Specific examples of the application of closed-loop DBS in PD are growing (Rosin et al., 2011; Little et al., 2013, 2016; Malekmohammadi et al., 2016). These cases have demonstrated symptom improvement, with substantial power savings and/or reduction in side-effects attributable to stimulation. The feedback substrates and closed-loop control algorithms involved have varied, but most have relied on the amplitude of beta activity as directly recorded in the basal ganglia-cortical loop. The amplitude of such beta activity correlates with bradykinesia and rigidity and is suppressed by both medications that exert effect on central dopaminergic activity, and high frequency DBS (Meidahl et al., 2017). Thus far, closed-loop control algorithms have either engaged an on-off activity pattern, with short ramping onset and offset (see Figure 2), or have employed a more gradual, proportional or hybrid control policy. One important consideration is the optimal reactivity of the closed-loop system, which may impact its efficacy, efficiency and ultimate therapeutic window. In healthy primates, and in patients with PD, beta activity is phasic (Tinkhauser et al., 2017). Longer bursts attain higher amplitudes, indicative of more pervasive oscillatory synchronization within the neural circuit. Shorter bursts predominate in healthy states; and in patients, the relative proportion of short and long bursts negatively and positively correlate with motor impairment, respectively. Therefore, it might be best to selectively terminate longer, (i.e., pathological) beta bursts through closed-loop DBS to both maximize power savings and to spare the ability of underlying neural circuits to engage in more physiological processing which may involve shorter bursts (Tinkhauser et al., 2017).

The role of more complex feedback signals, including multidimensional central and peripheral inputs, remains to be explored, as do the advantages of more sophisticated control algorithms. In particular, it may prove necessary to tailor control loops to afford improved amelioration of patient-specific patterns of impairment (Meidahl et al., 2017). On the other hand, one of the factors constraining the development of closed-loop DBS is the range of possible feedback signals, control policies, and stimulation patterns. Arguably, the field needs to focus on demonstrating an unequivocal gain over conventional DBS before closed-loop DBS techniques are further nuanced. The key will be to shift from acute trials in post-operative patients—where studies are confounded by stun effects and temporal constraints – to acute and then ultimately to chronic trials in patients who have already undergone conventional DBS. The latter will allow resolution of the stun effect, and will afford time for optimization of conventional DBS so that valid comparisons can be made. However, these types of chronic trials will require further development of enabling technology, together with a more informed understanding of the dynamics of target circuits.

In PD, beta activity is phasic with longer busts attaining higher amplitudes, indicative of more pervasive oscillatory synchronization within the neural circuit. Shorter bursts predominate in the healthy state, and the relative proportion of short and long bursts negatively and positively correlate with motor impairment, respectively.

At present DBS is characteristically provided as continuous, open loop, non-responsive input, and uses a “one-size fits all” set of parameters for a wide range of neurological disorders. It cannot respond to the patient's state (asleep/awake, rest/active), dominant symptoms (tremor, bradykinesia, gait impairment) or medication level. In contrast, closed loop (i.e., responsive) DBS (CL– DBS) for PD, using kinematic and/or neural biomarkers, has the potential to deliver more precise and customized neuromodulation, based on state, symptom and level of medication. To be successful CL-DBS requires feedback signals (control variables) that are accurate reflections of the disease state or symptom, and neuromodulation paradigms (control policy algorithms) that are customized for that disease/symptom. Tremor is especially well-suited for responsive or CL-DBS as it varies in amplitude or presence over time and differs between PD subjects. CL-DBS for tremor can be approached using control variables such as a peripheral measure of tremor (Kinematic Cl-DBS) or the STN LFP beta band (13–30 Hz) power (Neural CL-DBS) (Malekmohammadi et al., 2016). Attenuation of STN beta band power by medication or DBS is associated with improvement in both rigidity and bradykinesia and has been the control variable used in CL-DBS studies to date (Kuhn et al., 2008; Whitmer et al., 2012). However, the presence of tremor itself may attenuate beta band power and thus Neural CL-DBS may not be useful for tremor (Shreve et al., 2017).

