Stimulation Localization and Electric Field Modeling

Many open questions remain regarding the precise location of optimal stimulation. DBS is known to exert direct bioelectric effects on both neuronal cell bodies and axonal populations; however, the contribution of each to efficacy and side effects remains unclear. ^{Reference Jakobs, Fomenko, Lozano and Kiening6} Regardless, millimetric differences in stimulation appear to influence stimulation outcome. ^{Reference Elias, Boutet and Joel7} Commonly the stimulation “sweet spot” is identified using group-level data by (1) determining the precise location of implanted electrodes, (2) determining the area of stimulation (volume of tissue activated: VTA) and (3) assessing the relationship with clinical outcome measures.

Several software tools have been developed for localization of implanted electrodes. ^{Reference Horn, Li and Dembek8–Reference Lauro, Vanegas-Arroyave and Huang12} These employ various algorithms that reconstruct the desired electrode from the artifact visible on postoperative CT or MRI. ^{Reference Horn, Li and Dembek8,Reference Neudorfer, Butenko and Oxenford10,Reference Milchenko, Snyder and Campbell13–Reference Husch, V. Petersen, Gemmar, Goncalves and Hertel15} This reconstruction is then registered to the preoperative high-resolution MRI for visualization. The introduction of directional and segmented electrode (presently up to 16 channels per lead) technology has complicated this process. ^{Reference Krauss, Lipsman and Aziz16} Establishing the short-axis orientation of the electrode is required to fully utilize the directionality and steer the stimulation. Evidence suggests there is continued rotation of the electrode mainly during the first few weeks after implantation. ^{Reference Dembek, Hoevels and Hellerbach17} Efforts are currently being made to optimize and validate algorithms that can reconstruct this orientation from the artifact of directional markers. ^{Reference Dembek, Hoevels and Hellerbach17–Reference Dembek, Hellerbach and Jergas19}

Once the electrode is localized, modeling an accurate electric field representative of the stimulation (i.e., VTA) has been a persistent challenge. ^{Reference McNeal20–Reference Åström, Diczfalusy, Martens and Wårdell22} The “gold standard” for computational prediction of neurostimulation, developed in 1976, couples the electric field data to multi-compartmental neuron models; however, it is too computationally intensive for practical clinical application. ^{Reference McNeal20} There have been considerable efforts spent attempting to simplify the method of calculating the spatial extent of the electric field while accounting for varying electrical stimulation parameters. ^{Reference Chaturvedi, Luján and McIntyre21,Reference Rattay23–Reference Butson and McIntyre26} However, modeling how the field is influenced by surrounding tissue properties remains a considerable hurdle. Recently, finite-element modeling has been applied to neuron compartment models to manage the complexities of axonal activation during extracellular stimulation. ^{Reference Åström, Diczfalusy, Martens and Wårdell22} These models attempt to display a stimulation field that is relevant to the therapeutic effects of stimulation. However, the threshold of axon activation is dependent on its size. ^{Reference Ranck27} Moreover, which axons are responsible for therapeutic effect is not known. ^{Reference Ranck27} Furthermore, axonal polarity is determined by the potential difference between adjacent axons; thus, the entire distribution of polarity must be considered. Finally, these models fail to account for stimulation frequency, which is known to influence therapeutic effect. Simplified models are available for visualization; however, they are understood to be a limited representation of the true electrophysiology.

