagues state that realizing benefits from multi-contact technologies is a “risky and expensive problem” in which the increased number of electrodes and lead trajectories “multiplies the risk of hemorrhagic and infective complications as well as the financial cost” [110]. Multi-target systems introduce surgical challenges associated with custom electrode trajectories, including the risk of ventricular entry, vascular injury, and traversal through cortical sulci [22,23]. Risks can manifest as devastating consequences including psychiatric effects, paralysis, and even death [22]. Although the multi-contact method may lower some risks, it requires that safe trajectories exist, and that targets of interest are relatively nearby. Multi-target neurostimulation may allow for modulation of separate and distributed anatomical structures, unfit for multi-contact, but must prioritize patient well-being.

Findings are mixed on whether a second lead increases patient risk [129–131]. Advances in hardware and surgical techniques have contributed to a reduction in DBS complications over time [132,133]. Despite these improvements, adverse events must be considered in a cost-benefit analysis of distributed neurostimulation systems. Additionally, the patient burden must be considered. Modern implantable pulse generator (IPG) technologies have limited power and lead connections: the multiple IPGs necessary to increase the number of targets increases surgeries and device powering considerations (i.e.. recharging by the patient). Current IPG have restricted programming configurations and on-device processing capabilities. While neural engineers continue to address such technical challenges, the goal of this section is to evaluate surgical risks and to assess whether the clinical benefits discussed previously outweigh the associated costs.

We review five specific risks, in order of most to least severe, in the context of multiple electrodes (Figure 3). Rather than presenting a comprehensive meta-analysis of DBS-related adverse events, this section highlights key trends across existing meta-analyses (Table 6) and examines their implications for multi-target deep brain interfaces.

Figure 3. Challenges of Distributed Intracranial and Deep Brain Stimulation Interfaces.

(a) Stimulation side effects, often indistinguishable from disease progression or drug effects, include motor, affective, and autonomic disturbances. Manifestations of such effects are annotated on the figurine. (b) A representation of surgical imaging software shows the desired and actual target trajectories, with a warning indicating slight differences between the two. Clinical outcomes depend on lead placement. Pre-surgical planning optimizes electrode positioning, while suboptimal placement increases side effects. (c) Vascular damage can cause hemorrhage or edema, potentially leading to paralysis or death, seen on imaging [305]. Hemorrhage is seen around the lead on the right side of the image, indicated by a warning. (d) A patient lies on a surgical table, with a visible infection. Scalp and burr hole infections may require device removal and, in severe cases, be fatal. (e) A traditional macroelectrode Deep Brain Stimulation (DBS) stimulator, lock, and implantable pulse generator (IPG). Lead migration risks are mitigated by stability technologies, but rigid devices risk fracture. Figure created with Figma.

Table 6.

Prevalences of DBS Adverse Events, Reported in Meta-Analyses Published 2022- March 2025

Adverse Event	% Prevalence (n)	Short Term Effects (n)	Chronic Effects (n)	Study Identifier
Intracerebral Hemorrhage	0.6% (5)	“aphasia, speech initiation disorder, dysarthria, central facial palsy, paresis of the left hand, frontal lobe syndrome, delirium, and cognitive decline”	hemiparesis (2), cognitive disorder (1), diplopia (1), dysarthria (1), balance disorder (1)	[133]
	0.8% (3)		death (11)	[134]
	1.3% (7)	hemiparesis (5), abandoned procedure (2)	residual hemiparesis (1)	[135]
	2.9% (379)			[136]
Intraventricular hemorrhage	1.05% (1)			[137]
Subdural Hematoma	0.52% (2)	surgical drainage (2)		[138]
	1.05% (1)			[137]
Cerebral Edema	1.6% (13)	“cognitive decline, dysarthria, apraxia, speech difficulties, apathy, confusion, and delirium… seizures”	“mild cognitive impairment” (1)	[133]
	3.74% (51)	“headache, nausea, confusion, a decline in cognitive function, mood changes, aphasia, disorientation””	“gait instability, muscle strength loss/paralysis, and epilepsy”	[140]
Infection	2.66% (14)			[137]
	2.95% (22)	antibiotics (22), skin erosion/granulation tissue (19), full system explantation with replacement (2), partial system explantation (5), IPG explant (5)	full system explanation with no replacement (7)	[135]
	3.4% (27)	antibiotics, removal and re-implantation		[133]
	6.67%			[146]
	9.95% (20)	IV antibiotics (14), wound revision (7), partial hardware removal (12)		[138]
	10.1% (36)			[134]
Skin Erosion	1.7% (6)			[134]
Lead Fracture	0.39% (2)	extension wire replacement (2)		[135]
Lead Migration	0.52% (2)	surgical “revision” (2)		[138]
	2.1% (17)			[133]

3.1. Hemorrhage and Edema

Expanding lead counts with distributed neurostimulation raises concern for hemorrhage and edema, making complication risk a key consideration for system design. Advancements in surgical techniques have kept hemorrhage rates relatively low, typically ≤1%. A recent meta-analysis report revealed an intracerebral hemorrhage prevalence between 0.6% and 2.9% [134–137] and subdural hematoma prevalence between 0.52% and 1.05% [138,139]. In addition to intraparenchymal hemorrhage during lead placement there can also be subdural hemorrhage, which can occur from direct injury or CSF loss leading to avulsion of draining veins. This complication can potentially be reduced with lead implantation modalities that minimize or avoid CSF egress. Although often asymptomatic, severe cases can result in hemiparesis [134,136], cognitive and motor impairments [134], or rarely death [140]. Peri-lead edema characterized by fluid accumulation around DBS leads, although uncommon and often asymptomatic, has prevalence of 1.6% and 3.4%, with long-term consequences of epilepsy, and cognitive and motor impairments [134,141].

Although infrequent, hemorrhagic and edema complications must be carefully considered when increasing electrode count. Increased number of brain penetrations by temporary location-optimizing microelectrodes correlates with higher hemorrhage risk, ranging from 0.7% to 2.9% per trajectory [136,142,143], raising the question whether increased trajectories of macroelectrodes would follow the same trend. However, microelectrodes may serve as an imperfect proxy: DBS electrodes are typically 1.27–1.3 mm in diameter [144], while microelectrodes used for targeting are much smaller, often micron-scale [142]. Other material and geometry differences exist, limiting conclusions which can be extrapolated to increased macroelectrode brain penetrations. More concerning, simultaneous bilateral implantation is associated with increased hemorrhagic complications when compared to staged bilateral and unilateral surgeries [143], suggesting surgeries simultaneously implanting multiple targets must consider increased bleeding risk. Image-guided planning and stereotactic systems can reduce risk, but hemorrhage risk could constrain distributed systems relying on modern DBS lead-based devices. Further studies are needed to assess the trade-off between clinical benefit and risk from multiple electrodes.

3.2. Infection

As distributed neurostimulation systems introduce more implanted hardware, understanding how electrode number and surgical complexity impact infection risk becomes critical for ensuring long-term safety and viability. The association between infection risk and the number of implanted electrodes remains ambiguous. Evidence is limited to unilateral versus bilateral macroelectrode comparisons, which show no significant difference [131,145]. However, stereo electroencephalography (sEEG) may serve as a surrogate, and a pediatric study found a significant association between depth electrode count and surgical site infections [146]. Furthermore, if infection risk compounds with the number of leads, extra precautions may be needed for distributed systems. Such single-lead infection rates are well documented with prevalence ranges from 2.66% to 10.1%— of which scalp and burr hole infections account for 1.35%–17.8% (Bouwens van der Vlis et al., 2022; Bullard et al., 2020; Cabral et al., 2022; Doshi et al., 2022; Holewijn et al., 2024; Jitkritsadakul et al., 2017; Lapa et al., 2024; Zheng et al., 2024). Importantly, most infections arise at the IPG site rather than the scalp or burr hole, and intracranial infections are rare [139,140,145,148]. Therefore, while distributed DBS systems necessitating an additional IPG may increase the risk of infection, more research is needed to determine whether the number of implanted electrodes themselves contributes to a higher incidence of infection. If additional or larger incisions are needed, this would possibly increase infection risk.

3.3. Stimulation Induced Side Effects

As the number of leads and stimulation sites increases with multi-site stimulation, the potential for side effects may also rise, making it essential to thoroughly understand how stimulation parameters, target regions, and electrode placement all potentially contribute to adverse outcomes. Although speculative, multi-site stimulation systems may induce more complex side effect profiles due to the interaction of different target regions. To safely implement distributed neurostimulation systems, it is essential to thoroughly understand how stimulation parameters, target regions, location within the target, and interactions between targets contribute to adverse outcomes. Animal model studies and human cases of multi-site lesions from strokes or tumors could provide insights. With such knowledge, stimulation side effects may frequently be mitigated or avoided with changes in parameters and stimulation configuration. Single-target DBS is associated with well-documented motor, sensory, psychological, and autonomic side effects [142,149–151]. However, for side effects that develop over longer time scales, differentiating whether they arise from DBS, disease progression, or other treatments can be challenging [152]. These examples highlight the complex relationship between physiological, affective, and autonomic side effects and the interplay of intra-target location and stimulation parameters. A deeper understanding of these factors will be essential for mitigating risks as multi-site stimulation becomes more prevalent.