The effectiveness and efficiency of Kinematic compared to Neural CL-DBS and to open loop DBS were determined in six PD subjects (Malekmohammadi et al., 2016) using a dual threshold control policy algorithm, (Figure 3), for either beta band power, (Figures 3A,B), or tremor power, measured by a Bluetooth enabled smartwatch (Figure 3C). Voltage was increased if either beta or tremor power was above the upper threshold, remained the same if between thresholds, and was decreased if below the lower threshold. Both techniques improved tremor to a similar degree; Kinematic CL-DBS used 10.6% of the total energy that would have been delivered (TEED) during open loop DBS and Neural CL-DBS used 31.5% TEED. Kinematic CL-DBS was significantly more efficient than Neural CL-DBS (P < 0.05). Neural CL-DBS was efficacious in treating tremor if the initial voltage was in a therapeutic range; whereas Kinematic CL-DBS required tremor to be present and initial voltage was set at the lower therapeutic limit. This research demonstrates that CL-DBS, using either a neural or kinematic control variable, is more efficient than open loop DBS. Furthermore, it highlights the potential that additional development of patient- and symptom-specific control variables, and control policy algorithms will improve the efficiency, efficacy and specificity of DBS therapy for PD.

Tremor in PD is a symptom especially well-suited for responsive or closed loop DBS as it varies in amplitude or presence over time and differs between patients.

Levodopa-resistant posture, and FOG symptoms are disabling and difficult to address in patients with PD. Recent DBS trials have rarely addressed FOG and specifically “on-medication” freezing and falling. Freezing appears to involve some GABA-ergically-mediated activity of the globus pallidus interna (GPi) that leads to dysfunction of the pedunculopontine nucleus (PPN). Human DBS studies have targeted the PPN or the PPN plus STN with mixed—and in many cases, unsatisfactory outcomes (Stefani et al., 2007; Moreau et al., 2009; Moro et al., 2010). However, there is no consensus regarding the best PPN target and both rostral and caudal pedunculopontine nucleus subregions have been targeted (Thevathasan et al., 2017). There is great variability in clinical methodology used among surgical centers and the spread of stimulation and inconsistency in targeting suggests that neighboring brain stem regions may be implicated in any DBS response. Because of the intermittent nature of FOG, a feasibility closed-loop neuromodulation (CL-DBS) approach for bilateral PPN DBS plus conventional bilateral open-loop DBS (OL-DBS) of the GPi to manage on-medication FOG in PD has been initiated at the University of Florida. Five patients with advanced PD and refractory FOG were implanted with Medtronic Activa PC+S implantable neurostimulators leads. The patients were carefully evaluated, so as to define true “on” medication freezers. A closed-loop PPN DBS paradigm has been developed using LFPs occurring in GPi and PPN during normal walking and during maneuvers known to trigger freezing episodes. Assessments are blinded and videotaped including objective gait laboratory analysis. PPN CL-DBS was aimed to deliver stimulation at different frequencies (5, 25, 65, and 130 HZ) at the onset of gait. This PPN CL-DBS paradigm was also used for long-term PPN CL-DBS via the Nexus-E firmware, which allows similar Nexus-D operation, but is a completely embedded, enclosed system with no external triggers or machines. GPi stimulation settings were determined by clinical benefit consistent with current standard of care for optimizing pallidal DBS. The study remains ongoing and will report safety and feasibility of this approach along with clear neurophysiological changes in beta and theta bands that could be potential biomarkers to guide future adaptive DBS approaches.

A novel closed loop DBS system that targets GPi and PPN is currently investigated for treatment of refractory “on” FOG in PD patients. Approach appears safe and applicable.

Currently used DBS devices generally deliver “open loop” stimulation, continuously stimulating target structures regardless of changes in the brain circuits related to disease expression. Device programming is a labor-intensive process that is based on “trial and error” and requires significant clinical expertise, which are barriers to widespread clinical application. Additionally, continuous open-loop DBS for PD may result in suboptimal control of fluctuating motor signs, stimulation-induced adverse effects, and short battery life (Little et al., 2016).