Together with the localization of electrodes and estimation of the VTA, retrospective group-level analyses can identify stimulation “sweet spots” associated with improved outcomes. These probabilistic maps are generated by transforming the VTA to a common space and statistically weighting the location by clinical outcome. ^{Reference Elias, Boutet and Joel7,Reference Dembek, Baldermann and Petry-Schmelzer28} The resulting maps delineate zones of response in a voxelwise data-driven manner. By this method, it has been possible to pinpoint the most efficacious substrate to stimulate across large cohorts of patients and refine future targeting. ^{Reference Elias, Boutet and Joel7} Direct relationships between stimulation location and clinical improvement have been established for Parkinson’s disease (PD), ^{Reference Horn, Wenzel and Irmen29–Reference Horn, Reich and Vorwerk33} dystonia, ^{Reference Neumann, Horn and Ewert34–Reference Schönecker, Gruber and Kivi36} essential tremor (ET) ^{Reference Al-Fatly, Ewert, Kübler, Kroneberg, Horn and Kühn37–Reference Germann, Santyr and Boutet39} and obsessive-compulsive disorder, ^{Reference Baldermann, Melzer and Zapf40} among others. ^{Reference Elias, Boutet and Joel7} These associations have shown some advantages in improving the efficiency of DBS programming. In the case of subthalamic nucleus (STN) stimulation in PD, effective symptom control has been associated with stimulation at contacts around the dorsolateral border of the STN. ^{Reference Dembek, Roediger and Horn32} Despite accurate lead placement, optimization of stimulation settings requires in-depth evaluation and individualization of each setting. Modern DBS systems with more sophisticated designs have introduced techniques to shape and steer the electric field and increase the therapeutic window but also the burden of clinical programming. ^{Reference Dembek, Reker and Visser-Vandewalle41} Several recently developed data-driven tools utilize electrode localization information to suggest stimulation settings that optimize stimulation location at the established “sweet spot.” ^{Reference Waldthaler, Bopp and Kühn42,Reference Roediger, Dembek and Achtzehn43} These studies have shown improved DBS programming efficiency and non-inferior clinical results compared to standard clinician-optimized settings. ^{Reference Waldthaler, Bopp and Kühn42,Reference Roediger, Dembek and Achtzehn43}

The association between stimulation location and clinical improvement, however, must be interpreted cautiously. The VTA, in particular, remains a visual approximation of the presumed electrical field based on theoretical models and often ignores local impedance changes and intrinsic dynamics of neuronal populations. ^{Reference Krauss, Lipsman and Aziz16} Crucially, these findings lack large-scale prospective validation, and different published approaches result in a variety of different “sweet spot” localizations when used with the same data set. ^{Reference Dembek, Baldermann and Petry-Schmelzer28} Despite efficiency advantages, thus far, no VTA-based programming has been found to be superior to programming based on clinical grounds.

Normative Structural and Functional Connectivity

There has been a shift in focus from what is being stimulated at the local level to how stimulation exerts its effects at a network level. Accumulating evidence opposes the long-held belief that DBS exerts its effects via local modulation of the discrete gray matter nuclei in which they are implanted. ^{Reference Horn, Reich and Vorwerk33,Reference Horn and Fox44–Reference See and King46} Instead, suggesting that modulation of distributed brain networks is at least equally important for optimal outcomes. ^{Reference Lozano and Lipsman47,Reference Accolla, Herrojo Ruiz and Horn48} To appreciate the engagement of distributed brain networks, specialized MRI sequences are required. Specifically, diffusion-tensor imaging (DTI) models the structural connectivity (i.e., the white matter tracts physically connecting regions) by interpreting the diffusion of water constrained along the long axis of axons. Conversely, functional MRI (fMRI) acquired at rest measures low-frequency blood oxygen level-dependent (BOLD) signal oscillations between regions and establishes connectivity between covarying regions (i.e., functional connectivity). Both measures provide complementary, indirect information about the architecture of brain-wide networks. Through mathematical models, the networks of connected regions provide a theoretical foundation for neurobehavioral phenomena, termed connectomics. ^{Reference Sporns49}

Connectomes may be studied either between individuals using patient-specific DTI or fMRI or, more commonly, using aggregate group data from large cohorts of subjects. The latter approach generates a presumed generalizable average of brain connectivity, known as the normative connectome (Figure 1A). ^{Reference Fox50} The major advantage of using a normative connectome is the high quality of the data and its flexibility of use in populations where the necessary imaging (DTI or fMRI) may not be available. This is often the case in clinical environments, where advanced imaging acquisitions are not routine, costly and often reserved for specific research scenarios. Furthermore, for rare conditions that rely on collaborative efforts for subject recruitment, comparison of images acquired using different hardware and acquisition parameters may prohibit analysis. Normative connectivity is of particular interest in DBS subjects where the acquisition of high-resolution MRI has been challenging due to safety concerns related to the implanted metallic hardware. Normative connectivity provides the means to investigate network-related questions in large, previously inaccessible groups of patients, often retrospectively.