3.4. Device Targeting & Migration

Whether an increased number of leads contributes to greater targeting error remains an open question. A single-site retrospective review of 125 patients found no significant difference in targeting error between simultaneous and staged bilateral DBS implantation, or between first and second electrodes implantation [130], though small series have shown the second side placed in a single intraoperative setting may be less accurate possibly due to brain shift. Advances in individualized stereotactic systems, intraoperative imaging, and robotic-assisted surgery have helped maintain targeting accuracy in modern DBS procedures [132]. These technologies can consistently keep targeting errors within 2 mm [132,133,153]. Current robot-assisted (RA) surgical systems filter surgeon's natural tremor and undesired movements yet meta-analysis have shown comparable durations to non-RA surgeries, demonstrate a learning curve, and mixed results regarding accuracy benefits [154,155]; to our knowledge, no duration or accuracy data has been released for emerging semi-autonomous neurosurgical robots (ANR) systems [156,157]. Furthermore, if multi-target stimulation systems are miniaturized to keep implanted volume low— such as with nanoparticles and micro-device technologies— it remains unclear if 2 mm is an acceptable spatial error. Exactly where to place micro and nano technologies within a target and the effects of submillimeter deviations remains unknown.

A key factor in lead placement accuracy is brain shift, defined as the displacement of brain parenchyma during or after surgery. Neurosurgery teams routinely use post-operative imaging to assess lead placement and migration [132,158]. Lead shift was reported up to 0.6 mm, with an additional 0.3 mm shift contralateral to the burr hole; bilateral placements showed significantly greater shift (0.7 mm), likely due to longer procedure times and increased CSF egress [159]. Other studies have similarly identified procedure times [160] and intracranial air [161] as predictors of lead displacement, highlighting the importance of limiting CSF egress when placing multiple brain leads. While lead migration may be lessened by the StimLoc^™, other securing devices [162], or glues, meta-analysis considering modern DBS macroelectrodes have reported lead migration rates between 0% to 3.49% [136,139,140]. Additionally, these securing techniques may be affected by factors related to anesthesia and operative management. In contrast, sEEG electrodes are less prone to migration due to small dural openings and short implantation duration, allowing for higher electrode counts with variable placement accuracy. CSF egress with this modality can be controlled more easily by simply capping the bolt if a significant CSF leak emerges [163]. sEEG and DBS leads are both purposely placed to avoid the ventricles. Distributed neurostimulation systems may be similarly optimized in the development process to design features such as miniaturization of the dural opening or to changes to minimize the biomechanical related CSF distribution.

3.5. Hardware Robustness

The competing demands of preventing both lead fracture and lead migration remain a critical consideration when introducing additional electrodes or novel brain-interface designs. Fracture rates range from 0% to 9.3% [136,147,158,162]. Flexible leads reduce fracture risk but may increase migration, while rigid fixation limits migration but raises fracture risk [142]. On the other hand, multi-target systems could be employed to disperse failure risk across with planned redundancy, similar to how additional “rescue” implants are added when stimulation efficacy decreases.

The prevalence of adverse events in modern DBS surgery is well studied, leading to the development of technologies and techniques that mitigate risks. Sections 2 and 3 comprise our risk-benefit analysis of macroelectrode DBS systems employed for multi-target neurostimulation. Preliminary evidence suggests that multi-target stimulation may provide enhanced control of symptom related distributed brain networks. Several diseases in which dysfunction is inherently distributed throughout brain anatomical structures may benefit from a multi-target approach. However, the successful translation of emerging multi-target stimulation systems requires careful consideration of lessons learned from current technologies. While multi-target stimulation could be used to minimize side effects, stimulation across multiple connected brain regions may paradoxically induce stimulation side effects if parameter and contributions related to device location are not understood. Additionally, for multi-target DBS systems, precise electrode implantation is essential to prevent lead migration while also ensuring accurate positioning. Therefore, as neuromodulation systems evolve to incorporate multiple targets, engineers have begun to reimagine the scale, materials, and configuration of these systems to enhance therapeutic outcomes while minimizing risks.

4. Hardware Solutions for Distributed Neurostimulation

Advancements in DBS hardware, including wireless IPG battery recharging and directional, current steering electrodes have already led to improvement in the cost-to-benefit ratio of brain stimulation technologies [7,8]. However, due to considerations in Section 3, implanting more than four macroelectrode DBS or RNS leads is currently infeasible. Efforts in academia and industry can be directed to determine if emerging technologies can enable multi-target neurostimulation leveraging a greater number of targets and if doing so can improve therapeutic outcomes or be applied to optimally implement closed loop algorithms.

This section considers systems capable of multi-target stimulation with far-, mid-, and near- term translational readiness (Figure 4). Far-term applications include highly novel approaches such as injectable stimulators, system-on-chip devices, nanoparticle-mediated, and non-invasive neuromodulation, which are still in conceptual stages. These systems are primarily being developed in academic laboratories, and face challenges in funding, scale-up, and clinical validation. Mid-term refers to more transformative platforms that are in early pre-clinical or clinical trials which incorporate wireless communication, high channel counts, or advanced on-chip processing. Near-term translation may be achieved with enhanced designs and improved targeting algorithms of more traditional DBS and RNS systems on the market, introduced in Section 2: dual-target and multi-contact applications may continue to gain clinical adoption within the next few years.

Figure 4. Clinical Readiness of Devices Capable of Multi-Target Stimulation.

Systems are evaluated in terms of far-, mid-, and near-term translational readiness for widespread clinical adoption as a multi-target therapeutic. Four categories partition the readiness scale: academic research grade, industry research and development (in the form of pre-clinical trials), to-market clinical trials, and on-going clinical investigation. For devices in pre-clinical and clinical trials, company (black) and device (grey) names are included.

While near- and mid-term systems will be introduced, the goal of this section is to review academic literature on novel far-term approaches to multi-target neurostimulation, specifically those which deviate from traditional macroelectrode DBS and RNS designs. Such far-term technologies, with a 20+ year horizon, are currently underfunded and lack commercial momentum. However, this could change if scientists and clinicians begin to recognize the specific clinical value of distributed, multi-site stimulation. Broader awareness and early-stage support could accelerate innovation and help bridge the translational gap for these next-generation systems. Optogenetic and sonogenetic techniques, while capable of multi-site circuit manipulation using holographic light patterning [164,165], face regulatory [166] and safety concerns [167] due to human genetic manipulation and thus will not be considered.

4.1. Implantable Electronic Devices

Academic researchers have prototyped miniaturized, wireless implantable neural stimulators. These devices may be implanted within the skull, placed on the cortex, or inserted into deep brain structures (Figure 5). This section also reviews cortical and dural stimulators, which, though not designed for deep brain use, could be adapted for such applications. Furthermore, we distinguish between system-on-chip microdevices (Section 4.1.1) and millimeter size, multi-channel devices (Section 4.2.2). Our focus is on stimulation devices (Table 7), but we also include multi-site recording devices (Table 8) that could enable stimulation or closed-loop use. Additionally, two wireless modules (Table 9) offer short-term solutions to reduce wiring with more targets. The following section will review specifications for multi-target stimulation (Table 10) and potential solutions.

Figure 5. Invasive System-on-Chip Multi-Target Implants.

Not to scale. Devices are placed on a single coronal cut in a head model for clarity. Devices include (from left to right): (a) Neuralink’s N1 implant, (b) Neural Dust system, (c) Neurograin, (d) Traditional DBS, (e) ENGINI system, (f) DOT, and (g) Microbead. Device sizes and communication links are noted underneath their names. Components of wireless links are displayed on or above the head model: (a) N1 Bluetooth module, (b) ultrasound communicator, (c, e-g) RF and magnetoeletric coils. Devices (c) and (e) include coil repeaters placed above the brain. Bilateral DBS macroelectrode leads are included for comparison.

Table 7.

Technical Specifications for Stimulating Electronic Implants.