As previously noted, one approach is to utilize subthalamic beta band activity as the control signal for closed loop stimulation (Little et al., 2013). Advantages of this approach include conceptual simplicity, and that the control signal can be derived from the same electrode array as used for therapeutic stimulation, thereby obviating the need for additional sensor electrodes. Disadvantages include a relatively low amplitude signal with high stimulation artifact, and that the beta band activity is strongly affected by normal movement, not just the severity of Parkinsonian motor signs (Qasim et al., 2016). We have developed an alternative technique for closed loop stimulation in PD using a sensor that is permanently implanted in the subdural space over the primary motor cortex. This is attached, along with the ipsilateral subthalamic stimulating lead, to an investigational pulse generator (Activa PC+S, Medtronic) with sensing as well as stimulation capability that can be used to prototype feedback control algorithms. The control strategy is based on detection of a narrow band gamma oscillation (60–90 Hz) that has previously been shown to be a biomarker of the dyskinetic state (Swann et al., 2016) (Figure 4).

We have tested this method in three phases (1) in vitro experiments using previously collected cortical signals converted to analog voltages and applied to a saline bath containing a paddle lead attached to an external PC+S; (2) in the clinic, streaming data from an implanted PC+S to an external laptop that controlled the PC+S stimulus via radiotelemetry (Herron and Chizeck, 2014), and (3) in a second clinical application, which evaluated the control algorithm totally embedded in PC+S. In two test subjects, closed loop control (Nexus E) provided 26 and 39% reduction in energy delivered, without adverse effects. Patients in this study did not perceive stimulation- related changes. Cortical contacts were implanted through the same DBS burr hole and appeared to be stable in the long-term (as observed in other recent studies conducted at multiple sites).

A new technique aims to optimize electrophysiological signals used for closed loop DBS using a sensor permanently implanted in the subdural space over primary motor cortex.

Tourette syndrome (TS) is a developmental neuropsychiatric disorder characterized by involuntary and vocal tics, which is commonly co-morbid to other conditions, such as attention deficit hyperactivity disorder and obsessive compulsive disorder (Cheung et al., 2007; Kenney et al., 2008). The etiology of TS remains largely unknown; however, a commonly accepted hypothesis implicates dysfunction of corticostriatal and thalamocortical circuits (Albin and Mink, 2006). DBS is an emerging therapy for cases of severe and intractable TS, and next generation DBS devices, such as the Neuropace RNS and Medtronic Activa PC+S, provide tools with which to record electrophysiological signals that can be used to study network effects and pathophysiology in this disorder.

Two studies of TS electrophysiology using DBS have recently been conducted at the University of Florida. The first implanted dual DBS leads in the centromedian- parafascicular complex of the thalamus, and employed a NeuroPace RNS device both to study electrophysiology of tics and to develop a responsive neurostimulation paradigm. Responsive DBS was achieved in one subject, delivering stimulation in response to increases in low-frequency activity (5–15 Hz) in the thalamus that correlated with dystonic tics (Molina et al., 2017). A second study, using the Medtronic Activa PC+S devices targeted the same thalamic nucleus bilaterally, but also included bilateral cortical subdural strips over the motor cortex to study the thalamocortical network of tic generation (Shute et al., 2016).

Long complex tics were shown to be concurrent with a consistently detectable low frequency activity in the thalamus, as shown in other studies. However, stimulation artifacts resulted in the development of a responsive DBS based on the cortical signatures of tic generation, which involved increases in cortical beta activity (Shute et al., 2016). An ongoing NIH study is examining the physiology of tics and closed loop DBS in 10 human subjects. Resetting tic mechanisms to achieve a more normal pattern of oscillatory activity may reduce tics and also lead to less device discharge over longitudinal follow-up.

Pilot studies of responsive neurostimulation in TS patients have been conducted and a larger trial is currently underway.

