Closing the loop in psychiatric deep brain stimulation: physiology, psychometrics, and plasticity

Abstract

Deep brain stimulation (DBS) is an invasive approach to precise modulation of psychiatrically relevant circuits. Although it has impressive results in open-label psychiatric trials, DBS has also struggled to scale to and pass through multi-center randomized trials. This contrasts with Parkinson disease, where DBS is an established therapy treating thousands of patients annually. The core difference between these clinical applications is the difficulty of proving target engagement, and of leveraging the wide range of possible settings (parameters) that can be programmed in a given patient’s DBS. In Parkinson’s, patients’ symptoms change rapidly and visibly when the stimulator is tuned to the correct parameters. In psychiatry, those same changes take days to weeks, limiting a clinician’s ability to explore parameter space and identify patient-specific optimal settings. I review new approaches to psychiatric target engagement, with an emphasis on major depressive disorder (MDD). Specifically, I argue that better engagement may come by focusing on the root causes of psychiatric illness: dysfunction in specific, measurable cognitive functions and in the connectivity and synchrony of distributed brain circuits. I overview recent progress in both those domains, and how it may relate to other technologies discussed in companion articles in this issue.

Introduction: the target engagement challenge of psychiatric DBS

Widespread clinical use of deep brain stimulation (DBS) in psychiatry is hindered by interlocking challenges of target definition and engagement. Early open label data show that DBS can be life-transforming for individual patients [1–3], but DBS has not met endpoints in multi-center industry-led trials for major depressive disorder (MDD) [4, 5]. Success in those large trials is necessary to achieve full FDA approval and insurance reimbursement. At the same time, there is evidence that those early successes were not pure placebo. In obsessive-compulsive disorder (OCD), centers around the world have achieved consistent response rates of 60–70% despite wide variations in surgical and stimulation technique [6–9]. In MDD, both natural experiments and formal trials suggest that the stimulation itself is therapeutic. First, in early studies, patients and clinicians experienced double-blind therapy withdrawals when device components failed without warning or external signs [5, 10]. Those patients inevitably relapsed, even though they had no knowledge that their relapse was due to loss of stimulation. Second, Bergfeld et al. leveraged that knowledge in a DBS trial design where active electrical stimulation was withdrawn in a blinded, cross-over design after 1 year of treatment [11]. That was the only randomized trial of DBS in MDD to show statistical separation between DBS and a no-stimulation sham. It is possible that this separation reflected patients’ negative expectancy or a withdrawal effect, but that would still leave the evidence from device failures during formal clinical trials. Those relapses occurred across targets and centers, and it is hard to believe that all represent pure withdrawal syndrome given that they in some cases took months to manifest.

Collectively, these studies highlight that DBS can work – but is difficult to use, specifically because it is difficult to ensure therapeutic target engagement. That difficult arises in part from DBS’ tremendous flexibility. A standard DBS system provides 4-8 active electrodes (“contacts”) per hemisphere, each of which can be configured to source current into or sink it out of the brain target. Those different electrode configurations offer a wide range of “current steering” possibilities to optimize or avoid stimulation of specific nearby structures [12–14]. A similarly wide range of current intensities, pulse shapes, and frequencies can be combined with this electric field shaping, leading to millions of possible stimulation settings. It is not possible to thoroughly explore that space, and thus most clinicians limit themselves to a small range of possibilities [15, 16]. In DBS’ main indications, Parkinson disease and essential tremor, the exploration of programming space can be guided by patients’ symptoms. In both movement disorders, tremor almost immediately stops when reasonable settings are found. Target engagement is obvious – if the tremor stops, the stimulating field is reaching the required neural elements. The same method can be used to ensure the DBS electrodes are correctly placed during the initial surgery – if intraoperative test stimulation controls tremor without inducing side effects, the DBS lead is likely in a reasonable position.

In contrast, for psychiatric applications, there is little immediate feedback to guide clinicians. There are well documented acute effects of initiating stimulation at DBS contacts, including marked positive emotional responses [15, 17, 18]. Unfortunately, those acute responses often do not predict whether patients’ core symptoms will change over the subsequent weeks to months [15, 16]. If patients do not respond to the settings suggested by an initial acute response, clinicians must adjust programming based on their intuitions. There is an interesting question of why this is the case – other therapies, particularly ketamine and derivatives, can cause rapid improvement [19]. On the other hand, those treatments also do not last, e.g. patients who respond initially to ketamine require frequent retreatment [20]. What this highlights is that DBS and other antidepressant treatments likely have very different mechanisms of action. Identifying how those mechanisms might synergize is a fascinating question, but well beyond the scope of this review.

Experts with many years of experience in psychiatric DBS develop heuristics [21] or advanced research-grade methods [22, 23] to guide their next steps. Those approaches work, but are hard to scale or operationalize in protocols or manuals. This likely explains why DBS consistently succeeds in single-center psychiatric studies but fails in multi-center trials – at new centers with less expertise, patients may never have received the true study intervention, despite outward appearances [24–26].

More recently, wider recognition of the target engagement problem has birthed 4 core schools of thought or hypotheses on how to improve that engagement:

1. Better definition of the “target” itself, with a specific emphasis on the concept of surgical targeting. In this view, most problems with target engagement arise because the electrode was not placed in the correct brain sub-structure(s), and conversely, optimal identification of those structures would lead to much more reliable outcomes.