The major shortcoming of normative connectomes is their inability to capture the subtle but important differences specific to an individual. Fundamental differences exist between the brains of healthy individuals and the brains of those affected by neurological disease. ^{Reference Bassett and Bullmore51} For example, the sample used to define the human connectome did not include subjects older than 40 years. ^{Reference Elam, Glasser and Harms52} Therefore, a connectome derived from healthy subjects may not accurately depict the disease-specific circuitopathies influencing the outcome being studied. ^{Reference Klobušiakova, Mareček, Fousek, Výtvarova and Rektorova53,Reference Sala, Caminiti and Presotto54} As a compromise, it may be advantageous to create disease-specific connectomes using scans of individuals with the same disease as the population of interest. ^{Reference Wang, Akram and Muthuraman55} Further, matching the connectomes for other covariates, including age, sex or disease duration/severity, may improve the accuracy of findings. However, with each subdivision, the number of subjects with available images for connectome generation diminishes. It has been suggested that a minimum of 150–200 subjects is necessary for the stabilization of group connectivity estimates, thereby constraining the ability to generate disease-specific connectomes. ^{Reference Cohen and Fox56} A recent PD-specific functional connectome was created from 75 subjects undergoing DBS for PD, offering an alternative to normative connectomes capable of visualizing unique differences not seen in a healthy normative connectome. ^{Reference Loh, Boutet and Germann57} A few studies have directly compared the results from disease-specific and healthy control-derived connectomes with conflicting overall findings. ^{Reference Horn, Reich and Vorwerk33,Reference Wang, Akram and Muthuraman55,Reference Loh, Boutet and Germann57–Reference Germann, Elias and Boutet59} Larger and better-matched disease-specific connectomes will likely provide clarity.

The field of neuromodulation has leveraged normative connectomics to study networks mediating therapeutic efficacy (Figure 1A). ^{Reference Wang, Akram and Muthuraman55} The predictive ability of normative connectivity profiles was investigated. ^{Reference Horn, Reich and Vorwerk33} From the VTAs of 51 PD patients with STN stimulation, a connectivity profile was generated based on the human connectome project data (DTI and fMRI) to identify connections reliably associated with clinical motor improvement. This connectivity profile was used to predict the motor outcome of an out-of-sample cohort with 12–20% of variance explained. This concept has since been repeated for several other conditions with similar results. ^{Reference Baldermann, Melzer and Zapf40,Reference Horn and Fox44,Reference Horn, Kühn, Merkl, Shih, Alterman and Fox60,Reference Hollunder, Ostrem and Sahin61} Recently, small prospective studies have applied these findings in PD. Rajamani et al. developed an algorithm capable of suggesting optimal stimulation settings, which maximize engagement of tracts associated with improvement of four symptom domains (tremor, rigidity, bradykinesia and axial instability). ^{Reference Rajamani, Friedrich and Butenko62} In a preliminary analysis of five patients, this algorithm was able to suggest settings with comparable symptomatic improvement compared to standard-of-care determined settings. ^{Reference Rajamani, Friedrich and Butenko62} Similarly, in nine patients, Hines et al. demonstrated that prospective automated connectomic programming was safe and generated therapeutic effects similar to traditional programming strategies. ^{Reference Hines, Noecker and Frankemolle-Gilbert63} While promising, these studies are underpowered, and overinterpretation should be avoided.

As exemplified here, connectomes derived from large, high-quality datasets utilizing specialized MRI hardware ^{Reference Setsompop, Kimmlingen and Eberlein64,Reference Elias, Germann and Joel65} may be considered analogous to the anatomical reference atlases widely used in neurosurgical planning. ^{Reference Ewert, Plettig and Li58,Reference Schaltenbrand and Wahren66,67} Both feature high-fidelity information insensitive to the subtleties of patient-specific anatomy. As in traditional reference atlases, normative connectomes should serve as a guide to inform patient-specific imaging. ^{Reference Coenen, Schlaepfer and Varkuti68}

