Investigation of Architectures for Models of Neural Responses to Electrical Brain Stimulation

Electrical brain stimulation is used clinically to target pathological regions of the brain for treatment of diseases, such as Parkinson's disease, epilepsy and depression. Conventional treatments involve chronic implants that disrupt activity through a fixed periodic train of pulses or bursts of pulses applied to the affected region. However, stimulating one region of the brain necessarily affects other structurally and/or functionally connected areas. Understanding how connected regions of the brain are affected by stimulation at the implant site could improve treatment efficacy by informing optimal placement and stimulation patterns. In this study, we build predictive input-output models from intracranial recordings obtained from 10 epilepsy patients implanted with electrodes. Specific contacts within each subject were electrically stimulated (inputs), and evoked responses were simultaneously captured from all contacts (outputs). From these data, we constructed and compared four different dynamical models that contain causal linear and nonlinear components. All model architectures successfully predicted evoked responses to stimulation with single pulses and sequences of pulses. Results suggest that a linear time-invariant model in series with a quadratic non-linearity best captures the relationship between stimulation amplitudes and evoked responses.