2. Customization of the concept of “target” to the individual patient’s phenotype and physiology. In this view, because MDD, OCD, and most other psychiatric disorders are internally heterogeneous [27, 28], the best way to identify the stimulation target for an individual is extensive test stimulation of multiple potential targets within that same individual.

3. Expanding the definition of “target” to include timing and temporal specificity [29]. In this model, the emotional and therapeutic effects of DBS may differ depending on the patient’s brain state just before stimulation. It is a paradigm adapted from epilepsy, where systems that predict and attempt to abort impending seizures are clinically successful [30].

4. Focusing on a “target” defined by a particular physiologic state of the brain (usually in terms of electrical activity) and identifying stimulation parameters that produce that physiologic state.

These are not mutually exclusive models. For instance, an anatomic target defined according to hypothesis 1 might show symptom-related physiology as in hypothesis 4 [31, 32], and stimulation might then be tailored to alter that physiology [33]. A temporally specific stimulation (hypothesis 3) might need to be delivered in a patient-optimized location (hypothesis 2) [34, 35]. All, however, can be considered as examples of “closed loop” treatment. In a closed (feedback) loop system, an intervention is applied and adjusted based on the specific response of the system being treated. This contrasts with an “open loop” treatment, in which a one-size-fits-all intervention is applied, without outcomes measurement. These 4 models differ in the degree of feedback and customization used. Hypothesis 1, focusing mainly on patient-specific neuroimaging, uses a single measurement and in some frameworks would not qualify as closed loop. They all represent, however, an increased focus on patient-specific characteristics compared to traditional psychiatry.

Advances and challenges in psychiatric target engagement

All of the above hypotheses are explored in depth in other articles in this issue, primarily in the context of MDD. Mayberg et al. describe approaches to more precisely defining the anatomic target, building on years of tractographic imaging [22, 23]. Other authors have done similar work to identify putative response tracts for OCD [36, 37]. Krystal et al. describe approaches for identifying which target is best for improving a given patient’s symptoms (hypothesis 2), and describe development of detectors that can stimulate only when a patient requires DBS (hypothesis 3) [35]. Sheth et al. describe a different, but equally data-driven approach to identifying the best stimulation region for an individual patient (hypothesis 2) and choosing the stimulation contact/parameters for that patient (hypothesis 4) [38, 39]. In each case, leveraging unique methods and expertise available at a specialized academic center, the investigators have produced remarkable clinical outcomes that exceed early, open-label DBS trials.

Each of these advances also has practical limitations. The primary approach to hypothesis 1 is tractographic imaging – tracing white matter pathways believed to be important to DBS response (Fig. 1), then ensuring electrode placement within those pathways [22, 37]. For each successful tractographic study, there is a counterexample. The > 80% response rate of [22] is countered by a ~ 40% response rate from an independent group using comparable technique in the same target [40]. The correlations and predictive modeling for OCD presented in [36, 37] did not replicate in an independent sample with different imaging/tract tracing methods [41], and a prospective study of tract-targeted DBS did not find any clinical advantage over standard technique [42]. Thus, hypothesis 1 may be correct but also incomplete, in that anatomic understanding and optimized surgical placement are not per se enough to ensure response. Anatomy must be combined with an understanding of which patients might respond to DBS and how stimulation affects the target structure’s physiology. There is also a challenge of how to make tractography scalable. It is possible that [40] had poor outcomes because the investigators did not replicate a key feature of the advanced modeling used in [22]. It is further possible that this represents a fundamental limit of tractography. In general, there is a growing understanding that MRI as applied to psychiatric populations contains more noise and confounds that we initially appreciated [43, 44]. This is definitely true for diffusion tractography, where small changes in algorithm parameters can cause false positives and negatives that could substantially impact target inference [45–47]. On the other hand, if success in DBS requires advanced expertise, it will be hard to scale DBS to any reasonable population. Mayberg et al. (this volume) discuss approaches that may make advanced imaging more scalable.

Fig. 1. Example tracts/tract combinations correlated with DBS response.

a intersection of three white matter pathways associated with improvement in SCC DBS for MDD, from [23]. b tracts connecting STN to medial PFC and associated with improvement in VCVS and STN DBS for OCD, from [36].

Hypotheses 2-4 have been explored by leveraging a monitoring technique first developed for intractable epilepsy. To localize and treat seizures by removing the seizure-generating focus, neurosurgeons often place temporary monitoring electrodes throughout the cortex and subcortical regions (Fig. 2). Multiple teams of cognitive neuroscientists have leveraged these opportunities to study complex cognitive and emotional phenomena, particularly those that depend on uniquely human cortical regions [48–51]. Two teams extended that work further, by using the same mapping technique to localize and treat depression. In the first approach, focused on hypotheses 2 and 3, studies in an epilepsy population identified possible biomarkers of spontaneous mood fluctuation [52]. Clinicians then implanted a similar montage of monitoring electrodes in a patient with severe MDD, and calculated correlations between the electrical signals at those electrodes and the patient’s daily mood fluctuations [35]. They also tested stimulation at numerous candidate brain sites, mapping how the stimulation effects varied based on the patient’s prior mood [34]. For instance, sites that were mood-improving when the patient felt more depressed could be agitating when she felt closer to euthymic. The team then used the best of these sites for chronic stimulation, using a “detect and treat” paradigm very similar to that used in epilepsy (and using the Neuropace device that is also used to treat focal seizures). Whenever the power of a specific brain oscillation crossed a threshold, the stimulator delivered a brief burst of mood-improving stimulation to the ventral internal capsule, a site previously targeted in both successful and unsuccessful open loop DBS trials [3, 5, 11]. In the first case report, this responsive stimulation led to rapid and durable mood improvements.

