Deep brain stimulation for depression informed by intracranial recordings

Network-Minded Implant Surgery

The initial surgery consisted of intracranially implanting 4 DBS leads (bilateral SCC and VC/VS) and 10 sEEG electrodes (Supplementary Information). To allow mapping of relevant brain networks and connections, target locations for both sEEG electrodes and DBS leads were determined based on high-resolution structural connectivity imaging. DBS leads were positioned to span the region of SCC and VC/VS with the joint maximal probability of connectivity with white matter tracts critically associated with each respective target, as we have previously described (18) (Supplementary Information). The DBS leads were externalized (19) to provide access for stimulation delivery. sEEG electrodes targeted downstream depression-relevant frontotemporal network regions: bilateral dorsolateral prefrontal cortex (dlPFC), ventrolateral and ventromedial prefrontal cortex (vlPFC, vmPFC), orbitofrontal cortex (OFC), anterior cingulate cortex (ACC), and mesial temporal lobe (MTL) (20, 21).

For the first time, our study supplemented traditional stereotactic planning tools with a holographic augmented reality platform that enabled interactive visualization of the DBS and sEEG targets and white matter tracts connecting them. This “HoloSEEG” platform created a collaborative planning environment and provided a more complete three-dimensional appreciation of the extended networks that were targeted with this approach (Fig. 1).

Neurophysiological Data-Driven Stimulation

Following the initial surgery, the patient was admitted to the epilepsy monitoring unit, which we conceptually renamed the neurophysiology monitoring unit (NMU). We performed 9 days of continuous high-resolution recordings from the 180 intracranial channels during unconstrained behavior, performance of specific behavioral tasks designed to assess affect and cognition, and delivery of stimulation across a wide parameter space (Supplementary Information). At the conclusion of the NMU period, we performed a second surgery to remove the sEEG electrodes and internalize the DBS leads to implanted pulse generators.

The critical goal of the NMU phase was to gather data to enable generation of personalized stimulation parameter sets to implement and evaluate during chronic outpatient DBS. The traditional method of choosing optimal parameters–trial-and-error exploration of the stimulation parameter space – would be infeasible for several reasons. The first is the sheer size of the parameter space. Conventional two-lead monopolar DBS with 4-contact leads allows 2⁸ possible contact configurations. Factoring in multiple possible stimulation frequencies, pulse widths, and amplitudes, the possible combinations number in the thousands. Dimensionality of the parameter space in this case is exponentially larger: because we use four rather than two DBS leads, and because each segmented DBS lead has eight rather than four contacts, there are 2³² (>4 billion) possible contact configurations alone. Another reason is the mismatch in temporal dynamics between DBS parameter adjustments and observable changes. In DBS for movement disorders, DBS adjustments often have immediately measurable effects (e.g., tremor suppression). In psychiatric DBS, however, the time delay and consistency between adjustments and observable effects are unpredictable. Thus brute force exploration of this high-dimensional parameter space would be extremely time-consuming and may only lead to the identification of somewhat effective parameters (local maximum) but perhaps not the most effective parameters (global maximum).

This trial replaces the effort-intensive, trial-and-error conventional method (Fig. 2A “Forward” solution) with a data-driven, rigorous approach utilizing the critical feature unique to this platform: the intracranial recordings. These recordings provide measures of network activity across a range of mood states occurring through natural variation or induced by specific behavioral tasks. They also provide measures of the network’s response to stimulation across the vast parameter space, which can be efficiently explored with this platform. Combining this information, we can calculate an “Inverse” solution (Fig. 2A): the set of computed stimulation parameters determined most likely to produce the desired behavioral outcome.

In this case, we calculated the inverse solution by defining a desired brain state and identifying the stimulation parameter set that produced a network-wide electrographic pattern that most closely matched it (Supplementary Information). We created correlation matrices between this desired brain state and the multiple stimulation parameter combinations tested per DBS lead, where each matrix correlated spectral power across all intracranial contacts. This process produced a rank-ordered list of parameter sets across the four DBS leads. Then, to determine the relative contributions of stimulation parameter features to the correlation match value, we modeled each component as a beta weight in a generalized linear model and used iterative model fitting to identify the most-contributing components.