	Stim Dust	Microbead	Neurograin	Optical Neurograin	Magnetoelectric Bio-Implants	Intracortical Visual Prosthesis (ICVP), Wireless Floating Microelectrode Arrays (WFMA)	Digitally programmable Over-brain Therapeutic (DOT)
Publication Identifier	[182]	[176]	[169]	[188]	[198]	[306]	[205]
Clinical Readiness	Academic Research	Academic Research	Academic Research	Academic Research	Academic Research	Clinical Trial	Pre-Clinical Trial
Transmission Type	Ultrasound	RF	RF, Inductive	Optical	Magnetoelectric	RF, Near-field	Magnetoelectric
TX Coil Size (Diameter)	none	1 cm	2 cm	none	6 cm	―	6 cm
Repeater Coil Size	none	none	2 × 2 cm	none	none	none	none
On-Chip Coil Material	PZT	Al	Al	GaAs Photovoltaic	Metglas, PZT	Ag	Metglas, PZT
On-Chip Coil Size (mm2)	0.5625	0.07	0.25	none	6	―	22.5
Powering Frequnecy	2 MHz	0.5 – 3 GHz	915 MHz	333 – 375 THz	330 kHz	4.5 MHz	218 kHz
Down Link Rate (kbps)	―	―	1000	1000	5.16	1250	1.6
Up Link Rate (kbps)	―	―	10,000	―	―	140 kHz	5
Transmission Scheme	pulse-width edge detection, single bit on-off keying	pulsed power transmission	asynchronous sparse binary identification transmission (ASBIT)	Manchester-encoded time division multiplex	amplitude shift keying	pulse-width modulation	on-off keying, resonance decay backscatter
Device Size (um) †	≈ 2000 × 2000 × 2000	300 × 300 × 80	500 × 500 × 35	2900 × 1700	1000 × 800	≈ 55 × 55	9000 × 9000 × 11,000
CMOS Tech (um)	0.065	0.13	0.065	0.18	0.18	―	―
Chip Power (uW)	48	38	30	40	400	280	56,000
Working Distance (cm)	2.15	0.66	0.8	―	3	3	0.75
Max Stim Amplitude	400 uA	38 uA	25 uA	20 uA	3.5 V	80 uA	14.5 V
Max Stim Duration (us)	continuous	500	400	1000	1200	200	500
Max Stim Frequency (Hz)	2000	200	500	550	1000	100	500
Stimulation Location	Sciatic Nerve	Sciatic Nerve	Cortex	Cortex	Sciatic Nerve	Visual Cortex	Dura
Testing	Ultrasound gel bath, rodent sciatic nerve	Rat sciatic nerve, water bath, and between slices of raw beef.	Saline solution, liquid head phantom, rodent under ketamine anesthesia.	Tested pulse and burst mode stimulation in rat model with direct laser powering (no skull or scalp included).	“Porcine tissue”, hydra vulgaris cell culture, rat sciatic nerve.	Phase I clinical trial in one human subject.	Human craniotomy stimulation proof of concept. Pig chronic study.

^†for sizes with ≈, approximations were made assuming square or cube sized device

Note: “―” represents specification not indicated

Table 8.

Technical Specifications for Recording Electronic Implants.

	Neural Dust	FF-WINeR	WiPRoC	BioBolt ECoG	Microscale Opto-Electronically Transduced Electrode (MOTE)	ENGINI
Publication Identifier	[182]	[185]	[212]	[210]	[186]	[214]
Clinical Readiness	Academic Research	Academic Research	Academic Research	Academic Research	Academic Research	Academic Research
Transmission Type	Ultrasound	RF	RF	Galvanic	Optical	RF, Near-field
External Coil Size (Diameter)	none	45 mm (diameter)	24 mm	none	none	―
Repeater Coil	none	32 mm (diameter)	none	none	none	yes
On-Chip Coil Material	PVT	AWG wire coil	―	―	AlGaAs Photovoltaic	Au
On-Chip Coil Size (mm)	0.5625	1.2 (diameter)	3 × 3	none	none	3.5 × 3.5
Powering Frequnecy	1.85 MHz	60 MHz	144 MHz	none	545 THz	433 MHz
Down Link Rate (kbps)	0.5 Mbps	―	15	10	―	205
Up Link Rate (kbps)	―	―	―	―	―	―
Transmission Scheme	amplitude shift keying	on-off keying	analog-shift keying	frequency shift keying	pulse-position modulation	frequency shift keying
Device Size (um) †	800 × 3000 × 1000	1000 × 1000	3000 × 3000	≈ 11,000 × 11,000 × 11,000	250 × 57	4000 (diameter), 1000 (thickness)
CMOS Tech (um)	custom	0.13	0.18	0.25	0.18	0.35
Chip Power	0.12	93	160	365	1	92
On Board Battery	none	none	none	Li ion	none	none
Working Distance (cm)	0.88	1.6	1	direct contact	0.6	―
Testing	Tested recording in rat gastrocnemius muscle and sciatic nerve, comparing to wires ground truth.	Tested in ex-vivo dissected lamb head (w/o bone and skin)	Bench testing only.	Tested in non-human primate.	Bench testing only. Successful recovery of a pre-recorded neural signal.	Anti-buckling insertion and system tested in Agarose Gel.

Note: “―” represents specification not indicated

Table 9.

Technical Specifications for Bidirectional, Wireless Communication Modules.

	Bidirectional Communication Module	CMOS System-on-Chip TX/RX Module
Publication Identifier	[222]	[221]
Clinical Readiness	Academic Research	Academic Research
Transmission Type	RF	RF
External Coil Size (Diameter)	42 mm (diameter)	2 × 2 cm
Repeater Coil Size	20 mm (diameter)	none
On-Chip Coil Size (mm)	10 mm (diameter)	2.25 mm
Powering Frequnecy	3.1–7 GHz	250 MHz
Down Link Rate	500 Mbps	150 Mbps
Up Link Rate	100 Mbps	2.5 MHz
Transmission Scheme	on off keying or binary shift keying	on off keying, amplitude key shifting
Device Size (um)	1000 × 800	2400 × 2200 × 300
CMOS Tech (um)	0.18	0.18
Chip Power (uW)	10,400	2.6
Working Distance (cm)	1	15
Testing	Testing with “fresh animal tissues” to mimic skin, fat, bone, and brain layers.	Testing PCB, power transmitted through 1 cm thick chicken breast.

Table 10.

Implantable System Specifications

Specification	Potential Solutions
Low-power, small footprint	Analog Circuit Design System-on-Chip Technology
Data acquisition	Flexible electronics Organic conductors Multi-modal data (e.g. neurotransmitter chemical sensors)
Data downlink and uplink	Radiofrequency Ultrasound Transducers and Piezo Beamforming Photovoltaic and Light Emitting Diode Magnetoelectronic Transducers Galvanic Coupling
Device Coordination	Information Theory Algorithms Device Identification Schemes On-chip processing
System Verification	Simulations Benchtop testing Ex-vivo testing Accelerated aging simulations In-vivo testing

4.1.1. System-On-Chip Microdevices

System-on-chip microdevices for neural interfacing, with micron-scale dimensions and low-channel counts, have been prototyped in academia. Although all devices reviewed in this class are in the far-term of translational readiness, they may be well positioned for diseases exhibiting highly distributed network-wide dysfunction that could benefit from chronic subthreshold neuromodulation, such as epilepsy [44] and neuropsychiatric disorders. Brain tissue activation would be achieved with multiple devices spread throughout and across structures. The use of multiple micro-devices would increase the degree-of-freedom in stimulation configurations and decrease immune responses and glial scarring [168]. Notable devices with stimulation capabilities include: Neurograin [169–175], Microbead [176–180], and StimDust [181]. Additionally, other microdevices have been developed exclusively for neural recording: Neural Dust [182,183], Free Floating Wireless Implantable Neural Recording (FF-WINeR) [184,185] and Microscale Opto-Electronically Transduced Electrodes (MOTEs) [186].

The deployment of numerous neural microdevices within the brain necessitates wireless power transfer and communication methods, as wired connections are impractical. Custom analog integrated circuit design enables low-power operation in neural microdevices while maintaining a compact footprint [176,177,181,183]. For example, FF-WINeR integrates an analog front end with digital serialization and power management circuitry to minimize telemetry and resonance tuning circuitry [187]. Transistor-based analog circuit design enables low power consumption, with MOTE at 1 μW [186]. Low-power operating points and relatively small stimulation amplitudes permit subthreshold neuromodulation applications.

Microdevices lack batteries and continuously harvest transmitted power, although Neural Dust includes an energy storage capacitor [183]. Microdevices rely on radiofrequency (RF), optical, and ultrasonic waves for wireless power transfer and communication. Power is transmitted using two or three coil [169,184] systems. Although adding a third coil enhances power transfer efficiency, it also increases the invasiveness of the system. Optical interfaces offer an alternative to coil-based power. The Optical Neurograin uses a GaAs/AlGaAs photovoltaic cell to harvest power from a 3.2 mW laser [188], while MOTE also uses optical power harvesting and communicates via LED [186]. However, optical scattering and absorption limits the depth of optical penetration depth. Neural Dust and Stim Dust systems incorporate a piezoelectric crystal connected to a custom transistor and two gold electrodes to achieve wireless power delivery and communication with ultrasound backscatter [181,182]. Ultrasound waves have shorter wavelengths and less tissue attenuation than RF, but face challenges with skull penetration, piezo miniaturization, and beamforming. Several beamforming algorithms have been developed for single-interrogator, multi-mote systems, including regularized linearly constrained minimum variance and delay-and-sum versions [182,189].