Fig. 2. Limbic network mapping for closed loop DBS optimization, adapted from [38].

White dots illustrate the electrodes of linear stereo-EEG (SEEG) probes instrumenting prefrontal, cingulate, and mesial temporal cortices. All of these structures are implicated in the clinical effects of deep brain stimulation, and thus placing these temporary recording electrodes enables identification of the electrophysiologic signatures of clinically effective DBS.

In a second approach, emphasizing hypotheses 2 and 4, a different team applied a very similar monitoring electrode montage, and also monitored how those electrical signals covaried with the patient’s mood [38]. (The two studies found different mood-correlated signals, highlighting the general heterogeneity of such biomarkers and the need for patient-specific calculations.) In contrast to the direct “detect and treat” approach that used bursts of intermittent stimulation, this second study focused on optimizing traditional DBS (constantly delivered throughout the day at a fixed frequency and intensity). Having identified the brain signature of a patient’s better or worse mood, they then applied stimulation to many different brain electrodes, recording the response to bursts of intermittent stimulation. A small number of stimulation sites and intensities produced electrical signatures that resembled the patient’s “better mood” state. The best of these was then chosen for chronic stimulation, and again produced rapid and substantial improvement. Further, this second study included a blinded discontinuation, and demonstrated that symptoms worsened as stimulation was decreased below its therapeutic “dose”.

Both approaches to patient-specific, physiology-informed DBS are still in their earliest stages, having each reported a single successful case. Regardless of the ultimate clinical success, having this type of direct access to the physiology of psychiatric illness will almost certainly offer new insights into the biology of depression and other mental disorders. At the same time, these physiology-based approaches have significant scalability challenges. Each demonstration required tight cooperation between psychiatrists, surgeons, and a team of engineers tasked with building advanced machine learning models to interpret the neural signals. In epilepsy, seizure focus localization can be done visually – the seizure is obvious, and its onset zone can often be inferred simply from the height of the epileptic signals on different electrodes. These patient-specific neural markers were not so readily visible, and required advanced quantitative approaches with careful data inspection. It is not clear how to embed such expertise in every potential implanting center. An implant manufacturer could perhaps offer “data analysis as a service”, just as they currently support surgeons with a representative in every operating room. To the degree that this analysis directly drives stimulator programming, however, it could represent clinical decision-making by the manufacturer, raising substantial legal and regulatory risks.

More broadly, all the approaches just discussed depend on and reify two problematic constructs. First, they all emphasize the notions of “depression” and “mood”. Depression is not a single clinical entity. When considering internal heterogeneity within criterial (e.g., irritable vs. depressed mood, increased vs. decreased appetite), there are over 2000 ways to arrive at a diagnosis of MDD by finding 5 of 9 DSM-5. criteria. Atop this, comorbidity is the rule rather than the exception in treatment-resistant psychiatric disorders, across all diagnoses. This was even true in the DSM-5 field trials, where over 50% of patients diagnosed with MDD also met criteria for at least one other diagnosis [53]. It is likely not possible to reduce this type of heterogeneous construct to a single 0–10 scalar mood rating or a composite score on a larger scale. As also reviewed elsewhere [54, 55], this is likely why searches for electroencephalographic or genetic markers of antidepressant response have failed – they are trying to find a common biological basis for a collection of different disorders, with an average signal of zero. In theory, this objection can be overcome by patient-specific modeling, as exemplified by [35]. As long as a patient is internally consistent in their mood self-ratings, then it should be possible to identify neural correlates of the subjective experience of distress for that patient. Identifying those patient-specific markers circles back to the question of the extended data analyses required, as well as a need for proof that mood ratings are in fact internally consistent.

The second problematic construct is the general concept of “DBS”, which to date has meant one very specific form of stimulation: high frequency, almost always 130 Hz, and almost always delivered as tonic, 24/7 stimulation (with [35] as a notable exception). That paradigm was chosen largely because it works in DBS’ primary indications of Parkinson disease and essential tremor, i.e., its safety is well established. The trial results above show that this paradigm can work – in some patients. They also show that tonic high-frequency DBS fails to truly treat or reverse the root cause of psychiatric illness. As described above, when stimulation is withdrawn, patients almost always relapse (although there are exceptions [56]). This is true in both MDD and OCD [57], and has also been demonstrated in acute behavioral studies [58] and animal models [59]. This stands in contrast to transcranial magnetic stimulation (TMS), where a relatively small amount of stimulation can produce clinical benefits easily lasting a year or more [60, 61].

