Poster B69, Thursday, August 16, 3:05 – 4:50 pm, Room 2000AB

Quantitative Assessment of Cognitive Models with Neuroimaging Data

Frank H Guenther1, Ayoub Daliri2, Alfonso Nieto-Castanon1, Megan Thompson1, Jason A Tourville1;1Boston University, 2Arizona State University

Advances in neuroimaging techniques such as functional magnetic resonance imaging (fMRI) and electrocorticography (ECoG) over the past two decades have resulted in a greatly improved understanding of the neural mechanisms underlying human sensory, motor, and cognitive capabilities, leading to increasingly sophisticated neural models of these functions. Within the domain of speech, we have developed and refined a large-scale neurocomputational model, called the Directions into Velocities of Articulators (DIVA) model, which provides a unified mechanistic account of acoustic, kinematic, and neuroimaging data on speech [for a detailed review see Guenther (2016) Neural control of speech, MIT Press]. Functional neuroimaging has been a powerful means for evaluating and refining such models. To date, however, these evaluations have been almost exclusively qualitative. Quantitative evaluations have been hampered by the absence of a general computational framework for (i) generating predicted functional activation from a model that can be directly and quantitatively compared to empirical functional neuroimaging data, and (ii) testing between models to identify the model that best fits experimental data. Here we present a general computational framework to overcome these issues. Within this framework, the brain network responsible for a task is broken into a set of computational nodes, each of which is localized to an MNI stereotactic coordinate in the brain. Associated with each node is a computational load function that links the node’s activity to a computation involving quantifiable measures from the task. The instantaneous neural activity at each location in the brain (e.g., each voxel of an fMRI image or each electrode of an ECoG array) is then calculated by summing the contributions of all model nodes at that location, with each node treated as a Gaussian activity source centered at the node’s location. The parameters of the Gaussians (i.e., spread and magnitude of activation) are optimized to produce the best fit to the functional data. Model comparisons are based on the overall fit level and number of free parameters using the Akaike Information Criterion (AIC). This framework was used in conjunction with a large fMRI database of speech production studies (116 speakers) to illustrate the DIVA model’s ability to provide a unified account for whole-brain activity patterns seen during speech production under normal and perturbed conditions. Additionally, the activation foci from two prior meta-analyses of speech production neuroimaging experiments [Turkeltaub et al. (2002) NeuroImage, 16: 765-780; Brown et al. (2005) Hum Brain Mapp, 25:105-117] were used to construct two new, simplified models of speech production which were fit to the same dataset to illustrate how easy it is to create a quantitative model for fitting fMRI data. All models were then compared using the AIC, with the results highlighting the advantages of functional models like DIVA in accounting for data from multiple speaking conditions. [Supported by NIH grants R01 DC002852, R01 DC007683.]

Topic Area: Computational Approaches