Stochastic models in neurobiology: from a multiunitary regime to EEG data

• Ano de defesa: 2015
• Autor(a) principal: Aline Duarte de Oliveira
• Orientador(a): Jefferson Antonio Galves
• Banca de defesa: Pablo Augusto Ferrari, Eva Locherbach, Jacob Ricardo Fraiman Maus, Claudia Domingues Vargas
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade de São Paulo
• Programa de Pós-Graduação: Estatística
• País: BR
• Link de acesso: https://doi.org/10.11606/T.45.2016.tde-01062016-162919

In this thesis we study three different stochastic processes describing the brain activity. The first one is a continuous time version of the stochastic chains with memory of variable length. These stochastic chains take values in the set of neurons and assign, at time t, the value of the last neuron which spiked up to time t. Moreover, we assume neurons interact through a phenomena called chemical synapses. Briefly this means that when a neuron spikes, it loses all its membrane potential and at same time changes the membrane potential of the neurons which are influenced by it. Under this approach we proved the positive recurrent of the process and presented a perfect simulation algorithm able to generate a finite sample of the process under its invariant measure. In the second model we continue considering the chemical synapses interaction and add also an interaction through electrical synapses. The last one happens duo to the presence of specific channels which allow the passage of ions along the the membrane of two neurons and, as consequence, we have a sharing of potential between the neurons. Moreover, we consider also the constant lost of potential of the neurons for the environment which push each neuron to a resting state. For this model we study the long-run behaviour of the process with a finite number of neurons, the hydrodynamic limit for the system and investigate the possible invariant distributions for the limiting process. In the last model considered here we study the brain activity measured through EEG data. We investigate the predictive coding principle which says that neural networks are able to learn the statistical regularities inherent in a stimuli and reduce redundancy by removing the predictable components of the input. To test this conjecture we propose procedures to perform statistical model selection on the EEG data in order to retrieve structural features of stochastic sources. This is done through a case study in which the EEG data is recorded under the effect of two different stochastic rhythmic sources produced by two different context tree models. We present a suitable class of stochastic processes, called here as hidden context tree models, to model EEG signals evoked by rhythmic structures. Then, we propose a consistent statistical procedure to perform statistical model selection in this class and in our case study.