The remainder of this article discusses two alternative psychiatric DBS strategies that may overcome those problems. Diagnostic heterogeneity may be addressed by refocusing away from diagnoses toward more objectively measurable constructs. Those constructs may map more reliably to the types of circuits that are readily accessible to DBS. The limitations of tonic high-frequency stimulation might be addressed by developing new stimulation approaches that leverage known neuroplasticity principles. As described below, new paradigms are already emerging in neurostimulation for motor disorders, and may adapt well to psychiatric syndromes.

Measuring target engagement through functional domains

Rather than developing treatments for categorical labels such as “depression”, we could consider psychiatric disorders as composed of measurable, objectively definable ingredients (commonly termed “functional domains”). For instance, the computational psychiatry movement argues that disorders can be considered as failures in various evidence evaluation and decision making systems [62–66]. A computational psychiatry lens might attempt to characterize, for instance, whether a given patient’s experience of depression related more to difficulty in anticipating (and thus seeking) natural rewards such as social interaction, in applying executive function to sequence the steps and mental models needed to achieve those rewards, or in learning from experience that reward is possible. The National Institutes of Health Research Domain Criteria (RDoC) framework makes a similar argument, but with a perspective focused more on theoretical psychological constructs such as response to actual/potential threat or abstract cognitive systems such as working memory or response inhibition [27, 65, 67]. Critically, both frameworks emphasize a “precision medicine” approach, where once a patient’s specific deficit is identified, a brain system subserving that function is treated to remediate the deficit [24, 68–70]. That approach perfectly matches DBS’ capabilities. Almost all decision and emotional functions involve interactions between cortical and subcortical structures. Psychiatric DBS can target those cortico-subcortical interactions, often by stimulating white matter pathways that mediate them [22, 24, 71]. Modern DBS hardware offers the possibility of “current steering”, shaping the emitted electric field to precisely capture and modulate sub-circuits within those larger pathways [13, 72, 73]. In other words, if we can describe patients’ symptoms in terms of decisional or functional domains that are known to map to specific brain circuits, DBS can deliver treatment specifically to those circuits.

The concept of functional domain mapping stands alongside another line of work that argues that although categorical DSM-based diagnoses such as MDD are internally heterogeneous, they are likely composed of neurobiological subtypes or “biotypes”. Instead of working in a space of named psychological constructs (which have been arrived at by the same expert-based, fallible reasoning that led to the DSM constructs), one could consider a more data-driven approach of simply collecting physiological, imaging, or behavioral data from large numbers of patients. With a large enough sample, common clusters might emerge, and these could be considered as diagnostic entities. That approach is easy to describe, but challenging to implement in practice. First, the sample sizes required are quite large – emerging literature suggests that tens of thousands of well-measured participants are necessary for a purely data-driven, discovery approach [43]. Second, almost all such studies emphasize advanced machine learning methods, which can be extremely sensitive to specific analytic choices and which can easily deceive even a cautious investigator [74]. For instance, a reasonably sized attempt to define depression biotypes appeared to find a subtype with a particularly TMS-responsive circuit impairment [75]– but a follow on analysis by another group found that this result was likely by chance solely due to the pipeline used [76]. Many of the measures that might be used in biotyping are not test-retest reliable, i.e. they do not consistently produce the same answer twice when used on the same participant [77, 78]. Reliability problems can be overcome – but this again requires careful technique [79]. It also may well lead back to the concept of functional domains. For instance, a pipeline specifically engineered for reliability and individual differences measurement was most useful when interpreted in terms of known circuits linked by theory to specific psychological domains [80].

This domain-oriented approach could advance the art and science of DBS on several fronts. First, it could make outcomes more measurable. A major impetus behind the computational and RDoC frameworks is the relatively subjective nature of DSM-5 diagnostic criteria and their associated rating scales [27, 62, 64, 67]. Domain-oriented assessments are often obtained through direct, objective measurements of behavior, e.g. specific response patterns on psychophysical tasks [64, 81–84] or movement patterns captured through phone sensors [85–87]. In other DBS indications, the development of reliable, objective rating scales was central to reducing noise in multi-site clinical trials and thus proving clinical value [88, 89]. Objective behavioral readouts might also address the DBS parameter programming challenge discussed above. For instance, if behavior on a psychophysical task changes quickly in response to changes in stimulation settings, then that behavior could be used to measure DBS’ effects on the target circuit.

Second, functional domains are a path to animal models of psychiatric DBS. Major advances in other DBS indications, primarily Parkinson disease, have been aided by the ready available of rodent and non-human primate models [90]. Those models allow new stimulation paradigms to be prototyped, understood, and engineered for reliable outcomes [91, 92], allowing human trials to begin with a firmer scientific footing. Unfortunately, there are no reliable animal models of psychiatric disease, because those diseases are defined using constructs based in subjective human experiences that animals cannot report [62, 93]. Decisional and RDoC constructs, on the other hand, have established cross-species homologies [62]. It may be possible to identify DBS or DBS-like stimulation paradigms that affect functional domains in model animals, which would then act as “lead compounds” to seed more mechanism-directed human trials. Translation will be hindered by the reality that rodents have very different cortical and subcortical anatomy, lacking major prefrontal divisions believed to be important for DBS’ psychiatric effects [94, 95]. There is reasonable homology, however, in circuits subserving avoidance, safety learning, addiction, and compulsive behavior [94, 96–98], all of which relate to major diagnostic categories for which DBS has been studied [6, 7, 24, 99, 100].