LFP analysis is a well-established method for investigating disease, and network and stimulation dynamics in PD. Analysis of intraoperative basal ganglia LFPs can potentially predict the optimal stimulation contact (Ince et al., 2010). There is increasing interest in the use of DBS to treat signs and symptoms of TS, and multiple DBS targets have been proposed for the treatment of TS (Viswanathan et al., 2012). LFP analysis using signals captured from a combination of DBS electrodes and cortical recordings has previously shown thalamocortical activity changes during tics in TS (Shute et al., 2016). Limited data exist about the role of LFPs in the posteroventral GPi in TS, although open label stimulation at this site does seem to improve symptoms (Shahed et al., 2007).

LFP recordings from the GPi during an intraoperative testing paradigm capturing the resting state, voluntary movements, and tic activity have been analyzed and reported in 3 subjects with TS (Jimenez-Shahed et al., 2016). The LFP data filtered between 13 and 30 Hz indicated the presence of event related desynchronization with lower amplitude beta band oscillations during tic periods compared to the resting state. During tic periods, there was also amplitude enhancement in the gamma range (40–150 Hz) and higher frequencies (150–500 Hz) in all subjects. Subjects demonstrated individual changes in the spectral power of LFPs in both the low and high frequency oscillations (HFO) during different states. The resting state was further characterized by coupling between the phase of theta-low alpha oscillations to the amplitude of HFOs in two subjects, while tic activity was associated with beta-HFO phase-amplitude coupling (PAC) in all 3 subjects. These results were in contrast to the novel identification of beta-HFO PAC in the GPi of four un-medicated PD patients at rest. These findings suggest that tic activity can be neurophysiologically distinguished (at the subcortical level) from rest, voluntary movements and akinesia, and that the GPi is a viable target for neuromodulation to decrease tics in TS. Additionally, these data challenge the view that beta-HFO PAC is only a marker of akinesia, since tics represent a hyperkinetic state. These results also highlight the importance of investigating the HFO range of LFP activity.

Given these findings, we posit that continuing investigation of the LFP spectral characteristics and non-linear interactions between different LFP frequency bands across nodes within the basal ganglia-thalamo-cortical network in TS (and other movement disorders) will broaden understanding of the neurophysiologic abnormalities characterizing these conditions. This knowledge will also contribute to the identification of the most appropriate and sensitive signals to trigger closed loop stimulation when different movement patterns are present.

LFP recordings from the GPi demonstrate individual changes in the spectral power in both the low and HFO sub-band during different rest, voluntary movements or tics.

DBS of the subcallosal cingulate white matter (SCCwm) has been shown to elicit durable improvements in depressive symptoms (Riva-Posse et al., 2017). Here, preliminary work is presented that characterizes network-level electrophysiological changes in patients with treatment-resistant depression who were implanted with SCCwm-DBS. Using a prototype bi-directional DBS platform that employed the Activa PC+S (Stanslaski et al., 2012) in conjunction with dense array EEG, preliminary evidence of electrophysiologic changes in the SCC and downstream cortical regions was demonstrated that correlated with both the disease state and stimulation conditions.

In order to assess the validity of any oscillatory biometric captured from patients with chronically implanted prototype devices, preliminary evaluations focused upon limitations of oscillatory analyses in the context of technical device capabilities. In particular, it is known that amplifier limitations are sensitive to disease-independent tissue impedances and impedance changes. Thus, amplifier-related distortions present in bi-directional DBS devices during stimulation were modeled in order to identify potential spurious oscillatory results. Non-linear gain responses from an amplifier that is engaged to its output limits can result in strong high-frequency oscillations and low-to-high phase-amplitude coupling spuriously because of soft-clipping of the input signal. In light of this, we are developing an interactive tool that can be used to assess level of gain compression making analysis and interpretation of oscillatory results more reliable.

