Characterizing response to patterned electrical stimulation to improve neuromodulation therapies

Abstract

Neuromodulation therapies, such as deep brain stimulation (DBS), aim to modulate dysfunctional neural circuitry in patients with treatment-resistant neurological disorders for symptom alleviation. DBS applications extend from reducing motor impairments in Parkinson’s Disease, essential tremor, and dystonia, to improving mood-related symptoms in major depressive disorders, obsessive compulsive disorder, and bipolar disorder. Adaptive deep brain stimulation (aDBS) has become a critical advancement in neuromodulation, where stimulation therapy is adjusted based on current patient needs, and it has led to improved patient outcomes, reduced side-effects, and increased device battery life. Leveraging the unique opportunity of recording from invasive electroencephalography (iEEG) in humans, this thesis contributes to three distinct components of adaptive DBS: biomarker detection for brain-state estimation, characterizing stimulation response, and identifying candidate stimulation sites for effective neuromodulation. We first motivate the importance of brain-state estimation for adaptive neuromodulation and demonstrate two examples: an unsupervised sleep/wake classifier and a subject-specific pain state classifier. We first build an unsupervised sleep/wake classifier using a Hidden Semi-Markov Model on several-hour long datasets. We showed that our model outperforms other unsupervised models and discuss how unsupervised models could be leveraged in situations where data are plenty, but labels are sparse. We also build a subject-specific pain state classifier and identify unique neural features and models that predict subject-reported pain levels. We identify that models performed better in some subjects and discuss how differences in subject-reported pain may contribute to model performance. Next, we characterize stimulation response by performing patterned electrical stimulation by varying stimulation amplitude and frequency and measuring the responses to stimulation at many recording channels. We identified in what channels stimulation response could be detected and to what extent different stimulation patterns were distinguishable. We found that most subjects had few, select channels that are capable of distinguishing between different stimulation patterns, and that some but not all stimulation patterns are separable. These results contribute to our foundational understanding of neuromodulation, suggesting directions for future work to continue towards building a general model for stimulation response. Finally, we present results establishing that single-pulse stimulation response can be used to predict a channel’s sensitivity to patterned stimulation, providing an efficient method to identify channels responsive to stimulation. We then leverage this idea and demonstrate an experimental paradigm to create stimulation response maps across many stimulation channels. Given a target site for neuromodulation, these stimulation response maps help identify candidate stimulation sites, which could be a helpful tool in identifying patient-specific sites for neuromodulation.