Third, and closely related, most functional domains are trans-diagnostic: they are impaired in multiple diagnostic categories. Excessive avoidance responses to potential threats, for instance, have been suggested to underpin generalized anxiety, post-traumatic stress disorder, and possibly obsessive-compulsive disorder (OCD) [8, 98, 101–103]. Cognitive control, the ability to shift and inhibit responses in the face of changing rules, is impaired in patients with depression, OCD, eating disorders, addictions, and autism – among many others [104–110]. Thus, if DBS can be developed to remediate one of these domains, it could simultaneously unlock treatment of multiple indications (as opposed to the current process of separate clinical trials and development in each diagnosis). There is evidence that trans-diagnostic treatments can work, the most well-known being the clinical success of the Unified Protocol for Trans-Diagnostic Treatment of Emotional Disorders. This cognitive-behavioral therapy (CBT) protocol targets emotion regulation dysfunctions believed to be central to multiple mood and anxiety disorders, and has evidence for successful treatment of depressive, anxious, and compulsive symptoms [111, 112]. The same argument could be made for other CBT strategies such as exposure (effective in the treatment of multiple disorders involving overlearned avoidance [8, 113–115].

As a specific example of that domain-focused approach applied to DBS, we and others have shown how closed-loop DBS can be applied to enhance the domain of cognitive control. Cognitive control is, roughly, the ability to inhibit or replace a default behavior strategy with a more adaptive alternative. It can be impaired in a range of disorders that have also been proposed as DBS indications, including addiction, MDD, OCD, and PTSD [104–110]. Improvements in cognitive control may also be central to recovery from those disorders. For instance, patients who respond to DBS for OCD describe a subjective improvement in their ability to focus their thoughts and behaviors away from their OCD-driven patterns [116], and objectively gain more benefit from exposure/response prevention CBT [117, 118]. The predominant target for OCD DBS, the ventral internal capsule/ventral striatum (VCVS), accesses fronto-thalamic and fronto-striatal circuits that are central to cognitive control, and to response inhibition more generally [95, 119, 120]. (VCVS also appears in the literature under other names; see [71] for a detailed review.) That same target can be efficacious in MDD [3, 11, 24], again highlighting the potential trans-diagnostic value of cognitive control improvements.

To show that VCVS DBS enhances cognitive control, we first worked with patients who had DBS systems implanted at this target as part of prior clinical trials [3, 5, 121]. In that first study, patients performed a laboratory task that “stress tests” cognitive control, the Multi Source Interference Task (MSIT), with their stimulators either off or on at usual, clinician-determined settings. With DBS on, patients’ MSIT performance improved – they became 5–10% faster without making more errors (i.e., they were able to shift the canonical speed-accuracy tradeoff curve) [58]. This reaction time effect was specific to cognitive control: it did not occur during simple motor tasks or during value-based choice tasks. Further, when DBS was on, patients showed increases in a well-established electrophysiologic signature of cognitive control (frontal theta oscillations [108, 122–124]), particularly in lateral prefrontal structures that are strongly associated with response inhibition [47, 125].

We then replicated the effect in a second population, patients undergoing acute neurosurgical monitoring for treatment-refractory epilepsy [49]. A small number of such patients had electrodes in or near the VCVS target. Brief bursts of DBS-like electrical stimulation, delivered through those internal capsule electrodes, again decreased MSIT response times without increasing errors, and again produced increases in frontal theta power. Another group recently reported independent replication of the same effect, using a different task [126]. More importantly, in this second study we showed that an objective improvement in cognitive control (as measured through MSIT performance) could be used for closed-loop DBS. We developed a filtering algorithm that inferred a smoothed version of participants’ MSIT reaction time, regressing out the effects of trial-to-trial variability and differing trial difficulties [127]. We then used that algorithm to track MSIT performance trial-to-trial, and stimulated only when participants performed more slowly (i.e., when cognitive control functions were lapsing). This effectively doubled the performance improvement [49].

Finally, we recently showed that VCVS DBS’ effects on cognitive control can be used as a “read-out” to improve the reliability of DBS programming, i.e. to solve the central target engagement problem described above. In our study with epilepsy patients [49], we observed that the cognitive control enhancement had topographic specificity – it was stronger when stimulation occurred through electrodes placed more dorsally in the capsule, closer to white matter fibers originating in lateral PFC [47, 94, 128]. In other words, the enhancement is a read-out of whether particular sub-circuitry is being engaged by stimulation. We then inverted this problem, and asked whether, given only a patient’s behavior, we could reliably identify the best DBS contact to engage cognitive control circuits. In a simulation considering multiple Bayesian optimization algorithms, we showed that an automatic algorithm, with no clinician input or prior information, could reliably identify the best contact for a patient over 80% of the time [129]. With active clinician guidance (e.g., knowledge of whether a particular setting caused subjective improvement in addition to the objective behavior change), that accuracy would likely increase. Critically, this approach is also simple and scalable. All that is needed is a laptop-scale computer and two monitors, one to display a task to a patient and one to display information to the programming clinician. In comparison, every approach to closed loop DBS described above requires advanced imaging, extensive (and expensive) inpatient electrophysiologic monitoring, and/or expert analytic support. Those are unlikely to generalize outside specialized academic centers; the cognitive control approach can. There is an interesting side question of why it is possible to produce acute cognitive change and yet subjective improvement from DBS takes weeks; this will be an important mechanistic question for future studies.