Modern analytical approaches were employed to overcome challenges that result from both recording noise and low patient sample sizes. Next, we utilized a machine learning-based algorithm to extract oscillatory biometrics in the chronic PC+S recordings. This approach, when learned on limited, noisy data in a training set, is able to predict the current level of depression severity (as measured by weekly rating scales) with approximately 50% correlation in a testing set of patients. This strategy appears to have technical, analytical, and practical advantages given the limited number of patients, sparseness of sampling, and the known trajectory of clinical depression response(s). This approach, when complemented with standard Fourier-domain approaches, can help to identify useful models for biometrics construction that can be based on correlative analyses.

Despite promising initial results, the exact mechanism of action of SCCwm-DBS still remains unclear, but appears to involve cortical regions beyond the site of stimulation (Mayberg et al., 2005). Using multimodal electrophysiology, we are therefore investigating rapid, transiently induced cortical oscillatory responses to connectomics-optimized SCCwm-DBS. In all patients with simultaneous PC+S LFP and dense EEG, alpha power change patterns can be used to clearly separate stimulation at the optimal SCC contact from stimulation at an adjacent contact on the DBS lead just 1.5 mm away. The complementary accuracy and reliability of this method in confirming current tractographic approaches is the focus of ongoing studies employing a larger cohort.

Various approaches can be used to identify putative biometrics of depression and anatomic sites and networks that be employed in future closed-loop DBS to objectively confirm proper anatomical targeting and potentially optimize therapeutically effective parameters of DBS in treatment of depression.

Involvement of the central thalamus has been implicated in the pathophysiology of certain types of traumatic brain injury (TBI). As this area links to other relevant brain networks and regions, the involvement of the central thalamus may play an important—and reciprocal—role in other brain responses to TBI, as well. Neural networks mediating arousal project to the central thalamus, with efferent projections to the anterior forebrain (Schiff, 2016). Pre-clinical animal (non-human primate—NHP) and rodent) models have elucidated these central thalamic to forebrain connections and their functional activation using electrophysiologic methods and optogenetic techniques combined with fMRI (Liu et al., 2015; Baker et al., 2016). Proof of concept that DBS of the central thalamus can produce improvements in cognitive performance and arousal supports the further development of clinical studies (Giacino et al., 2012). Additional studies in human patients with minimally conscious state showed that central thalamic DBS produced improvement(s) in both motor and cognitive function. Notably, a single-subject in minimally conscious state demonstrated motor and cognitive improvement with central thalamic DBS; this patient regained capacity to identify objects and speak after an initial titration phase along with improvements in attention and organized activity of the upper limb. Oral feeding was greatly improved with DBS as well. A clinical trial targeting patients with persistent cognitive difficulties after TBI is planned. Six patients with severe to moderate brain injury (GOSE Outcome level 6–7) at least two years following their initial injury with persistent cognitive impairment limiting regaining vocation and social reentry will be enrolled. The primary outcomes measure will test attention and working memory function. EEG will be used as a secondary physiological marker of cognitive impairment in order to examine physiological mechanisms of DBS effects and to potentially develop adaptive technologies.

Preclinical and clinical data support the role of the central thalamus and its projections in arousal and attention.

Ventral Striatum (VS) DBS has FDA humanitarian device exemption approval for treatment of intractable OCD, and has been shown to incur 60% clinical benefit in recent meta-analysis (of open-label studies) (Alonso et al., 2015). However, management and programming strategies are increasingly challenged to provide persistent benefit while limiting the occurrence of DBS-induced behavioral side effects, most notably hypomania (Widge et al., 2016). Programming adjustments have been made largely on the basis of acquiring acute beneficial effects on outcome measures of “anxiety reduction,” “improved mood” and “increased energy.” Induction of “mirth” has served as a guidepost measure to programming, and has been regarded as predictive of good response to treatment (Haq et al., 2011). However, DBS-induced mirth also represents a potential risk for development of hypomania or mania. Thus, there is a clear need for an adaptive DBS (aDBS) system that can correctly assess hypomania as distinguished from a euthymic mood state, and automatically adjust stimulation accordingly.