Outpatient Phase

We implemented the NMU-derived stimulation parameters in the outpatient phase of the trial. We began with an 8-month period of open label optimization, in which we observed the effects of the chosen stimulation parameters without blinding. The subject then entered the double-blind discontinuation phase. We saw the patient in clinic weekly during this time for close monitoring. At a timepoint unknown to the subject and blinded raters, we began reducing stimulation by 25% per visit on a randomly chosen DBS target pair. A 25% increase in the Montgomery-Asberg Depression Rating Scale (MADRS) over 2 consecutive visits, relative to pre-discontinuation, and Clinical Global Impression – Improvement scale (CGI-I) of 6 (“much worse”) were escape criteria that would trigger exit from the discontinuation phase and reinstatement of open label stimulation. Subject blinding was maintained by asking the subject to not access the patient programmer and to recharge the DBS pulse generator daily so that they would be unaware of differences in recharge duration that would tip off amount of battery usage.

Inverse Solution for DBS Parameter Generation

We delivered stimulation across a wide range of parameter space, including single pulses and pulse trains of 1 sec, 15 sec, 5 minute, and 20 minute duration (Supplementary Information). The 1-second stimulation data were used for inverse solution generation. These short stimulation trains did not produce an observable behavioral response, but 15 second stimulation and longer did. In general, SCC stimulation effects, when present, consisted of feelings of calmness and mental clarity. VC/VS effect, when present, were more energizing, consisting of increased talkativeness and feeling “online”. His symptom scores decreased rapidly during the 10 day NMU stay, decreasing by 56% (MADRS) to 67% (HAM-D) relative to his baseline severity (Figure S1).

We performed the inverse analysis in three ways to identify and confirm a solution. We first chose the subject’s baseline resting activity on day 9 of the NMU stay as the desired state. His symptom ratings were most improved at this point, and his performance on behavioral tasks was at its best (Supplementary Information). Subjectively, he stated that if he were able to consistently feel this way, he would function much better in life, emotionally and cognitively. Whether he had reached this state due to the accumulated stimulation delivery by that time or due to other non-stimulation-specific effects, it represented a desirable state. From the rank-ordered list generated by the inverse analysis targeting this state, we chose the top-ranked set as our prioritized set of stimulation parameters. This modeling approach does not allow us to quantify the significance of the difference between items in the list, however, so we are agnostic regarding the potential effect of choosing any other high-ranking set.

We performed two additional analyses to add confidence to this choice. First, we chose an alternate desired state: activity during one of the behavioral tasks that induced a positive mood state. This choice provided an objectively defined target state to balance the subjectivity of the original choice. The highest ranked match using this alternative was identical in all but one stimulation parameter to the original set. Second, we followed the same steps to match to an undesired brain state, which we chose as the baseline recording on day 2 in the NMU when his depression severity was still elevated (Supplementary Information). This rank-ordered list was in nearly the exact inverse order of the other two, providing further confirmation of the result.

Evaluation of the beta weights in the general linear model provided further information regarding the stimulation parameters that most influenced the match score. Contact configuration – the combination of DBS contacts through which stimulation was delivered – was consistently the most significant contributor. The second most significant contributor was stimulation frequency. Pulse width and amplitude were consistently least significant.

Clinical Outcome

We implemented the NMU-defined stimulation parameter sets during an initial eight-month period of open-label DBS. Of note, this sEEG-informed strategy changed the workflow for clinical programming visits. Because the stimulation sets were generated based on NMU data, little clinic time was spent exploring the parameter space to build stimulation sets for this complex 4-lead configuration.

During this time, the participant reported steady improvements in mood, increased interest in pleasurable activities, and closer emotional connections to loved ones. Professionally, he noticed increased concentration and improved performance at work, and decreased anxiety when making presentations to colleagues and clients. His partner reported increased talkativeness, emotional engagement, and interest in shared activities. Standard rating scales demonstrated progressive improvement in depression severity leading to remission of symptoms by week 22 relative to the initial surgery, or approximately 18 weeks following initiation of chronic DBS (Fig. 2B). He did not report treatment-emergent adverse events such as anxiety, hypomania, or suicidal ideation. Neuropsychological testing at the end of open-label optimization revealed improvement in semantic fluency, abstract visual reasoning, and long and short delay recall of non-contextual verbal material. Whether these cognitive improvements were a direct result of stimulation or secondary to mood improvements is difficult to determine.

At 37 weeks he entered the double-blind, randomized discontinuation phase of the trial to distinguish between true and sham response. He was randomized to SCC discontinuation first. Stimulation amplitude was reduced by 25% per visit beginning on the 3^rd visit in double-blind fashion. The stability of his scores during this lead-in period (lack of nocebo effect) argues against unblinding. During the discontinuation, he reported steadily worsening mood and anxiety. His symptom scores increased during this period until he met rescue criteria. At this point, stimulation was reinstated at pre-discontinuation levels, and his depression symptoms again quickly remitted. He did not undergo discontinuation of the VC/VS target, as our protocol limits the discontinuation phase to a single target for patient safety.