Effective communication strategies, grounded in information theory, are essential for deploying large-scale networks of microdevices. Additionally, achieving synchronized modulation across multiple sites requires a detailed understanding of how neural targets interact and contribute to disease pathology. Adjusting the settings of one lead can influence the neural activity at others, making precise programming challenging [190]. For example, coordinated reset and other complex stimulation patterns across nodes in a multi-site system require precise inter-node synchronization [190,191]. To address these challenges, several hardware solutions have been engineered. For example, the StimDust watchdog circuit controls data transmission and waits for the completion of downlink commands before initiating load impedance backscatter uplink [181].

Nurmikko and colleagues have conducted extensive research on communication schemes for neural technologies. Their Neurograin system uses a novel algorithm, “asynchronous sparse binary identification transmission”, where binary event data is transmitted with device specific identifiers [171]. To regulate stimulation parameters and initiation, a “daisy chain” protocol is used. Each chip receives instructions but only activates upon detecting its assigned three-bit sequence, which encodes one of seven pre-set stimulation commands [174]. A seven-bit identifier is implemented by selectively melting fuses. In this system, a shorted fuse completes a simple circuit, while an open fuse enables capacitor charging, generating distinct binary outputs representing permanent identifiers [188]. A Manchester decoding circuit enables precise modulation of pulse width stimulation [188]. An external neuromorphic spiking neural network decodes the transmitted event-based data [171]. Benchtop tests of the MOTE reveal pre-recorded neural spikes can be recorded, transmitted, and reconstructed [186]. The MOTE employs pulse-position modulation, where voltage amplitude is encoded in time between current spikes via an amplifier and relaxation oscillator circuit, which drives a charge pump circuit and LED.

The inventors of Neurograin and Microbead have developed specialized injection tools to enable precise targeting and minimally invasive implantation of devices. Khalifa and colleagues envision the Microbead as an injectable device in which cortical, peripheral, or deep brain stimulation structures can be targeted [179,180]. The injection system requires a 0.8 mm diameter Burr hole to deliver devices to depths ranging from 1.5 to 6.6 mm [179]. A custom implantation system has also been developed for the Neurograin. The implantation tool, an array of polyethylene glycol (PEG) holders, is fabricated with a polydimethylsiloxane (PDMS) mold and allows devices to be implanted to a maximum 1.4 mm depth [173]. Injectable implantation systems may allow for feasible surgery lengths, precise targeting, and minimally invasive deployment [157].

Benchtop and ex-vivo testing is conducted as a first step to verify designs. Communication with the Microbead was tested with a benchtop set-up of a 5 mm slice of raw beef, and in a water bath to demonstrate environments with high leakage currents [176]. For RR-WINeR, after simulated optimization of coil design, the design specifications were verified in an ex-vivo lamb head model [185]. Additionally, an accelerated aging analysis used an Arrhenius equation based “acceleration factor” to estimate that the encapsulation— vapor deposited parylene C followed by a dip in desiccated PDMS— would last 28.7 years [184]. Benchtop tests serve as a first step in characterizing power efficiency, communication schemes, and long-term stability of hardware.

Further validation requires in-vivo experimentation, and animal model studies play a crucial role in assessing microdevice performance during early development. For example, the Microbead system has been tested in rat sciatic nerve and induced muscle twitching [176,177]. Similarly, to test the Neural Grain, a relay and thirty devices were implanted in a rat model, and stimulation induced head angle changes and whisker oscillations [174]. Rats were trained to press one of two levers when it received stimulation, allowing researchers to evaluate the effects of different chip designs, stimulation frequencies, and pulse durations [174]. The Optical Neural Grain device was tested in rats shown to influence neuron firing rates. Although the device was directly powered by a laser during testing without skull or scalp obstruction, Monte Carlo simulations confirmed that optical transmission through biological tissue remains feasible [188]. Stim Dust was found to successfully stimulate rodent sciatic nerve [181]. Similarly, Neural Dust was benchmarked against wired “ground truth” measurements in the rat gastrocnemius muscle and sciatic nerve [182].

Material selection is critical to biocompatibility and microelectrode performance. Application-specific integrated circuits (ASIC) are fabricated on silicon substrates using standard photolithography and microfabrication techniques, ensuring consistent and reproducible production. Unlike the Microbead and Neurograin, which are self-contained on silicon chips, Stim Dust includes several components mounted on a flexible polyimide substrate [181]. Microdevices feature both on-chip and protruding electrodes, with electrode materials varying across designs. These include low-impedance inkjet “poly-ink” on Al + Zn/Ni/Au electrodes [176], planar PEDOT:PSS electrode with liquid crystal polymer encapsulation [169], tungsten wire with epoxy sealed connections [169], and PEDOT:PSS covered gold electrodes [181]. Material composition and electrode geometry directly influence impedance and stimulation amplitude. Continued advancements in materials science will be essential for optimizing distributed neural interface designs.

4.1.2. Milli-Meter Scale, Multi-Channel Devices

Unlike micro-devices, milli-meter scale systems with mid-term and near-term translational readiness have been developed: Nia Therapeutics and Cortec Neuro have run pre-clinical trials for their millimeter skull based IPG technologies; Motif Neuroscience’s dural stimulator has undergone pre-clinical testing; the Intracortical Visual Prosthesis (ICVP) and Neuralink N1 Implant are in clinical trials [192–196]. Larger size and power of devices in this class enable neural stimulation, relevant to a wide array of applications, including Parkinson’s Disease and other movement disorders. Furthermore, multi-channel devices have greater on-device processing capabilities and have already demonstrated on-implant machine learning end edge computing capabilities [193].

Compared to micro-devices, the hardware of millimeter sized devices reflects an increase in on-device processing capabilities. For example, the N1 Implant’s ASIC, operating at 6 mW, features hundreds of programmable analog amplifier 'pixels' that feed into analog-to-digital converter (ADC), serialization, and digitization circuits [157]. The ICVP incorporates a 5 MHz clock that drives an internal state machine, enabling real-time command execution and precise control of electrode drivers [197]. The unnamed magnetostrictive neural interface leverages transistor device variability and a “temporal majority voting” circuit to uniquely address individual chips using a physical unclonable function, all while operating with a single transmitter [198].

The ICVP consists of up to sixty-three wireless floating microelectrode arrays (WFMA) designed to stimulate the occipital cortex for visual restoration. Each WFMA incorporates an ASIC, an off-chip coil for wireless power transfer, and electrodes with adjustable surface area and length to optimize cortical stimulation [199]. A clinical trial of twenty-five devices in one patient has achieved progress in shape discrimination [200], light vs dark thresholding [201], direction of motion encoding [202] and mapping phosphene position [192]. Retinal prosthetics that aim to reproduce complex neural firing patterns to restore visual phosphenes require precise spatial and parameter control. The wireless and distributed design of the ICVP enables miniaturized, low-power devices to interact with multiple targets in the visual cortex, facilitating previously unattainable real-time neural decoding and functional mapping in freely moving subjects.

Millimeter-scale devices, like microdevices, use various wireless methods, with emerging magnetoelectric interfaces offering robust, energy-efficient communication. Neuralink’s N1 device uses near-field magnetic induction with two coils, one for wireless power and one for data uplink, while incorporating advanced materials and fabrication techniques [203]. The device, measuring 1–30 mm in diameter and up to 10 mm in height, is encased in silicone and hermetic glass for durability and biocompatibility. Neuralink holds patents for a resonant coil fabricated by rolling metal-coated polyamide and for flexible polymer electrode arrays wired to a skull-based transmitter for signal processing and wireless communication [203,204].