A pilot study funded by the NIH BRAIN Initiative aims to develop and to test a prototype aDBS system for intractable OCD that uses LFP signals to automatically adjust DBS parameters to both improve symptom management and reduce stimulation-induced behavioral side effects. One of the challenges of this study was to identify a label for the classifiers that was objective, reliable, and fitting the National Institute of Mental Health (NIMH) Research Domain Criteria (RDoC) Constructs. Positive Valence (e.g., for elevated or euthymic mood) and Negative Valence Constructs (e.g., to encompass depression, anxiety, and disgust associated with worsening OCD) were selected, with facial affect being used to represent the motor output of emotional state; a measure considered to be superior to clinician ratings of affect for tracking changes in real time. The automated facial affect recognition (AFAR) platform (Tian et al., 2001) will be utilized, and time-locked with inputs from LFPs, EEG, motion, physiology, and changes in DBS programming in order to build classifiers for use in this, and future studies.

Ongoing studies are underway toward the development of adaptive technology that can provide accurate classification of acute fluctuations in obsessive ideation and/or compulsive behaviors that can be used to improve DBS programming for use in treatment of OCD.

Epilepsy is an epidemiologically common neurological disease and over 1/3 of patients have drug resistant epilepsy. In select cases, neurosurgical intervention can cure focal epilepsy, but many patients are not surgical candidates because their seizures are poorly localized or originate from brain regions that cannot be safely resected. Seizures are generally treated as random events, and there are currently no proven surrogate biomarkers for seizures. Clinicians must select a treatment and titration plan (drug, dose, stimulation parameters, etc.) based on clinical guidelines, and then wait for a treatment failure to make further treatment adjustments.

Advances in neural engineering have led to implantable devices capable of therapeutic electrical stimulation (Fisher and Velasco, 2014), seizure detection and forecasting (Cook et al., 2013). Electrical stimulation of the anterior nucleus of the thalamus (Fisher et al., 2010; Salanova et al., 2015) and responsive stimulation of detected electrophysiological abnormalities (Morrell and RNS System in Epilepsy Study Group, 2011; Heck et al., 2014; Bergey et al., 2015) were shown to reduce seizure burden in well-designed pivotal trials, but patients rarely achieved complete seizure freedom. Recently, multiple groups have demonstrated the feasibility of seizure forecasting in both humans (Park et al., 2011; Cook et al., 2013; Brinkmann et al., 2016) and canines (Howbert et al., 2014; Brinkmann et al., 2015) using intracranial EEG. Seizure forecasting has also been demonstrated in ambulatory patients with an implantable device that provides continuous intracranial EEG (iEEG) integrated with a personal assistant device (PAD), running both seizure detection and forecasting algorithms (Cook et al., 2013). Next generation implantable devices are now poised to transform epilepsy management by integrating brain sensing, active brain probing, seizure detection, electronic seizure diaries, seizure forecasting, and intelligent therapeutic stimulation.

The fact that seizure probability fluctuates opens the possibility of directly tracking seizure probability and dynamically adjusting therapy to prevent seizures. Furthermore, it has been shown that the neuronal assemblies activated during seizures may be consolidated in post-seizure slow-wave sleep, i.e., physiological learning mechanisms may strengthen the seizure engram (Bower et al., 2015). This observation suggests targeting neuronal dynamics during post-seizure slow-wave sleep as a brain state dependent therapy. Thus, devices providing continuous brain sensing, probing, and intelligent stimulation based on seizure probability and brain state can open new therapeutic options. In addition, the ability to directly interact with the nervous system could open a new era of brain research.

For example, the RC+S (Bourget et al., 2015) is a rechargeable device, with sensing, stimulation, embedded computational payloads, and continuous iEEG telemetry that affords capability of distributed computing & analytics on a hand-held PAD and cloud environment. It is also rechargeable. Currently pre-clinical development and validation of RC+S in dogs with naturally occurring epilepsy is underway and a human feasibility trial is planned for early 2019 (Figure 5).