Centering DBS and DBS programming around a functional domain also has unique challenges. We have a worked example for cognitive control, but that depended on having the DBS target already developed through research grounded in classic categorical diagnoses. To extend to other key functional domains, such as emotion regulation [130, 131] or reward motivation [83, 132], new targets and reliable behavioral tasks may need to be developed. It is also not clear what we might do if a patient has symptoms in different domains. Some domains might be addressable through a single implant. For instance, the VCVS target that we have developed for cognitive control can also access the ventral striatum, which is heavily implicated in reward and motivation [119, 133]. In other cases, one might need to place leads into very anatomically separated structures or circuits. Because almost all such circuits will at some point converge through associative PFC, there will almost certainly be interactions when attempting to treat two domains at once. The pilot multi-target work by Sheth et al. (this issue) may help to clarify the feasibility of a multi-target approach.

More generally, RDoC and related initiatives are still relatively young, and it is not yet clear that their specific categorizations correspond to natural kinds or reliable behavior-to-circuit mappings. There is also a major gap in patient assessment. Although laboratory-style psychophysical tasks can be used for the type of within-patient optimization just described, they are very poor at capturing individual differences, i.e. actually identifying the patients who have strong cognitive control dysfunctions [134, 135]. There are attempts underway to develop more reliable trans-diagnostic assessments, including for cognitive control [80, 84, 136], but they are not yet validated. Without those assessments, it will be hard to conduct formal clinical trials of domain-focused DBS or any other domain-focused neurostimulation. Further, although trans-diagnostic domains are reasonably accepted as a research construct, they do not fit into standard medical device regulatory frameworks. Agencies such as the US Food & Drug Administration or the European Medicines Agency are set up to evaluate “labeling” of a specific device for a specific code or set of codes in the International Classification of Diseases (ICD). ICD does not include domains or constructs, and as such, it may be impossible to gain regulatory approval for treating them. The best near-term approach may be to “embed” domains within ICD diagnoses, e.g. to seek approval for treating MDD, but in patients whose MDD is specifically driven by cognitive control impairments. In all three cognitive control studies discussed above, some participants whose cognitive control improved also reported subjective well-being improvements, particularly in their ability to direct their attention away from ruminative inner narratives [49, 58].

Refining neurostimulation for neuroplasticity

Tonic high-frequency stimulation works for Parkinson’s disease and tremor but is almost certainly not the optimal approach to psychiatric disorders, any more than anticonvulsant medications could be considered optimal or even specific treatments for mania or agitation. Even in Parkinsonism, we know there are better options. A “coordinated reset” paradigm can disrupt pathological oscillatory circuits and provide symptom relief that lasts for hours even after stimulation ends, in simulations, non-human primate models, and pilot human studies [91, 137, 138]. An elegant mouse study showed that pulse train parameters can be tuned to preferentially excite specific interneuron sub-groups, and that doing so produces bradykinesia improvements that again outlast stimulation [92]. The underlying principle in both cases is timing-dependent neuroplasticity – that if a target circuit’s physiology and specific resonances are understood, then rationally designed pulse trains can be administered to evoke neuroplastic processes for stable, long-term symptom relief.

Two plasticity-oriented paths are particularly promising for psychiatric DBS: synaptic connectivity modification and oscillatory synchrony control. For the first approach, synaptic modification, many mental disorders are believed to involve dysconnectivity in specific monosynaptic circuits. Anxiety and trauma/stress related disorders are often framed as failures of top-down prefrontal control over amygdala and subcortical networks [98, 101, 139]. The dominant model of OCD and addiction involves hyperconnection of cortico-striato-thalamic loops, particularly loops associated with habit [8, 109, 140]. Neural engineers focused on rehabilitation of stroke, spinal cord injury, and paralysis have repeatedly shown that specific synaptic connections can be strengthened or weakened through timing-dependent plasticity. These studies exploit the classic Hebbian principle that a pre- and post-synaptic neuron can become more strongly connected if they are reliably active in a pre-before-post timing pattern, or can instead become less connected if a post-before-pre pattern repeatedly occurs [141]. There are at least three paths to producing those desired timings. First, the activity of a pre-synaptic region (cell ensemble) can be monitored, and stimulation can be delivered to the post-synaptic region whenever the pre-synaptic partner becomes active [142] (Fig. 3a). Depending on the specific lag between detection and stimulation, this paradigm can strengthen or weaken a target connection [143–147]. The same detect-and-stimulate paradigm can be applied in psychiatrically relevant circuits [148]. Second, paired pulses or pulse trains with the desired inter-regional timing can be directly given to the two sides of the synapse (Fig. 3b). This “paired pulse” approach can specifically facilitate motor connections [149–151], and we recently showed that it can strengthen or weaken fronto-amygdalar connections linked to fear and safety behaviors [152]. Third, stimulation can be paired with activity that naturally activates the target synapse, or at least one of the structures on either side of it. For instance, stimulation paired with natural muscle movements or patient attempts to create those movements has caused functional gains in spinal cord injury [153–155]. Similar approaches have been proposed for, e.g., augmenting cognitive-behavior therapy with neurostimulation [156], and could in theory be adapted for psychiatric DBS. There are proof-of-concept demonstrations of stimulation that can enhance learning, if given at the moment when a relevant memory is being encoded [51, 157, 158].