Not only devices with far-term translational readiness, but also one mid-term device undergoing pre-clinical trials, the Digitally Programmable Over-Brain Therapeutic (DOT) [205], implement magnetostrictive antenna. A magnetostrictive layer (e.g., Metglas) converts an alternating magnetic field into mechanical energy, which a piezoelectric layer (e.g., lead zirconate titanate (PZT)) then converts into electrical potential [206]. Magnetostrictive antennas offer robust performance under misalignment with minimal voltage loss and operate at lower frequencies, reducing specific absorption rate (SAR) and tissue heating compared to traditional methods [198]. The DOT device, an epidural neural stimulator, features a laminated Metglas–PZT magnetoelectric antenna connected to iridium oxide electrodes and a custom PCB, all encased in glass [205]. DOT was later adapted into the MagnetoElectric BioImplanT (ME-BIT), a within-ventricle “endocisternal interface” which replaces surface electrodes with catheter-based electrodes inserted through the ventricles to access spinal and brain targets [207]. In human cadavers and sheep studies, DOT effectively stimulated both cortical and deep brain structures [207]. A chronic study in pigs further evaluated its long-term stability and efficacy [205]. Additionally, cortical stimulation with DOT was achieved during human craniotomy. Another unnamed academic research prototype employs a similar antenna design [198]. In rat sciatic nerve, this device transmitted 3 cm with 9 uW [198].

An alternative communication scheme utilizes galvanic coupling, which enables electrical signal transmission through biological tissue. The BioBolt transcranial implant traverses the dura mater and skull, establishing direct skin contact to facilitate electrical signal transmission [208]. A charge balanced current is injected into the skin at frequencies above 100 kHz, enabling Galvanic communication with an external digital signal processing unit [208]. The dura facing surface of the device interfaces with a 16-channel flexible ECoG array composed of Parylene C with platinum electrode.

Analog circuit components enable safe operation, power efficiency, and reliable neural signal processing. To ensure safety of charge injection into “activated iridium oxide film electrodes”, the ICVP engineers employ current mirrors and an electrode simulation [209]. BioBolt, to decrease noise by 10% and power by 20%, includes a low power amplifier which uses the substrate body as a second “quasi-floating” gate [210]. To further improve power efficiency, the system uses a custom successive approximation register (SAR) analog-to-digital converter and a binary “boost signal” that indicates whether the quantized output falls within 0–0.5 or 0.5–1 V [210]. To optimize power, the Wireless Power Receiver on Chip (WiPRoC) incorporates a “multi-mode buck boost resonant regulating rectifier” (B2R3) that uses feedforward and feedback control to dynamically operate as a buck, boost, or hybrid rectifier, to ensure a stable DC power supply from a variable RF input [211]. Additionally, WiPRoC employs a fractal routing geometry, instead of a conventional loop, to reduce undesired RF coupling and further improve power use [212].

While many brain implants consider electrical recording and stimulation, chemical sensors may emerge as tools to close the control loop and inform stimulation patterns. The Empowering Next Generation Implantable Neural Interfaces (ENGINI) system has been designed to not only record electrical activity but also perform dopamine chemical sensing [213]. Fast scan cyclic voltammetry signals allow for subtraction of capacitive from faradaic current to estimate the neurotransmitter concentration. Niobium and gold interfaces were employed to prioritize chronic sensing applications [214].

Novel millimeter-scale device technologies— including communication modalities, coordination schemes, and analog circuit design— enable multi-channel stimulation command reception and real-time processing of electrophysiological and chemical brain signals at the edge, optimizing stimulation parameters for closed-loop neuromodulation. Human clinical trials with millimeter-scale recording devices provide translational experience and pave the way for future distributed brain stimulation systems.

4.1.3. System Portability

Next-generation implantable multi-site technologies rely on external devices for power delivery and communication. It remains crucial for all hardware to be portable and practical for patient use. Powering coils have been imagined as non-invasive external components, with or without invasive internal relays. However, system probability remains essential [215]. This section will review solutions for extracranial wireless communication modules (Figure 6).

Figure 6. Practical and Portable Brain Implant Wireless Module Designs.

Neurostimulation implants with far-, mid-, and near- term translational readiness have achieved wireless capabilities. However, patient usability and safety must be optimized in designs. Extracranial wireless modules which interact with implants inside the skull have been imagined as wearables, with wireless transmitter and receiver coils incorporated into (a) hat, (b) headband, or (d) headphone. (c) Extracranial wireless modules can be mounted on an arm band and connected to external wireless coils near the head. (e) Alternatively, sub-scalp devices can lie between the skin and skull to communicate with intercranial system-on-chip devices. Figure created with Figma.

Extracranial communication antennas and coils may be embedded in wearables, including hats, headphones, and headbands. The performance of a 5 MHz RF antenna “headband”, made of copper foil and ethyene-propylene-diene, was simulated [216]. A headphone mounted controller was engineered to interface with the WIMAGINE cortical ECoG [217]. CorTec’s Brain Interchange One includes an arm-worn processing unit wired to a magnetically attached skin module over the implant, enabling wireless power and data transfer [195]. Nia Therapeutics’ wireless Smart Neurostimulation System connects to an earpiece, similar to cochlear implant systems [193]. Despite technological advancements, patient convenience and comfort must be optimized for translation. To our knowledge, no publicly available research has assessed patient satisfaction with these configurations. Wireless communication modules must be designed to minimize inconvenience, considering factors like ergonomics, adaptability to daily activities, individual comfort, and hygiene, especially if regular wear or maintenance is required.

Sub-scalp devices, placed in the subgaleal space located between the scalp and the skull, provide an alternative to extracranial wearables. Such device placement has already been employed in recording devices [218] and transcranial stimulators [219]. Khalifa and colleagues propose a battery powered sub-scalp transmitter for powering neural implants [220]. Although the authors note further improvements are required in temperature management, miniaturization, and materials, the prototype suggests the sub-scalp space may provide a portable, minimally invasive area to interface with multiple implanted devices. Sub-scalp devices that directly connect to tablets or personal devices circumvent the need for additional hardware outside the patient’s body and may be minimally obstructive to patient activities. Additionally, such devices would circumvent problems created by wearables and headsets such as misplaced or lost components, misuse, or damage by the environment.

4.1.4. Wireless Communication Modules

While neuroscientists and physician teams await research- and clinical- grade wireless microdevices, bi-directional, wireless communication modules (Table 10) could allow for wireless communication to existing neurostimulators, MEA, and other electrode technologies. Communication systems, compatible with but lacking electrodes or recording hardware themselves, have been prototyped specifically for neural interface applications [221,222]. Such devices may promote multi-target system development in animals and short-term use in humans, but ultimately integrated systems will be necessary.

4.2. Magnetoelectric Nanoparticles

Magnetoelectric (ME) nanoparticles are injectable materials capable of wireless transduction of magnetic fields into electric stimulation to target deep brain. This technology, with far-term translational readiness, remains under development in academia but in the future may provide a mechanism for neuromodulation, like microdevices, which could benefit epilepsy and neuropsychiatric disorders. Unlike optogenetic technologies, ME nanoparticles do not require genetic modification to realize brain stimulation, although they are invasive and require an external field to transduce stimulation power. Current ME research includes small animal studies where magnetic fields are applied by coils surrounding the animals’ cages [223,224]. However, human applications will require fine spatial precision. An additional challenge is implantation, and proposed strategies include direct stereotactic injections, vascular delivery with magnetic steering, and ME-antibody conjugation for targeted transport [225,226].

Advances in ME materials science has led to magnetostrictive (CoFe2O4) and piezoelectric (BaTiO3) particles, rectangular shape of CoFe2O4BaTiO3 crystals, and core–double-shell (Fe3O4–CoFe2O4–BaTiO3) hexagonal magnetoelectric nano discs [223,224,227,228]. The ability of particles to stimulate cells is typically validated in neuron cell culture or small animals. Studies have demonstrated ME influence on mice running speeds with basal ganglion targets [224] and induction of reward-seeking behavior with ventral tegmental area targets [223]. Although still in early stages of development, ongoing research into nanoparticle materials and geometries will be critical in determining translational feasibility.

4.3. Non-Invasive Multi-Target Technologies

Three non-invasive technologies use transcranial energy interference to selectively target deep brain structures while minimizing surface and cortical stimulation. Unlike nanoparticles or implantable devices, these modalities require stationary setups, limiting ambulatory use. Nevertheless, their non-invasive nature, requiring no surgery, makes them promising alternatives for neuromodulation therapy. Since lacking an implanted component, non-invasive multi-target modalities could be used for neuromodulation in the cases of neuropsychiatric conditions or epilepsy and include therapy sessions like single-site transcranial magnetic stimulation therapies.

4.3.1. Transcranial Magnetic Stimulation

Transcranial Magnetic Stimulation (TMS) stimulates the brain with electric fields induced by magnetic fields. Traditional figure-eight coils are limited to single-target, cortical stimulation typically reaching only a few centimeters beneath the skull.

To overcome these depth constraints, magnetic temporal interference (TI) has been proposed. Magnetic TI allows for deep brain neurostimulation without undesirable cortical stimulation. The method has been implemented with solenoid [229] and Hesed [230] coil designs. In one instance, a Hesed coil, designed to target the nucleus accumbens, mitigated problematic cortical electric fields while maintaining deep electric fields at 4 to 6 cm [231].

