A Large Multiscale Neural Model Inversion Framework for Quantifying Excitation-Inhibition Balance

1821 

Abstract Submission 

Guoshi Li¹, Pew-Thian Yap¹

¹University of North Carolina at Chapel Hill, Chapel Hill, NC

Guoshi Li  
University of North Carolina at Chapel Hill
Chapel Hill, NC

Pew-Thian Yap  
University of North Carolina at Chapel Hill
Chapel Hill, NC

Excitation-inhibition (E-I) balance is a fundamental property of neuronal circuits and abnormal E-I balance has been hypothesized to be a key driver for multiple neurological and mental disorders. However, no computational models have been established so far to measure E-I balance in large-scale neuronal circuits based on noninvasive individualized neuroimaging. To close this important gap, we developed a large Multiscale Neural Model Inversion (LMNMI) framework for E-I estimation based on resting-state functional MRI (rs-fMRI) by extending a previous small-scale MNMI model (Li et al., 2011). The validity of the LMNMI model was evaluated by both ground-truth simulation and empirical analysis of E-I imbalance in Alzheimer's disease (AD).

In the LMNMI framework (Fig. 1A) the neural network dynamics is described by a discrete linearized neural mass model (NMM) of Wilson-Cowan type with a time step of Δt (Δt=TR) (Galán, 2008). Each network node contains two mutually coupled excitatory and inhibitory neural populations and the excitatory neural populations are connected via long-range fibers. First, structural connectivity from diffusion MRI is used to construct a sparse network by removing weak inter-regional connections. Second, empirical BOLD signals are Wiener-deconvolved to obtain composite neural activity y ̂(t). Third, Kalman filter is applied to estimate neural activity x(t) and the predicted error of composite neural activity is calculated. Lastly, connection parameters (W) are optimized by minimizing the prediction error using a gradient descent algorithm. For ground-truth simulation, synthetic neural activities were generated using the NMM with model parameters drawn from a uniform distribution. For empirical analysis, we used rs-fMRI data from the ADNI dataset including 48 normal control (NC) (26/22 males/females, 73.4 ± 6.5 years), 48 mild cognitive impairment (MCI) (27/21 males/females, 73.9 ± 10 years) and 48 AD subjects (27/21 males/females, 73.6 ± 8.6 years). Regional BOLD time series were extracted using the Desikan-Killiany atlas (Desikan et al., 2006) based on 46 regions covering the default model, salience, frontoparietal control and limbic networks (Yeo et al., 2011). One-way analysis of variance (ANOVA) was used to compare the E/I means of the three groups followed by post-hoc analysis with two-sample t-test. Multiple comparisons were corrected by controlling FDR with q < 0.05.

The Kalman filter was able to accurately track ground-truth neural activity (Fig. 1B) and the parameter estimation error rapidly converged to the minimum (Fig. 1C). The computation time for a network with 50 regions took about 35 minutes (20,000 iterations) when run on a standard computer. The estimated connection parameters closely matched the ground-truth parameters for a representative synthetic subject (r > 0.6; Fig. 1D-F). Application of the LMNMI model to the ADNI dataset indicated that recurrent excitation (W_EE) in both MCI and AD significantly decreased compared to NC (p<0.05, FDR corrected) for most of the brain regions (Fig. 2A). In contrast, no significant difference was observed for recurrent inhibition (W_IE) among the three groups (Fig. 2B). Consequently, the overall E/I ratio in both MCI and AD was significantly reduced from NC (p<0.05, FDR corrected) in a number of regions including the right precuneus and right putamen (Fig. 2C). The wide-spread reduction in excitation is consistent with the progressive disruption of synaptic transmission during AD progression (Sheng et al., 2012).

We developed a new LMNMI framework for large-scale E-I estimation based on rs-fMRI and validated its efficiency and accuracy using both ground-truth simulation and empirical analysis. This framework offers a highly efficient yet biologically realistic method to construct brain-wide individualized neural network for disease diagnosis and the identification of circuit dysfunction in neurological and psychiatric disorders.

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) ²

Connectivity (eg. functional, effective, structural)

fMRI Connectivity and Network Modeling ¹

Methods Development

Computational Neuroscience

Degenerative Disease

FUNCTIONAL MRI

Modeling

Other - Excitation-inhibition balance