Fig. 3. Neurostimulation approaches to evoke targeted neuroplasticity.

a Activity-dependent or power-dependent stimulation. When activity in a pre-synaptic area reaches a certain level, stimulation is delivered to the post-synaptic target, possibly with a delay “T”. Following principles of timing-dependent plasticity, values of T in the 10–100 ms range will tend to strengthen synapses between the two regions, whereas much longer value will tend to weaken those same synapses. b Paired-pulse stimulation. This paradigm applies the same timing-dependent principle as (a), but rather than waiting for the pre-synaptic activity to naturally achieve a target level, paired-pulse stimulation enforces that pre-synaptic activity through a depolarizing pulse. The relative timing T between pre- and post-synaptic stimulation again determines whether stimulation strengthens or weakens the connection between areas. In a paired-train variant described in [152], each region receives trains of pulses, rather than a single pulse. If those trains match the target region’s endogenous oscillatory frequency, they can have greater effects on cross-regional connectivity. c Oscillation-locked stimulation. Individual pulses or trains are delivered at specific phases (points in the rise/fall cycle) of an endogenous oscillation. Depending on the target phase, this paradigm can suppress or enhance an oscillation within a single region or the oscillatory synchrony between regions.

The second high-level approach, oscillation control, leverages a growing body of work suggesting that psychiatric disorders (and brain disorders more generally) involve disruptions in large-scale electrical oscillatory processes [159–162]. Almost every brain region shows electrical oscillations in its local field potential (LFP). Theorists argue that these oscillations represent rhythmic changes in excitability, such that an input arriving at the correct phase (rise/fall cycle) of an LFP oscillation is more likely to be encoded, processed, and used to affect behavior [163, 164]. If oscillations synchronize and coordinate across brain regions (oscillatory coherence), then those regions should become more able to efficiently communicate. In support of that theory, both our group and others have demonstrated decoding of cognitive and emotional behaviors specifically from cross-regional oscillatory synchrony [165, 166]. Invasive LFP recordings from patients with both neurologic and psychiatric diseases demonstrate oscillatory hyper-synchrony in circuits believed to be over-active and contributing to symptoms [33, 167, 168].

Given those findings, if within-region or cross-region oscillatory synchrony could be controlled, and if that control could shift given circuits/networks into new stable oscillatory states, then this oscillation-informed stimulation might be a powerful approach to inducing plasticity in networks of psychiatric illness [29, 169, 170]. A particular promising approach is “phase locked” or “phase aware” stimulation, wherein single pulses or brief bursts of stimulation are given at specific phases of an ongoing oscillation (Fig. 3c). Because oscillatory phases reflect moment-to-moment changes in excitability, stimulating at preferred phases should cause greater depolarization and larger physiologic/behavioral effects. For instance, in both TMS [171, 172] and DBS for Parkinson disease [173, 174], bursts of stimulation have differential effects depending on an underlying oscillation’s phase at the moment they are delivered. Repeatedly delivering phase-locked stimulation can increase [175, 176] or decrease [174, 177–179] the power of a target oscillation. Thus, to control local oscillations within a region, stimulation might be given directly to that region, either at high-excitability phases (enhance oscillation power) or low-excitability phases (suppress oscillation power).

Phase-locked stimulation could also be used to control cross-regional communication, by controlling LFP oscillatory synchrony between target regions. To achieve that synchrony control, stimulation could be given to a “downstream” region at times dictated by the phase of an “upstream” region. In motor circuits, that approach increased synaptic connectivity between the stimulated sites [180]. In a rodent model of learned helplessness, increasing cortico-thalamic coupling via cross-regional phase-locked stimulation altered animals’ behavior in a putatively antidepressant fashion [181]. Critically, different sub-circuits oscillate and synchronize at different frequencies, creating separable “communication channels”. For instance, in visual circuitry, “bottom up” processing tends to occur in theta and gamma bands, whereas “top down” feedback and attentional control tend to occur in circuits synchronized at beta frequencies [182]. Thus, control of cross-regional LFP synchrony at specific frequencies could enable DBS that affects specific sub-circuits, potentially increasing therapeutic effect and reducing off-target effects [29, 159, 169]. That control may not require constant, continuous DBS-like stimulation, but likely would require intermittent delivery to maintain the desired connectivity pattern.