Separate advancements in TMS have enabled multi-target cortical stimulation. For example, the ConnectToBrain project introduced their multi-site TMS (mTMS) system, capable of simultaneous multi-target stimulation [232]. Future research is needed to develop systems that integrate both deep stimulation via temporal interference and multi-site capability. Combining these abilities could offer a minimally invasive alternative to DBS, enable distributed neuromodulation without surgery, and reduce risks like hemorrhage and infection. Systems may enable neuromodulation and gradual monitoring and adjustment of side effects over repeated sessions. However, hardware challenges such as coil heating and bulky components limit current usability.

4.3.2. Transcranial Electrical Field Temporal Interference

Transcranial electric field TI applies alternating current at two frequencies to the scalp, creating an interference zone that targets deep brain structures. The method has been explored in safety studies, targeting the substantia nigra [233] and hippocampus [234]. While multi-target transcranial electric field TI stimulation has been proposed in theoretical protocols [235,236], to our knowledge simultaneous stimulation of multiple brain regions has yet to be demonstrated experimentally. A major challenge is that traditional methods require two electrodes per target, but physical space on the scalp is limited. To address this constraint, one group developed a two-electrode system capable of generating steerable, bi-modal field patterns by delivering two frequencies to each electrode [237]. However, like the traditional approaches, the bimodal stimulation technique limits targets to linear paths between electrode pairs. Future research should explore bimodal and custom-shaped electric field patterns that enable flexible, simultaneous multi-site neuromodulation.

4.3.3. Focused Ultrasound

Focused ultrasound (FUS), traditionally used to ablate dysfunctional tissue, can enable non-invasive neuromodulation without permanent structural damage when parameters remain below thermal thresholds. Clinical trials (Table 11) have demonstrated single-target deep brain excitation, and also inhibition, of electrical and metabolic activity with low-intensity FUS [238–245]. Many recent studies leverage the thin bone of the temporal acoustic window to mitigate skull attenuation, a key challenge in earlier research. While FUS neuromodulation remains in its exploratory stages, the low incidence of adverse events and absence of structural damage on MRI follow-up suggest its potential as a minimally invasive therapeutic.

Table 11.

Deep Brain Focused Ultrasound Neuromodulation, Human Clinical Studies

Target	Application	Outcome	Focal Length (mm)	Frequency (kHz)	Pulse Repetition Frequency (Hz)	Duty Cycle (%)	SPTA Intensity (W/cm2)	SPPA Intensity (W/cm2)	Mechanical Index	Thermal Index	Peak Negative Pressure (Mpa)	Acoustic Pressure (MPa)	Pulse Width (ms)	Session Time	Publication Identifier
inferior frontal gyrus	healthy controls			500	40	0.5	0.199		1.53					120 s	[245]
amygdala	healthy controls	fMRI BOLD signal of amygdala, functional connectivity	55 or 65	650	10	5	0.72	14.4					5	20 min, 30 s on / 30 s off	[242]
ventral intermediate nucleus, thalamus	essential tremor	neuropsychiatric tests		650	10		0.71973	14.39	0.75		0.61		5	eight, 10-min sessions, 30 s on / 30 s off	[244]
ventral capsule / ventral striatum	healthy controls	fMRI BOLD functional connectivity, arterial spin label perfusion	80	650	10	5	0.72	14.4				0.74	5	12 sessions, 20 min, 30 s on / 30 sec off	[243]
					125	5						0.74	0.4
					125	50						0.24	4
entorhinal cortex			55 or 65		100	5						0.68	0.5
hippocampus	temporal lobe epilepsy	seizure reduction		548	0.14	18 to 50	0.5 to 1.1				0.14 to 0.42			6 sessions, twice weekly	[241]
anterior cingulate cortex	chronic pain	patient reported pain measures		650	1.42	50	0.66	31	1.2	0.64			5	40 min sessions. 30 ms bursts	[247]
ventral intermediate nucleus, thalamus	essential tremor	accelerometer power		650		10						0.72	1	1 ms every 100 ms for 15 s cycles, repeated for 90 min sessinos	[239]
subcallosal cingulate cortex	drug resistant depression	fMRI BOLD signal of subcallosal cingulate cortex		650		0.8						1	30	30 ms ON, 4 ms OFF, for ten 2 min sessions	[240]

Blood Oxygenation Level Dependent (BOLD); Functional Magnetic Resonance Imaging (fMRI); Spatial Peak Temporal Average (SPTA); Spatial Peak Pulse Average (SPPA)

A first step towards simultaneous multi-site FUS, the Relative Through-Transmit (RTT) system, enables sequential targeting with real-time feedback to compensate for skull and tissue dephasing [238,240,246]. The system optimizes phase coherence by testing transducer-receiver pairs and applying singular value decomposition, overcoming inter-subject variability in skull attenuation. RTT FUS has been tested in applications for chronic pain [247], essential tremor [239], and major depressive disorder [240]. The authors leverage the system’s MRI compatibility to demonstrate metabolic changes using BOLD imaging, a strategy also used in other FUS studies [242]. Functional imaging could facilitate clinical translation by quantifying treatment efficacy, mapping interactions between stimulation sites, and enabling real-time adjustments in response to adverse effects. Beyond metabolic monitoring, imaging also enhances targeting precision. One group developed an image-guided fiducial tracking system with T1/T2 MRI registration to improve spatial accuracy [241]. While the RTT system allows flexible and precise targeting, it has not yet demonstrated simultaneous multi-target stimulation. This remains an open question for future research. As multi-target FUS technologies advance, robust targeting, adaptive feedback, and therapeutic monitoring could aid safe and effective translation of deep brain neuromodulation.

4.4. Computing Hardware for Distributed System Control

Edge computing and neuromorphic hardware offer the potential to enhance both current and future multi-target neural interfaces. Applications include neural stimulation triggered by predicted seizure propagation paths or individualized seizure and tremor biomarker decoding [248–250]. These methods may enable data-driven, closed-loop algorithms to dynamically regulate multiple control parameters based on complex, multi-site neural responses where both speed and low power consumption are essential. Emerging tree-based and biologically inspired hardware accelerators, that incorporate analog and in-memory computing, could further advance the efficiency and scalability of future neural interfaces.

Two on-chip edge computing systems, SCALO and NeuralTree, enable fast, low-power neural data analysis for brain computer interfaces. SCALO’s performance was validated for BCI prosthetic applications with a movement intention task, achieving 20 intentions per second from 384 electrodes [251]. Additionally, seizure propagation predictions were calculated with spectral correlations between 11 implants, achieving a false negative rate of 12.5% and a data rate of 506 Mbps; such seizure propagation path forecasts could be implemented into brain stimulation algorithms [250]. The SCALO system implements seizure projection, movement intention, and spike sorting with devices which can record, stimulate, and process data distributed among the brain [251]. Each device includes a separate uplink and downlink radio to connect nodes to each other, to processing elements, or external systems which allow distributed computing: template matching, machine learning, and hashing schemes. SCALO utilizes a dot product compression hashing scheme to optimize communication efficiency, in which portions of signals are transmitted and checked for similarity before the complete signal is sent [250]. The method reduces transmission by one hundred percent and has an error less than 8.5%, outperforming a microcontroller and non-hashing version.

The NeuralTree system connects 256 input channels for customizable feature extraction tasks including seizure detection and BCI movement intention [249]. The system was tested with two ECoG arrays in an epilepsy rat model, detecting seizures with 95.6% sensitivity and 96.8% specificity [249]. NeuralTree benefits from relatively low power (453 uW) and size (4x2 mm2 chip). Chip based AI utilizes TensorFlow with an Adam optimizer to minimize not only cross entropy loss but also a power consumption metric to further minimize power. In-memory computing stores features for biomarker extraction including line length, Hjorth activity, spectral energy, phase locking value, and phase amplitude coupling [249]

Analog hardware inspired by the biological brain, “neuromorphic” chips, utilize hierarchical routing methods and spike-based computing to process neural data. The Dynamic Neuromorphic Synchronous Processor (DYNAP) neuromorphic chip runs an “Address Event Representation” in which artificial neurons send their unique event-triggered identifier with a two-stage address organization (cluster-based and neuron-based) [252]. DYNAP performance in detecting high frequency oscillations and interictal epileptiform discharges from ECoG data was comparable to a digital algorithm, digital SNN, and expert reviewer [248]. Furthermore, the neurons of DYNAP represent the upward and downward character of these biomarkers, effectively compressing neural data for later reconstruction. The system achieved superior power usage (1.3 V supply voltage, 100 uW at a 100 Hz spike rate), size (43.79 mm2), and bandwidth (38 million events per second) [252]. Adaptive exponential integrate-and-fire artificial neurons are implemented in-silico with a biology-inspired pulse generator circuit and SRAM to store four different spiking patterns. Hierarchical routing enables efficient and low power signal processing on three levels (within core, between cores, and along a “long distance” mesh) [252].

