Bayesian optimization to select electrical stimulation parameters for neuromodulation

Neuromodulation devices, such as deep brain stimulation, vagal nerve stimulation, spinal cord stimulation, transcranial magnetic stimulation, and direct cortical stimulation, have many parameters, such as therapeutic location, stimulation frequency, amplitude, and pulse width. For many indications, the approved range of these parameters is limited, but it may be that significant improvements in therapy are achievable by personalizing stimulation to each patient's particular needs and physiology. However, with such a large parameter space, it is difficult to know how to approach optimizing and recording the results in an efficient way that accounts for previous information while suggesting new settings. Bayesian optimization is a black box optimization approach that is efficient and powerful. It builds a response surface model by fitting the response of a biomarker to tested stimulation parameters. It then uses that model to suggest future settings to balance exploitation, using settings that are near settings that are known to be good from previous testing, and exploration, testing parameters in the space that have not been tested to ensure that regions that could contain a good setting are tested. In this chapter, we will detail how to use Bayesian optimization and how we have used it for optimizing deep brain stimulation settings of the anterior nucleus of the thalamus for treatment of epilepsy, cortical stimulation for treatment of depression, deep brain stimulation settings for Parkinson's disease, and spinal cord stimulation settings for spinal cord injury.