Data-driven neural mass modelling

Apr 06, 2019

Lara Escuain Poole, Presentation date: March 28, 2019

Author: Lara Escuain Poole
Title: Data-driven neural mass modelling
Directors: Prof. Antonio Javier Pons Rivero, Prof. Jordi Garcia-Ojalvo
Presentation date: March 28, 2019
Link to text: https://www.tdx.cat/handle/10803/666615


Abstract: The brain is a complex organ whose activity spans multiple scales, both spatial and temporal. The computational unit of the brain is thought to be the neurone. At the microscopic level, neurones communicate via action potentials. These may be observed experimentally by means of precise techniques that work with a small number of these cells and their interactions, and that can be modelled mathematically in a variety of ways. Other techniques consider the averaged activity of large groups of neurones in the mesoscale, or cortical columns; theoretical models of these signals also abound. The problem of relating the microscopic scale to the mesoscopic is not trivial. Analytical derivations of mesoscopic models are based on assumptions that are not always justified. Also, traditionally there has been a separation between the clinically oriented analysts that process neural signals for medical purposes and the theoretical modelling community. This Thesis aims to lay bridges both between the microscopic and mesoscopic scales of brain activity, and between the experimental and theoretical angles of its study. This is achieved via the unscented Kalman filter (UKF), which allows us to combine knowledge from different sources (microscopic/mesoscopic and experimental/theoretical). The outcome is a better understanding of the system than each of the sources of information could provide separately. The Thesis is organised as follows. Chapter 1 is a brief reflection on the current methodology in Science and its underlying motivations. This is followed by chapters 2 to 4, which introduce and contextualise the concepts discussed in the remainder of the work. Chapter 5 tackles the interrelationship of the microscopic and mesoscopic scales. Although efforts have been made to derive mesoscopic equations from models of microscopic networks, they are based on assumptions that may not always hold. We use the UKF to assimilate the output of microscopic networks into a mesoscopic model and study a variety of dynamical situations. Our results show that using the Kalman filter compensates for the loss of information that is common in analytical derivations. Chapters 6 and 7 address the combination of experimental data with neural mass models. More specifically, we extend Jansen and Rit's model of a cortical column with a model of the head, which allows us to use electroencephalography (EEG) data. With this, we estimate the state of the system and a relevant parameter of choice. In chapter 6 we use in silico data to test the UKF under a variety of dynamical conditions, comparing simulated intracranial data with simulated EEG. Extracranial estimation is always superior in speed and quality to intracortical estimation, even though intracortical electrodes are closer to the source of activity than extracranial electrodes. We suggest that this is due to the more complete picture of the cortex that is visible with the set of extracranial electrodes. Chapter 7 feeds experimental EEG data of an epileptic patient into Jansen and Rit's model; the goal is to estimate a parameter that governs the dynamical behaviour of the system, again with the UKF. The estimation of the state closely follows the experimental data, while the parameter shows sensitivity to the changes in brain regimes, especially seizures. These results show promise for using data assimilation to address some shortcomings of brain modelling techniques. On the one hand, the mutual influence of neural structures at the microscopic and the mesoscopic scales may become better characterised, by means of filtering approaches that bypass analytical limitations. On the other hand, fusing experimental EEG data with mathematical models of the brain may enable us to determine the underlying dynamics of observed physiological signals, and at the same time to improve our models with patient-specific information. The potential of these enhanced algorithms spans a wide range of brain-related applications.