The processing and control of non-linear dynamics across multiple brain regions will require advancements in both hardware and software, integrating system identification, control theory, and machine learning to enable precise and adaptive neural modulation. Current systems are validated on benchtop set-ups with remote or pre-recorded data to characterize device performance [248,249]. Further development remains necessary to integrate these chips into implantable systems for large animal testing and potential human translation. Recent works have already begun to address necessary specifications, which include high bandwidth wireless data transmission, device memory structure, and thermal considerations [250]. The three examples of on-chip hardware can process high dimensional neural data for biomarker detection and brain state classification. Edge computing and neuromorphic hardware can distribute computational loads across nodes, optimizing performance, portability, and power efficiency. By reducing reliance on large, centralized computing systems, these methods enable real-time therapeutic control in diverse naturalistic settings.

5. Software Solutions for Distributed Neurostimulation

As the advancement and translation of novel hardware will take time, we turn to supporting software and algorithms that can enhance the application of current systems and facilitate the future generation of devices. The complexity of brain networks and the influence of distributed neural stimulators may necessitate computational methods to optimize stimulation parameters and target locations. Researchers developing multi-electrode peripheral nervous system and retinal stimulators have faced similar challenges in location and parameter optimization, offering valuable insights. Additionally, machine learning algorithms have been specifically designed for single-lead DBS, demonstrating the potential for data-driven parameter tuning. Digital twins can integrate these functions into comprehensive software platforms, leveraging patient data to refine algorithms and provide recommendations for both current and future neuromodulation technologies.

5.1. Digital Twins and Biophysics Simulations with Patient Data Incorporation

This section will discuss biophysics simulations which explore the relationship between microscopic cellular activity and macroscopic local field and brain wide phenomenon. One form of biophysics simulations, digital twins, update model parameters based on real-word data feedback [253]. Such computational strategies may help uncover the brain’s nonlinear and non-intuitive responses to distributed stimulation.

Recent modeling frameworks have extended these concepts to predict brain-wide effects and guide personalized neuromodulation strategies. The Virtual Epileptic Patient (VEP) Project’s pre-surgical planning tool uses models at different scales to predict whole brain activity for the identification of seizure foci [254]. The model integrates patient-specific brain segmentation and structural connectomes from MRI into a forward-scale neural field model (NFM), a centimeter-scale neural mass model (NMM), and an inverse Bayesian estimator enabling efficient, electrodynamics-based prediction of brain tissue electric fields without the computational burden of finite element methods [254]. Another project, Cleartune, similarly incorporates patient imaging data into a finite element brain models and a surrogate optimizer to predict fiber activation and optimize electrode current amplitude [29,88]. Additionally, Cleartune embeds symptoms profile such as tremor, bradykinesia, and axial symptoms allowing for personalized treatment strategies [29]. Earlier, Buston and colleagues developed the concept of probabilistic stimulation atlas (PSA) in which biophysical models of volume of tissue activated (VTA) are optimized by patient specific imaging data and associated with clinical measures (symptoms reduction and adverse stimulation effects) [255–257]. Recent developments include simulations with realistic subcellular morphologies and different cell types, and the use of convolutional neural networks to bypass certain computably costly simulation steps [258–260].

Researchers have used patient-specific biophysics simulations to map activation patterns from single-target neurostimulation and assess modulation of pathological circuits, with potential for guiding multi-target control strategies. For example, a VTA simulation of single-target DBS for Tourette’s syndrome suggested that two different stimulation sites produced separate influences within the dysfunctional CSTC loop circuit [97,100]. With this knowledge, future studies could consider the simultaneous stimulation of both targets. Would the effects of the two targets be additive, or would a non-linear combination arise? Advanced simulations and future clinical research are needed to answer these questions. Nevertheless, digital twins and data-driven VTA simulations offer tools to estimate how different electrode targets and stimulation parameters relate to symptom improvement and side effects.

Biophysical models without incorporated patient data have been used to study multi-target stimulation for epilepsy [261], movement disorders [262], and schizophrenia [263]. To our knowledge models which recursively incorporate patient data have not considered multiple target neurostimulation cases. For example, a PSA analysis of Tourette’s syndrome excluded patients with multiple targets due to small sample sizes, highlighting the issue of the “chicken-and-egg” problem [264]. Digital twin systems may provide a tool to estimate associations between combinations of targets and parameters associated with symptom reduction and adverse stimulation effects. The first step could include combining data from dual target exploratory studies for PSA analysis.

5.2. Multi-Site Stimulation Configurations

In addition to supporting target and parameter decisions, biophysics simulations can be used to design stimulation configurations for multi-target systems. Distributed stimulation must be coordinated to achieve a specific goal, based on underlying network pathology. Tass and colleagues have developed coordinated reset stimulation (CRS), a novel scheme that uses delayed feedback across multiple targets [265,266]. Computational models suggest that CRS may desynchronize pathological oscillations while using less power than traditional DBS [266]. The authors model 1,000 artificial leaky integrate-and-fire neurons with spike timing–dependent plasticity, where synaptic weights strengthen when presynaptic spikes precede postsynaptic ones and weaken otherwise. A non-linear relationship between frequency (f) and number of stimulation sites (M) arose from competing influences on spike-timing-dependent plasticity (STDP) between f, M, and interstimulus interval, which was defined as $1/\left(M*f\right)$ [267]. Notably, the authors found that optimal CRS performance occurred at stimulation frequencies between 10–100 Hz, distinct from the dominant network oscillation frequencies. Moreover, networks with either few (M=2) or many (M=24) stimulation sites exhibited sustained desynchronization, whereas intermediate configurations (e.g., M=12) paradoxically led to short-term synchronization [267]. The results suggest that high-frequency CRS applied across numerous sites may act as stochastic perturbations or “noise,” weakening pathological synapses through STDP and promoting long-lasting network desynchronization at relatively low stimulation amplitudes.

CRS was studied in humans through a six-patient clinical trial targeting the STN for Parkinson’s disease, using two-hour sessions across three separate days [268]. Due to technological constraints, CRS was delivered across multiple DBS contacts within a single brain target rather than across multiple structures. Despite these constraints, multi-contact CRS stimulation was associated with a significant reduction in UPDRS scores and a beta power biomarker compared to baseline [269].

Similarly, animal studies of CRS have been constrained by hardware limitations, often restricting stimulation to multiple contacts within a single target rather than across distinct brain regions. Nevertheless, applications for epilepsy have been explored, with one study implementing a multi-electrode cortical array [270] and another applying four electrodes across different hippocampal regions [271]. The former found association between asynchronous, low frequency, distributed stimulation and greater seizure reduction when compared to high frequency or single macroelectrode stimulation [270]. The authors hypothesized that CRS facilitated population decoupling and seizure mitigation. The latter found CRS was associated with significant reductions in beta and gamma frequencies, increased theta modulation, decreased spike relative phase, and a decreased network connectivity [271]. Similarly, in substance use disorder models, CRS to four NAc DBS contacts desynchronized neural activity, delivered less charge, and correlated with reduced alcohol consumption [272].

Other multi-site configurations, such as delayed Local Field Potential (LFP) replay stimulation, have been tested with computational models [273]. The LFP Replay algorithm uses a filtered LFP signal to drive a harmonic oscillator that determines stimulation timing. A basal ganglia biophysics model predicted that both LFP replay and CRS improve thalamic output.

Zheng and colleagues have developed stimulation algorithms which employ graph theory metrics [274–276]. One scheme triggers stimulation when a threshold metric (line length, amplitude, or slope) is crossed in CA1, CA3, STN, or motor cortex; if multiple thresholds were met, stimulation followed a serial time-division approach [275]. In the follow-up study, the integration of Granger Causality improved the false positive [276]. A third study used Causal Flow, a metric of net directional connectivity, to account for variability in seizure networks across animals [274]. Ultimately, the authors found that, “combined stimulation with matched frequency could significantly decrease the duration of evoked electrographic seizures, which proved more effective than using single target stimulation” [276]. The authors suggest that seizure reduction with multi-site stimulation may stem from nodal inhibition blocking both input and output influences [275].

Distributed stimulation is limited by wired connections, few electrodes, external power, and external and offline processing. Future research may leverage edge and neuromorphic computing to address these issues. Accordingly, early work in human and animal models supports the need for coordinated stimulation configurations when employing multi-target stimulation.

