In silico development and validation of Bayesian methods for optimizing deep brain stimulation to enhance cognitive control

Abstract

Objective: Deep brain stimulation (DBS) of the ventral internal capsule/striatum (VCVS) is a potentially effective treatment for several mental health disorders when conventional therapeutics fail. Its effectiveness, however, depends on correct programming to engage VCVS sub-circuits. VCVS programming is currently an iterative, time-consuming process, with weeks between setting changes and reliance on noisy, subjective self-reports. An objective measure of circuit engagement might allow individual settings to be tested in seconds to minutes, reducing the time to response and increasing patient and clinician confidence in the chosen settings. Here, we present an approach to measuring and optimizing that circuit engagement. Approach: We leverage prior results showing that effective VCVS DBS engages cognitive control circuitry and improves performance on the Multi-Source Interference Task (MSIT), that this engagement depends primarily on which contact(s) are activated, and that circuit engagement can be tracked through a state space modeling framework. We develop a simulation framework based on those empirical results, then combine this framework with an adaptive optimizer to simulate a principled exploration of electrode contacts and identify the contacts that maximally improve cognitive control. We explore multiple optimization options (algorithms, number of inputs, speed of stimulation parameter changes) and compare them on problems of varying difficulty. Main results: We show that an Upper Confidence Bound (UCB1) algorithm outperforms other optimizers, with roughly 80% probability of convergence to a global optimum when used in a majority-vote ensemble. Significance: We show that the optimization can converge even with lag between stimulation and effect, and that a complete optimization can be done in a clinically feasible timespan (a few hours). Further, the approach requires no specialized recording or imaging hardware, and thus could be a scalable path to expand the use of DBS in psychiatric and other non-motor applications.

SAFE-OPT: A Bayesian optimization algorithm for learning optimal deep brain stimulation parameters with safety constraints

“…Data-driven, particularly Bayesian, optimization has recently been proposed as an algorithmic strategy to efficiently search for the optimal stimulation setting for each individual patient. [10][11][12][13][14] These approaches have been shown to be robust to noise, are more efficient than more naïve grid or random search approaches, and are capable of navigating high-dimensional parameter spaces. [15][16][17] Data-driven optimization has the potential to automate the process of selecting stimulation settings, paving the way for continually learning the optimal stimulation setting for an individual patient without ongoing supervision from clinicians.…”

Sequential estimation of nonstationary oscillator states and stimulus responses

“…Bayesian inference has had great success in biology and biomedical engineering, from simulationbased inference of neuroscience models using experimental data (Gonçalves et al 2020) to parameter optimization of wearable devices (Kim et al 2017). In neuromodulation, deep brain stimulation parameters have been efficiently optimized using Bayesian optimization, a sequential design strategy that performs global optimization on an objective function (Connolly et al 2021;Grado, Johnson, and Netoff 2018;Nagrale et al 2023). The statistical model of this typically unknown objective function is a Gaussian process, a stochastic process over random variables jointly distributed as a multivariate normal distribution (Rasmussen and Williams 2005;Garnett 2023), where the random variables are experimental measurements.…”