Electrophysiology-Guided Closed-Loop Stimulation

Advancements, including current steering, improved direct targeting and leveraging large normative datasets, offer the potential to enhance DBS therapy by refining the stimulation target. Despite this, PD patients still experience cognitive and speech impairment, suboptimal gait and axial motor control and residual fluctuations between on/off medication states related to stimulation. ^{Reference Zibetti, Merola and Rizzi69,Reference Rizzone, Fasano and Daniele70} Furthermore, limited battery life subjects these patients to subsequent exchange procedures. ^{Reference Almeida, Rawal and Ditty71} A promising advancement to overcome these DBS drawbacks is closed-loop or adaptive DBS (aDBS), where stimulation parameters are continuously modulated based on a relevant biomarker (Figure 1B). ^{Reference Neumann, Gilron, Little and Tinkhauser72} This provides the correct stimulation only when needed, lessening side effects and long-term habituation while possibly preserving battery life. A simple example of this work is position-sensitive spinal cord stimulation for pain, where the voltage is adjusted according to body position using an accelerometer. ^{Reference Schultz, Schultz and Webster73} So far, aDBS has been primarily studied in epilepsy using causal neural signals to change the stimulation before symptoms emerge. ^{Reference Hoang, Cassar, Grill and Turner74}

Adaptive DBS in PD has focused on local field potentials (LFPs) recorded from implanted electrodes (Figure 1B). The spectral power in the beta frequency range (11–30 Hz) at the STN or globus pallidus internus (GPi) has been correlated with bradykinesia and rigidity. ^{Reference Little, Pogosyan and Neal75–Reference Tinkhauser, Pogosyan and Debove77} Averaging the beta activity over long periods allows the DBS power to be modulated based on the dynamics related to drug therapy and motor on-off states, leading to a reduction in battery usage by ∼ 50% and a reduction in on-state dyskinesias. ^{Reference Rosa, Arlotti and Marceglia78} Conversely, capturing the bursting nature of spontaneous beta activity can identify states of severe bradykinesia and rigidity. ^{Reference Deffains, Iskhakova, Katabi, Israel and Bergman79} Using the beta signal to trigger or increase stimulation to terminate the pathological signal has achieved a 50% reduction in power requirements and a reduction in adverse effects on speech compared to conventional continuous stimulation. ^{Reference Little, Pogosyan and Neal75,Reference Little, Beudel and Zrinzo80,Reference Little, Tripoliti and Beudel81} Better control of bradykinesia and rigidity has also been reported with beta signal-based aDBS. ^{Reference Arlotti, Marceglia and Foffani76} The arrival of commercially available DBS devices with simultaneous neural sensing and stimulation capabilities has enabled wide-scale study of aDBS in chronically implanted patients. ^{Reference Jimenez-Shahed82,Reference Fung, Sumarac and Sorrento83} Opportunities to develop more sophisticated aDBS approaches using machine learning based on the precise correlation between long-term electrophysiological signal recording and symptom severity are arising across a variety of diseases.

Importantly, LFP-based open-loop programming is already largely adopted due to the availability of DBS devices able to wirelessly stream from the brain (Medtronic Percept PC and RC, Dublin, Ireland). Multiple lines of evidence have confirmed that stimulating at the contact with the strongest LFP (e.g., beta in PD, alpha in ET) is associated with the best outcome. Likewise, the reduction of LFP with ongoing stimulation can be used to confirm the proper contact selection and identify the amount of stimulation needed to obtain such benefit. ^{Reference Arlotti, Marceglia and Foffani76} This latter approach is particularly favorable in conditions with a delayed clinical effect of stimulation, such as dystonia. ^{Reference Picillo, Lozano, Kou, Munhoz and Fasano5} Notably, the same device is capable of closed-loop stimulation using three different control algorithms: single threshold, single inverse threshold and dual threshold. The closed-loop capability is presently available only in Japan until approval from other health authorities is granted.

