The Neurostimulationist will see you now: prescribing direct electrical stimulation therapies for the human brain in epilepsy and beyond

Abstract

As the pace of research in implantable neurotechnology increases, it is important to take a step back and see if the promise lives up to our intentions. While direct electrical stimulation applied intracranially has been used for the treatment of various neurological disorders, such as Parkinson's, epilepsy, clinical depression, and Obsessive-compulsive disorder, the effectiveness can be highly variable. One perspective is that the inability to consistently treat these neurological disorders in a standardized way is due to multiple, interlaced factors, including stimulation parameters, location, and differences in underlying network connectivity, leading to a trial-and-error stimulation approach in the clinic. An alternate view, based on a growing knowledge from neural data, is that variability in this input (stimulation) and output (brain response) relationship may be more predictable and amenable to standardization, personalization, and, ultimately, therapeutic implementation. In this review, we assert that the future of human brain neurostimulation, via direct electrical stimulation, rests on deploying standardized, constrained models for easier clinical implementation and informed by intracranial data sets, such that diverse, individualized therapeutic parameters can efficiently produce similar, robust, positive outcomes for many patients closer to a prescriptive model. We address the pathway needed to arrive at this future by addressing three questions, namely: (1) why aren't we already at this prescriptive future?; (2) how do we get there?; (3) how far are we from this Neurostimulationist prescriptive future? We first posit that there are limited and predictable ways, constrained by underlying networks, for direct electrical stimulation to induce changes in the brain based on past literature. We then address how identifying underlying individual structural and functional brain connectivity which shape these standard responses enable targeted and personalized neuromodulation, bolstered through large-scale efforts, including machine learning techniques, to map and reverse engineer these input-output relationships to produce a good outcome and better identify underlying mechanisms. This understanding will not only be a major advance in enabling intelligent and informed design of neuromodulatory therapeutic tools for a wide variety of neurological diseases, but a shift in how we can predictably, and therapeutically, prescribe stimulation treatments the human brain.

Keywords: direct electrical stimulation; epilepsy; human; intracranial; neural activity.

Figure 1

AI-rendered neurostimulation future. A Dall-E rendering of what a future clinic visit could look like when prompted, “Draw me a picture of a doctor’s office in 100 years, where a neurologist will be sitting with AR Goggles and a tablet, and a patient will be sitting on the chair. There will be a screen in the background that has a 3D picture of the patient’s brain and their brainwaves.”

Figure 2

Stimulation parameters to induce neural responses. (A) Varying stimulation parameters used to induce behavioral and physiological changes. (B) Stimulation frequencies. The combination of stimulation parameters can therefore represent a massive and daunting space when stimulating intracranially to get a targeted response.

Figure 3

Neural responses in sleep vs. wake. (A) Illustration of stimulation responses via stereo-EEG (sEEG) electrodes implanted in the brain. (B) Changes in a response to single pulse electrical stimulation (SPES) dorsolateral prefrontal cortex site during wake vs. sleep. (C) Measured changes in brain responses to SPES with stimulation at different sites of the brain during sleep and anesthesia compared to when the participants were awake (Zelmann et al., 2023).

Figure 4

Data reduction and predicting responses across participants and frequencies. (A–B) Responses to single pulse stimulation (A) or trains of stimulation (dark grey bar, B) or for different frequencies and currents (N = 3). (C) Current × frequency map of voltage responses across individuals for a voltage response (N = 10). (D) Local stimulation responses across 21 participants with stimulation at the color coded sites shown in the MNI-mapped electrode locations on the left in the dorsolateral prefrontal cortex, illustrating the consistency of responses across individuals and stimulation sites (Paulk et al., 2022). (E) Example dorsal anterior cingulate (dACC) and rostral anterior cingulate (rACC) stimulation responses after a train of stimulation to specific current and frequency levels across multiple individuals. (F) Cingulate and amygdala responses decomposed into principal components, with 94.5% of the variance of local, nearby voltage responses (bottom) explained by principal components 1 and 2. Linear models (LM) incorporating current and frequency with principal component (PC) coefficients can be used to predict responses (bottom; Basu et al., 2019). (G) General framework of the idea of an output attractor state type funneling of brain responses such as the CCEP or high frequency activity changes in spite of a wide variety of input stimulation parameters.

Figure 5

Location and network shapes neural responses. (A) Individual participant where stimulation took place (left column) and white matter tracts mapped across the brain for different targets. (B–C) Similarity scores calculated between 1D connectivity matrices between stimulation sites and recording sites, with similarity between a given network and the stimulation network (Crocker et al., 2021). The 1D connectivity matrices involve concatenating all connection values between the stimulation site and the recording sites (with a subset shown in A), with the connection values either being the strength of the stimulation response from that stimulation site (top), the connectivity measured from the ongoing resting state activity (left bottom 1D matrices), or the DTI connectivity as measured from the DTI weighting from the stimulation and recording ROIs (right bottom 1D matrix).

Figure 6

Effect of neurostimulation on seizure control over time. In this graph of 62 patients who underwent RNS implantation at the Massachusetts General Hospital up until 2021, the cohort demonstrates an overall decrease in the number of long episodes per month compared to the initial baseline during the first 3-months after implantation. Long episodes often represent prolonged ictal activity captured on the ECOG recording of the RNS device. Each color is a different patient.