(534a) Designing Stochastic Model Predictive Control Based Neural Interface to Restore Communication between Brain Regions

Cerebellar (ischemic) stroke results in the loss of brain functions, which ultimately leads to severe disability in stroke survivors. An abundance of evidence shows a cerebellar stroke damages the postsynaptic dendrites of neurons, even in the absence of neuronal loss, and the synapses through which neurons communicate. In this work, we theoretically explore a possibility of whether an optimal controller can be used as an interface between two brain regions to restore the lost communication due to the loss of synapses. Particularly, we consider two interconnected networks (Network 1 and Network 2), each consists of 100 spiking neurons, as the representation of two distinct brain regions. Neurons in each network are synaptically connected within the network with probability 0.1 (sparse connectivity). Similarly, neurons in Network 1 are synaptically connected to neurons in Network 2 and vice-versa with probability 0.1. Neurons in both networks receive Poisson inputs which represent the inputs from other cortical areas of the brain. We first simulate the whole network and record the firing activity (timing of action potentials) of neurons in Network 1 before any loss in the number of synapses between Network 1 and Network 2. We then randomly remove a fraction of synapses (~30%) from Network 1 to Network 2 which results in altered dynamics and a poor communication between these two networks. Our objective is to restore this lost communication by developing a neural (brain-brain) interface where an optimal controller uses the firing activity of neurons from Network 1 to stimulate a fraction of neurons in Network 2. We particularly consider closed-loop optogenetic stimulation of neurons in Network 2 because of its capability in activating neurons selectively within a population. To design our controller, we first stimulate the neurons in Network 2 using a variety of stimulation patterns and record the firing activity (the timing of action potentials) of neurons in Networks 1 and 2. We then use these input-output data to develop a point-process generalized linear model (GLM) of neurons in Network 1. Using this model, we formulate an optimal control problem within the framework of stochastic model predictive control to restore the firing patterns of neurons in Network 1. We solve our control problem numerically and demonstrate the capability of the designed controller in restoring the lost communication between these two networks.