5.2. Optimization Algorithms for Location and Parameter Optimization

It is important to note that being able to record from one or more brain regions does not necessarily indicate an ‘optimal’ site for stimulation. Clinicians are primarily outcome-focused, aiming to relieve symptoms rather than analyzing the electrophysiological properties of a given area. If stimulation achieves its therapeutic goal, the underlying neural activity may be considered secondary. This approach explains why DBS became a standard treatment for Parkinson’s Disease without initially relying on recording technologies. However, as neuromodulation expands to treat disorders with less well-defined targets or more distributed circuits, the ability to monitor neural activity could serve as a valuable tool. Recording may not always identify the best site for stimulation, but it can provide insights to optimize stimulation parameters, reduce side effects, and potentially guide future innovations in closed loop or multi-target systems. In addition to clinical expertise in the tuning of neurostimulation parameters, emerging machine learning algorithms (Table 12) play a role in future therapy management.

Table 12.

Optimization Algorithms for Clinical Brain Therapeutics.

Method	Method Description	Input	Output	Data Sources	Publication Identifier
Surrogate Optimization	Random and sparse parameter space search to optimize a constrained function.	Stimulation Current	Movement Disorder Symptom (Tremor, Bradykinesia, Axial Symptoms) Reduction	Finite Element Model of electrode current to fiber activation and symptom reduction.	[29]
Gaussian-Process based Bayesian Optimization	Tests inputs to develop a descriptive function and uses all previous attempts to determine the next test input.	Stimulation Frequency, Stimulation Charge Delivery, Stimulation Pulse Duration	Seizure Duration	Mice model test data of over 1000 parameter combinations.	[279]
Forward & Reverse Model-Free Neural Networks	Two neural networks are utilized. A measurement predictor network (MPN) learns a forward model from training data while a stimulus generator network (SGN) learns a reverse model from MPN data.	Recorded Retinal Electrical Receptive Field Response	Desired Electrode Stimulation Amplitude Pattern	The MPN is trained on electrical receptive field values from a microelectrode array. Random stimulation of the retinal generates testing data.	[285]
Greedy, Stochastic Learning Algorithm	Reduce error between a template neural response and recorded neural response. Minimize error with random selection of parameters around previous best value with an updating factor to prevent local minima.	Recorded Visual Cortex Pattern of Firing Rates	Visual Cortex Stimulation Amplitude Necessary to Match Desired Template	Template patterns were created from showing an animal a visual stimulus and recording neural responses.	[286]
Multi-Objective Particle Swarm Optimization	Searches the parameter space in parallel, moving “particles” based on velocity, particle optimum, and group optimum. Multiple objectives are considered by a “Pareto front” vectors in addition to a combined objective.	Electrode Position	Electrode contacts and current to maximize target pathways and minimize side effect pathways.	Multi-physics (COMSOL) and cable model (NEURON) software.	[278]

Various algorithms have been proposed to optimize stimulation across complex parameter spaces. One such optimization algorithm, the “inverse solution” approach, tunes parameters to reach a brain state predefined from recording data. For example, in a depression case study, dual-target DBS guided by sEEG recordings was used to adjust stimulation parameters toward brain activity patterns resembling non-depressed states [108]. Parameter combinations were tested and rank-ordered by similarity to the desired activity, streamlining the typically time-consuming trial-and-error process. Another algorithm, particle swarm optimization (PSO), supports both combined and objective-specific Pareto optimization, with the latter offering deeper insight into individual model objectives (Pena et al., 2018). One study employed PSO in a biophysical neuronal model to determine optimal DBS electrode current intensity with objectives of maximizing target activation while minimizing side effect area stimulation [277,278]. Gaussian-process Bayesian optimization has also been applied as a distinct approach to efficiently tune multiple stimulation parameters in neural interfaces. A seizure duration minimization model employed Bayesian regression to iteratively optimize stimulation frequency, charge delivery, and pulse duration in an in-vivo model (Bonizzato et al., 2023; Stieve et al., 2023). Future research may explore whether Bayesian optimization can be extended to simultaneously optimize both stimulation parameters and electrode location. Finally, graph theory-based optimization has used metrics such as clustering coefficient, betweenness centrality, and eigencentrality, derived from patient diffusion tensor imaging and MRI, to identify high-influence brain regions for stimulation in epilepsy, OCD, MDD, and movement disorders [280,281]. While clinical expertise remains essential in therapy management, adapting optimization algorithms can complement physician-driven decision-making by providing data-driven insights to refine and optimize distributed brain stimulation systems.

Network control theory (NCT) combines control theory, dynamical systems, and graph theory to model how external inputs can drive transitions between brain states by leveraging the structural connectivity of neural networks [282,283]. NCT represents connections between brain nodes as a matrix, informed by diffusion tensor imaging [284]. The method is agnostic to type of stimulation device with “control energy” representing the amplitude of stimulation. Such methods may provide tools to appropriately deploy distributed neurostimulation technologies through individualized mapping and treatment planning.

Engineers developing multi-site visual prosthesis have designed optimization algorithms that could be adapted for distributed deep brain neurostimulation. One study implemented a dual artificial neural networks framework where one network learned a forward model predicting retinal responses to stimulation, while the other learned an inverse model to optimize stimulation parameters for precise cell activation [285]. Another group optimized parameters for a 32-electrode visual cortex interface using a “greedy, stochastic learning algorithm” [286]. The model was trained by recording cortical activity in mice during presentation of natural visual stimuli. Following optimization, electrical stimulation was able to induce brain activity patterns closely resembling those evoked by natural vision. These examples from visual neuroprosthetics demonstrate the potential of multi-electrode parameter optimization techniques to control spatially distributed neural responses precisely. By adapting similar data-driven strategies, distributed neurostimulation systems, particularly in deep brain applications, could benefit from enhanced precision, improved therapeutic outcomes, and reduced stimulation-induced side effects.

Similarly, peripheral nerve interfaces, which often include a variable number of contacts, have driven the development of optimization algorithms to determine the number and location of electrodes. PSO has been applied to either maximize the number of stimulation sites or minimize the distance to fascicle centers [287]. Another work used a biophysics model of external field stimulation and the McIntyre-Richardson-Grill artificial neuron model to represent the median and sciatic nerves and determine the optimal number of implants [288]. The method considered the number of artificial fibers activated per electrode. A similar work bypassed computationally intensive biophysical simulations by using a multilayer perceptron to predict fiber activation from electrode position, with validation via Monte Carlo methods [287]. Biophysics simulations are only needed to train the algorithm. Multi-site stimulation outside the deep brain provides valuable insights into clinically relevant optimization algorithms for multi-electrode neuromodulation. However, significant differences between the deep brain, peripheral nervous system, and retina highlight the need for future research to adapt and refine these methodologies for distributed deep brain neurostimulation applications.

6. Future Outlook

Distributed deep brain neurostimulation seeks to leverage tens to thousands of stimulation sites in an effort to achieve network-level modulation by dynamically adjusting dysfunctional synchrony underlying complex and, in many cases, treatment-resistant diseases. Growing evidence across neurological and neuropsychiatric conditions supports the view that these disorders frequently stem from network-level dysfunction rather than from isolated abnormalities in a single brain region. Small off-label human studies using multi-target deep brain stimulation reveal that this approach may mitigate symptoms while minimizing adverse effects.

This strategy would require coordinated stimulation in which precise adjustments in amplitude, frequency, and pulse width across multiple targets are processed in real-time. However, current neurostimulation technologies are largely limited to a small number of stimulation sites with static frequency settings per lead, underscoring the need for advancements in adaptive, multi-target neuromodulation technology. Cross-disciplinary collaboration, and academia-industry partnerships are necessary to bridge this gap. Physicians can continue to explore target combinations, quantify patient outcomes, and provide insights to system specifications for specific disease applications; specifically, they can provide larger and randomized controlled clinical trials, and meta-analysis on risks and benefits of the multi-target approach using current and near-term technologies. Academic neural engineers are well positioned to formulate solutions to fundamental design challenges, with realistic prototype validation; to address cell-level biology and immunology questions relevant to device interfaces; and to design advanced data processing algorithms for system control. Industry can use this knowledge to develop full, clinical-grade systems, which prioritize patient safety; partner with surgical teams to lead pre-clinical and clinical trials of near-term technologies; fulfill regulatory requirements; and scale products for patient translation.

Patient safety remains a key barrier to clinical translation, and a central priority. Nevertheless, emerging technologies— including micro-devices, millimeter-scale implants, nanoparticles, and non-invasive methods— may collectively be used to overcome these limitations. System portability could be enhanced through sub-scalp devices and edge computing hardware. Moreover, as stimulation systems grow in complexity, computational tools will likely be necessary to support clinical use, safety, and ongoing monitoring. Distributed deep brain neurostimulation would enable patterned stimulation and expanded degrees of freedom. Current research strongly supports its potential to address more specific and less common symptom profiles. Multi-target approaches have the potential for improving care for individuals with neurological and neuropsychiatric conditions unresponsive to other treatments.