There is a need for developing better tools to precisely identify and analyse seizures. Neuronal ensembles during seizures using scalp EEG have shown that neurons that fire together (at the onset of seizure), wire together (during slow wave sleep).

DARPA's innovations in neurotechnology are making possible real-time, seamless translation between human brains and machines. These technologies have enabled initial approaches that enable analyses of the ways that diffuse and varied signals from arrays of firing neurons affect brain function. DARPA's current investments in neural interfaces and related technologies build upon this understanding of neural encoding and decoding, developing multi-scale computational models with high spatial and temporal resolution to target a variety of neurologic functions, and these research approaches are synergized by studies that address engineering challenges of designing implantable, closed-loop systems. Key DARPA-funded capability demonstrations include controlling and receiving feedback from a prosthetic arm using sensors in the motor and somatosensory cortices; receiving feedback from a virtual hand using implants in peripheral nerves; and improving declarative memory through stimulation of specific brain regions. DARPA's work is establishing a basis of what is currently possible, revealing how neurotechnology could facilitate symbiotic human-machine interfacing, and engaging assessment and address of the ethical, legal, social, and policy issues fostered by emerging neurotechnologies (see also, below).

The Restoring Active Memory (RAM) project aims to develop implantable therapies to treat veterans with TBI by increasing their ability to encode information. In this study, the hypothesis that targeted electrical stimulation can modulate neural encoding states and subsequent memory outcomes was tested. Using recordings from neurosurgical epilepsy patients with intracranially implanted electrodes, multivariate classifiers were trained to discriminate spectral activity during the encoding of words that predicted whether patients would later remember or forget these words; and stimulation was applied to various brain regions to modulate performance.

It was found that stimulation modulates performance, with large variability across the population and across the various brain regions stimulated. Hippocampal stimulation tended to impair recall performance while lateral temporal cortex stimulation significantly improved recall performance. In addition, stimulation increased recall performance if delivered when the classifier indicated low encoding efficiency but had the reverse effect if stimulation was delivered when the classifier indicated high encoding efficiency (Figure 12). These data suggest strategies for therapeutically treating memory dysfunction using closed-loop brain stimulation.

A hippocampal memory prosthesis is defined as a closed-loop system that bypasses damaged hippocampal region(s) to restore or enhance memory functions (Berger et al., 2011). Like closed-loop DBS systems, it consists of a recording unit (e.g., multi-electrode array), a signal processing/control unit, and a stimulator (e.g., stimulating electrodes). Differing from DBS systems, which typically deliver stereotypical stimulation patterns (e.g., HFS or LFS) to target regions to modulate neural activities, hippocampal memory prostheses utilize neural code-based stimulation patterns to reinstate neural signal transmission and thus mimic brain functions. Hippocampal memory prostheses have been developed and tested in rodents (Song et al., 2007, 2009; Berger et al., 2011, 2012; Hampson et al., 2012a,b), nonhuman primates (Hampson et al., 2013), and human epilepsy patients (Song et al., 2016; Hampson et al., in preparation).

This technique is now being applied to human studies in which multi-electrode “macro-micro” depth electrodes are implanted in the hippocampus of epilepsy patients undergoing Phase II invasive monitoring for seizure localization (Figure 13). Hippocampal CA3 and CA1 neural ensembles are recorded while patients perform a delayed match-to-sample (DMS) task. Multi-input, multi-output (MIMO) nonlinear dynamical models are built to describe the transformation from CA3 (input) spatio-temporal patterns (codes) of spikes to CA1 (output) spatio-temporal patterns (codes) of spikes using CA3 and CA1 data recorded from success trials of DMS tasks, and further drive stimulations to the CA1 region (Song et al., 2013, 2016). In combination with the modeling from human data, results from preclinical testing in rodents and nonhuman primates (Hampson et al., 2013) demonstrate that (1) MIMO models accurately predict CA1 codes in real-time from ongoing CA3 codes; and (2) closed-loop electrical microstimulation of CA1 using the MIMO-predicted CA1 codes improves DMS performance in those same subjects, indicating improvement of working memory function.