Individual Functional Response to Stimulation

As our understanding of the mechanism of DBS grows deeper, it is clear that although stimulation acts on the local region around the electrode, the therapeutic effects and side effects that manifest are the result of brain-wide network engagement. ^{Reference Lozano and Lipsman47,Reference Accolla, Herrojo Ruiz and Horn48} As discussed, this engagement may be investigated using the connectivity of the stimulation site to the rest of the brain using normative connectomes. ^{Reference Fox50} Although applicable to large groups of varied data, this method is limited by assumptions that patient-specific circuitopathies are accurately represented by normative data, ^{Reference Bassett and Bullmore51,Reference Horien, Shen, Scheinost and Constable84} registration and localization errors of estimated VTAs are minimal, ^{Reference Lofredi, Auernig and Ewert85–Reference O’Gorman, Jarosz and Samuel88} and the modeling of electrical stimulation is a representation of true electrophysiology. ^{Reference Krauss, Lipsman and Aziz16} Conversely, the acquisition of neuroimaging data demonstrating an individual’s response to stimulation has provided key insights into the mechanism of DBS and has potential clinical application as a personalized biomarker for stimulation success. ^{Reference Boutet, Jain and Gwun89,Reference Boutet, Madhavan and Elias90}

MR-related adverse events and safety implications

MRI scanning of implanted metallic DBS systems is subject to stringent safety guidelines, which have restricted its use. ^{Reference Medtronic103} In 2009, it was reported that close to 50% of centers surveyed did not use MRI in patients with implanted DBS. ¹⁰⁴ MRI restrictionsss arise from safety concerns mainly related to the potential for radiofrequency (RF) induced heating of the metallic hardware during the scanning (Figure 2). ^{Reference Rezai, Lozano and Crawley96–Reference Rezai, Phillips and Baker99}

Figure 2. Summary of risk factors contributing to DBS device heating. Intrinsic and extrinsic risk factors are listed and labeled on a depiction of an implanted DBS device (A) and an illustration of an MRI suite (B). B_1+RMS = root-mean-square value of MRI effective component of RF magnetic (B ₁) field; IPG = implantable pulse generator; RF = radiofrequency; SAR = specific absorption rate. Reproduced with permission from Boutet et al. Radiology 2020. ^{Reference Boutet, Chow and Narang102}

There are five reports of injury involving RF currents in the DBS literature, three of which are MR-related. ^{Reference Boutet, Chow and Narang102} This prompted the US Food and Drug Administration (FDA) to issue a warning regarding the MR imaging of patients with DBS devices in 2005 (Figure 3A). ^{Reference Gupte, Shrivastava, Spaniol and Abosch97} This led to strict vendor guidelines including restricting the field strengths (e.g., 1.5T) and limiting parameters related to the amount of RF being delivered (e.g., specific absorption rate [SAR], root-mean-square value of the MRI effective component of the RF magnetic [B₁] field [B1+rms]). Between 2005 and 2011, several human studies aimed to confirm the safety of the FDA guidelines, with no adverse events reported in over 3000 DBS patients scanned at 1.5T (Figure 3A).¹⁰⁴ As discussed above, there is considerable interest in investigating DBS patients outside of these restrictive guidelines, including using fMRI during active stimulation. As a result, recent studies in phantoms have examined the safety of these systems outside vendor guidelines, reporting acceptable temperature rises under local experimental conditions (Figure 3A).^{Reference Kahan, Papadaki and White98,Reference Bhidayasiri, Bronstein and Sinha105–Reference Hancu, Boutet and Fiveland110} This was further expanded in vivo in a publication where 102 subjects with fully internalized DBS were scanned with 3T structural T1-weighted and fMRI outside vendor guidelines, confirming safety.^{Reference Boutet, Rashid and Hancu111} Importantly, these findings may not be generalizable to different MRI or DBS systems, and local institutional safety testing should be performed until generalizability can be confirmed.^{112,Reference Baker, Tkach, Phillips and Rezai113}

Figure 3. Graphs depicting MRI-DBS-related publication over time. (A) A line graph representing the cumulative number of DBS-related MRI safety studies published from 1992 to 2019. The studies were categorized into phantom, animal, human and technique safety studies. Modified with permission from Boutet et al. Radiology 2020. ^{Reference Boutet, Chow and Narang102} (B) A line graph representing DBS-fMRI studies over time. The rate of publication increasing, particularly in the last 5 years. Modified with permission from Loh et al. Brain Stim. 2022. ^{Reference Loh, Gwun and Chow118} DBS = deep brain stimulation; FDA = Food and Drug Administration; fMRI = functional MRI; T = tesla.

