ESTIMATING PARAMETERS OF A MULTI-CLASS IZHIKEVICH NEURON MODEL TO INVESTIGATE THE MECHANISMS OF DEEP BRAIN STIMULATION

Abstract

The aim of the research is to provide a computationally efficient neural network model for the study of deep brain stimulation efficacy in the treatment of Parkinson's disease. An Izhikevich neuron model was used to accomplish this task and four classes of neurons were modeled. The parameters of each class were estimated using a genetic algorithm with a fitness function based on spike frequency as a function of input current. After computing the optimal parameters the neurons were interconnected to form the network model. The estimated parameters were capable of replicating the normal firing characteristics for each type of neuron, but failed to replicate richer spiking characteristics such as post-inhibitory bursting and tonic firing. Without these characteristics, the network was unable to produce biologically feasible results. Findings indicate the Izhikevich model relies heavily on manual tuning and must be trained under an extensive set of conditions to allow for the majority of spiking characteristics to be learned. The use of the Izhikevich model in a network simulation will always be limited to the characteristics trained on a single neuron. When connected to the network the neuron may be exposed to a variety of unlearned conditions and therefore may not be capable of replicating biologically realistic behavior.