Plasticity-oriented DBS may have unique advantages as psychiatric DBS’ clinical focus expands beyond depression to other difficult-to-treat illnesses. Some disorders specifically involve pathological plasticity, and thus a plasticity-oriented approach might be needed to treat them. Addiction is often framed as a plasticity problem, where reward- and habit-related circuits become sensitized to drugs of abuse and retain that sensitization even during prolonged abstinence [140, 183]. Cravings that drive relapse are argued to arise because those circuits are never de-sensitized. If DBS-like stimulation could induce plasticity that opposes drug-driven sensitization, cravings might persistently disappear. Rodent models support that approach [184], although their translation to humans will depend on developing the circuit specificity described above. Similarly, the transition from acute to chronic pain is associated with strengthening of connections in emotion and habit circuits, which are theorized to underlie the catastrophization and avoidance commonly seen in chronic pain [185, 186]. A plasticity-oriented approach could in theory reverse those processes. Because the above approaches all require only intermittent stimulation, they also could avoid a known side effect of DBS for pain (in some targets) – kindling of seizures over time [187, 188]. In general, it may also be possible to reduce side effects from DBS using a plasticity-oriented approach. While there is not firm evidence for this contention, there is a wide belief that common side effects (worsening symptoms, hypomania, agitation) are due to engagement of the "wrong" circuits that happen in some patients to be near to the therapeutic target. Because plasticity-directed stimulation (A) more specifically engages the target circuit and (B) delivers less total energy, it may be possible to limit those off target effects.

While plasticity-oriented DBS-like approaches are promising, they have two major limitations: the targets and parameters are understudied. To the first point, although there are clear examples of plasticity in psychiatric pathogenesis and disease maintenance (see above), the precise circuits/synapses that should be targeted for treatment are not fully known. Given the internal heterogeneity of most psychiatric disorders, those circuit targets likely vary between patient subgroups. Novel approaches to biotyping and clustering patients may resolve that heterogeneity [80, 84, 189], but are not fully developed or validated. In the near term, patient-specific circuit mapping as described in [35, 38] may be necessary to enable plasticity- or oscillation-focused DBS. Second, the precise activity-to-stimulation timings (and other parameters, e.g. stimulation intensity) that will maximize plasticity or oscillatory entrainment are not known. They likely differ from circuit to circuit, because of the underlying physiologic differences that cause those circuits to have different preferred frequencies. It is possible, as shown in [143, 152] to exhaustively test a range of possible inter-region/inter-stimulus timings. The challenge is that this is extraordinarily time consuming. Because neuroplasticity processes are mainly observable at timescales of minutes to hours, we run into the same programming problem that has hindered clinical DBS. If there are hundreds to thousands of parameter combinations, and each may take hours to observe, it becomes impossible to fully search parameter space and find the optimum.

That parameter search might be aided by algorithms that can rapidly analyze subtle physiologic changes. Long-term neuroplastic and oscillatory entrainment effects likely arise because each individual stimulation event triggers a specific physiologic event. For instance, in [152], the paired-pulse parameters that produced long-term plasticity also caused roughly 1 second of oscillatory coherence between the targeted structures, just after stimulation was given. That is, an acute response to specific stimulation parameters predicted the chronic effects of multiple deliveries of those same parameters. If the same principle holds true across circuits, then rapid detection and analysis of those acute physiologic responses could more quickly identify plasticity-inducing stimulation parameters. Such responses are likely too subtle to be directly detected by a human observer, but could be revealed by real-time analysis. In that vein, multiple investigators have demonstrated real-time algorithms, in many cases coupled with machine learning systems, that map the stimulation “response surface” far efficiently than a human could. Examples include a specific set of mapping sequences for depression-related circuits [190], tremor response mapping for DBS in Parkinson’s [191], and our own work described above in optimization for cognitive control [129]. Similar concepts have been leveraged to develop oscillation-locked stimulation in a Parkinson’s model and suppress pathological beta oscillations [178].

Conclusions and future directions

Deep brain stimulation is a powerful technology for turning circuit insights into treatments, and for helping patients whose symptoms have not responded to non-invasive therapies. Applying and scaling DBS in psychiatry, however, requires better understanding of which patients should receive DBS at which anatomic targets, and how stimulation parameters should be adjusted to engage those targets and their connected circuits. The past few years have seen a proliferation of potential techniques to verify target engagement, based on advanced neuroimaging and/or extensive invasive physiologic recording. Those techniques are both promising and potentially limited, in that they may not directly treat the underlying cause of a patient’s symptoms. Further, the underlying limitations of our diagnostic classification system make it difficult to ensure that we are selecting patients likely to benefit from DBS. I have presented two alternate approaches to overcome these limitations, one grounded in functional cognitive measurement as a means to target engagement, the other emphasizing neuroplasticity as a means to greater long-term efficacy. Each may solve problems that current closed loop approaches cannot address, but because these are relatively new approaches, each has technical challenges yet to be overcome. Nevertheless, the combination of these new stimulation paradigms with innovations in surgical target mapping offers the promise of a much more precise and patient-customized approach to DBS in psychiatry.

Author contributions

ASW wrote the article.

Funding

Preparation of this article was supported by R01MH125429, R01MH124687, R01MH123634, UH3NS100548, R01NS120851, R01MH119384, R21DA052568, and the MnDRIVE Brain Conditions Initiative. The views are entirely those of the author(s), not any funding body.

Competing interests

ASW receives DBS device donations from Medtronic and consulting fees related to DBS for psychiatric illness from Abbott. He holds multiple granted and pending patents in the area of closed-loop deep brain stimulation.