Advancements in DBS hardware and neuroimaging techniques have partially overcome the challenges of performing MRI in patients with implanted and active DBS systems. However, heat at the electrode tips remains the primary risk associated with DBS devices. Electrode and device designs incorporating components to act as heat sinks and improve insulation have been studied. ^{Reference Zhao, Liu and Xu114,Reference Elwassif, Datta, Rahman and Bikson115} Other studies report modification to RF coils, resulting in lower SAR and heating and custom MRI acquisition parameters to limit the use of RF pulses. ^{Reference Boutet, Chow and Narang102} The progressive development in this field has led to the evolution of vendor guidelines and considerable growth in the number of publications (Figure 3B). ^{Reference Boutet, Chow and Narang102} Select modern DBS devices have been deemed conditional at 3T and for whole-body scanning. ^{Reference Boutet, Chow and Narang102} When imaging outside of published guidelines, emphasis should be placed on minimizing SAR and B1+rms. ^{Reference Boutet, Chow and Narang102} The current best practice guidelines based on DBS vendor recommendations, reviewed literature and institutional experience are summarized in Figure 4.

Figure 4. Summary recommendations of best practices for MRI in patients with DBS devices. These recommendations are based on guidelines from DBS vendors, reviewed literature and institutional experience. B_1+RMS = root-mean-square value of MRI effective component of RF magnetic (B ₁) field; DBS = deep brain stimulation; RF = radiofrequency; SAR = specific absorption rate. Reproduced with permission from Boutet et al. Radiology 2020. ^{Reference Boutet, Chow and Narang102}

A modern era of functional imaging in response to DBS

Improved understanding of MRI-DBS safety has facilitated a growing number of studies utilizing fMRI to provide unique mechanistic and therapeutic insights. ^{Reference Boutet, Hancu and Saha106,Reference Hancu, Boutet and Fiveland110,Reference Boutet, Rashid and Hancu111,Reference Sui, Tian and Ko116,Reference Davidson, Tam and Yang117} As of 2022, there were 37 studies investigating fMRI during active DBS, with over half published within the last 5 years (Figure 3B). ^{Reference Loh, Gwun and Chow118} The majority of these studies focused on STN-DBS in PD; perhaps more notable was the paucity of GPi-DBS studies. ^{Reference Loh, Gwun and Chow118} Encouragingly, emerging and experimental indications for DBS are represented in fMRI-DBS studies, reflecting the utility for characterizing the effects of stimulation, refining patient selection and targeting. ^{Reference Loh, Gwun and Chow118} These studies have demonstrated the large-scale networks modulated by DBS and the variation in stimulation-evoked response related to the stimulation site, stimulation parameters, patient characteristics and degree of treatment efficacy. ^{Reference Loh, Gwun and Chow118} Reviewing this literature suggests a movement beyond characterizing just the brain response to stimulation but increasingly making associations between these responses and clinical outcomes. ^{Reference Loh, Gwun and Chow118} This hints at the translational potential of fMRI in DBS, particularly in identifying biomarkers of therapeutic success. As discussed, the process of stimulation parameter optimization is burdensome and especially difficult in instances with delayed clinical feedback, such as dystonia, depression and Alzheimer’s disease. ^{Reference Picillo, Lozano, Kou, Puppi Munhoz and Fasano4,Reference Picillo, Lozano, Kou, Munhoz and Fasano5} fMRI during active DBS appears to be an accessible and rapidly acquired objective marker with the potential to guide programming (Figure 1C). ^{Reference Boutet, Madhavan and Elias90} Furthermore, the fMRI response to DBS has been correlated with long-term motor improvement, suggesting intraoperative fMRI may help refine electrode placement.

A study using a large cohort of PD-DBS patients found that fMRI acquired prospectively at 3T demonstrated a characteristic pattern of brain response to clinically optimal stimulation. ^{Reference Boutet, Madhavan and Elias90} Specifically, over 200 fMRI sessions were conducted in 67 PD-DBS patients during active cycling stimulation. The sessions were acquired in both clinically optimal and nonoptimal settings by adjusting the contact and amplitude (Figure 5). Comparing the stimulation-dependent functional response obtained during these sessions allowed the training and validation of a machine learning classification model able to identify optimal stimulation settings with a 76% out-of-sample testing accuracy. ^{Reference Boutet, Madhavan and Elias90} This model relies on features within 16 regions throughout the brain, one of the more prominent being the decrease in stimulation-dependent BOLD signal in the ipsilateral primary motor cortex in addition to thalamic and cerebellar changes observed (Figure 5). These findings, central to the classification model, are consistent with previously reported imaging and electrophysiological work. ^{Reference Jech, Urgošík and Tintěr100,Reference Loh, Gwun and Chow118–Reference Hesselmann, Sorger and Girnus124} At the network level, it appears that both DBS and levodopa attempt to normalize the PD-related spatial covariance pattern. ^{Reference Asanuma, Tang and Ma125,Reference Moeller, Nakamura and Mentis126} Interestingly, inspecting the pattern of engagement from nonoptimal stimulation shows the involvement of non-motor circuits, including the visual cortices and operculum. ^{Reference Boutet, Madhavan and Elias90} This is likely a manifestation of target-adjacent stimulation. As the location of stimulation moves away from the dorsolateral STN, surrounding white matter tracts and associative/limbic territories of the STN may be recruited. ^{Reference Hollunder, Ostrem and Sahin61,Reference Vanegas-Arroyave, Lauro and Huang127,Reference Alexander, DeLong and Strick128} These consistent neuroimaging findings suggest a common neuroanatomical network mediating therapeutic improvement, and engagement outside of this network possibly explains the presence of adverse events.

Figure 5. (A) Experimental design for postoperative DBS contact and voltage screening using fMRI. fMRI is acquired on each contact and a range of clinically relevant voltages. The resulting images are analyzed using a machine learning classification model, and the most optimal settings tested are identified. The model identifies a pattern of network engagement specific to stimulation at the clinically optimized contact (visualized in panel B) and voltage (visualized in panel C). DBS = deep brain stimulation; fMRI = functional MRI. Modified with permission from Boutet et al. Nat. Comm. 2021. ^{Reference Boutet, Madhavan and Elias90}

A central finding in fMRI studies during active DBS is that changes to stimulation (e.g., cycling the device on/off) will manifest in rapid fMRI responses, preceding changes in symptoms that may take hours/days to reappear. ^{Reference Boutet, Madhavan and Elias90,Reference Elias, Germann and Boutet129–Reference Dimarzio, Rashid and Hancu133} These studies have further shown an association between acute stimulation-dependent activity and long-term clinical outcome, an association robust enough to build predictive models of outcome. ^{Reference Boutet, Madhavan and Elias90,Reference Elias, Germann and Boutet129,Reference Loh, Elias and Germann130} These desirable features position fMRI as a biomarker of stimulation success. As hinted at earlier, these biomarkers are considered increasingly important in the expanding parameter space of modern DBS devices. The trial-and-error testing of individual parameters is not sustainable. Within the constraints of short clinical visits, optimization is particularly challenging when clinical features lack immediate feedback. The current literature suggests it is conceivable that contact and amplitude settings could be efficiently optimized using fMRI. ^{Reference Boutet, Madhavan and Elias90} Beyond this, the fMRI signal appears to be specific to individual diseases and stimulation targets, showing unique patterns with predictive capability in GPi-DBS for dystonia, ^{Reference Loh, Elias and Germann130} subcallosal cingulate DBS for depression, ^{Reference Elias, Germann and Boutet129} ventral striatum DBS for obsessive-compulsive disorder ^{Reference Slepneva, Frank and Norbu134} , the anterior thalamic nucleus for epilepsy ^{Reference Sarica, Yamamoto and Loh135} and the ventrocaudalis nucleus for pain. ^{Reference Rezai, Lozano and Crawley96} These conditions require days to weeks between setting changes for a visible clinical response. ^{Reference Picillo, Lozano, Kou, Munhoz and Fasano5}