Synthetic neuronal datasets for benchmarking directed connectivity metrics
- Published
- Accepted
- Subject Areas
- Biophysics, Neuroscience, Computational Science
- Keywords
- neuronal modeling, computational modeling, hemodynamic response function, EEG forward modeling, Granger causality
- Copyright
- © 2014 Rodrigues et al.
- Licence
- This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, reproduction and adaptation in any medium and for any purpose provided that it is properly attributed. For attribution, the original author(s), title, publication source (PeerJ PrePrints) and either DOI or URL of the article must be cited.
- Cite this article
- 2014. Synthetic neuronal datasets for benchmarking directed connectivity metrics. PeerJ PrePrints 2:e737v1 https://doi.org/10.7287/peerj.preprints.737v1
Abstract
Background: Datasets consisting of synthetic neural data generated with quantifiable and controlled parameters are a valuable asset in the process of testing and validating directed connectivity metrics. Considering the recent debate in the neuroimaging community concerning the use of directed functional connectivity metrics for fMRI data, synthetic datasets that emulate BOLD dynamics have played a central role by supporting claims that argue in favor, or against, certain metrics. Generative models often used in studies that simulate neuronal activity, with the aim of gaining insight into specific brain regions and functions, have different requirements from the generative models for benchmarking datasets. Even though the latter must be realistic, there is a tradeoff between realism and computational demand that needs to be contemplated and simulations that efficiently mimic the real behavior of single neurons or neuronal populations are preferred, instead of more cumbersome and marginally precise ones. Methods: this work explores how simple generative models are able to produce neuronal datasets, for benchmarking purposes, that reflect the simulated effective connectivity and, how these can be used to obtain synthetic recordings of EEG and fMRI BOLD. The generative models covered here are AR processes, neural mass models consisting of linear and non-linear stochastic differential equations and populations with thousands of spiking units. Forward models for EEG consist in the simple three-shell head model while fMRI BOLD is modeled with the Balloon-Windkessel model or by convolution with a hemodynamic response function. Results: the simulated datasets are tested for causality with the original spectral formulation for Granger causality. Modeled effective connectivity can be detected in the generated data for varying connection strengths and interaction delays. Discussion: all generative models produce synthetic neuronal data with detectable causal effects although the relation between modeled and detected causality varies and less biophysically realistic models offer more control in causal relations such as modeled strength and frequency location.
Author Comment
This is the first version of a submission to PeerJ for review.
Supplemental Information
Figure 1: Data simulation
Strategies used for synthetic neural data modeling. LFPs are simulated by four distinct generative models and the resulting time-series can be used by EEG or BOLD forward models to produce the respective signals.
Figure 2: Three layer spherical head model
Three layer spherical head model for one current dipole with radius rq and scalp electrode with radius r. Adapted from (Mosher, Leahy & Lewis, 1999).
Figure 3: GGC DOI for LFP signals with increasing coupling strength
GGC DOI for the time-series simulated with increasing coupling strength with the generative models: A) AR models, B) KIe,i sets, C) Kuramoto oscillators, D, E and F) Izhikevich columns.
Figure 4: GGC DOI for LFP signals with increasing interaction delay
GGC DOI for the time-series simulated with increasing interaction delay with the generative models: A) AR models, B) KIe,i sets, C) Kuramoto oscillators, D) Izhikevich columns.
Figure 5: Model orders for LFP signals with increasing interaction delay
Figure 5: Model orders estimated with BIC and AIC for the time-series simulated with increasing interaction delay with the generative models: A) AR models, B) KIe,i sets, C) Kuramoto oscillators, D) Izhikevich columns. The lagged observations for these interaction delays with 250 HZ sampling rate are [1, 5, 10, 15, 20, 25].
Figure 6: GGC DOI for EEG signals with increasing coupling strengths
GGC DOI after EEG forwarding the time series from the generative models with varying coupling strengths: A) AR models, B) KIe,i sets, C) Kuramoto oscillators, D, E and F) Izhikevich columns.
Figure 7: GGC DOI for EEG signals with increasing interaction delays
GGC DOI after EEG forwarding the time series from the generative models with varying interaction delays: A) AR models, B) KIe,i sets, C) Kuramoto oscillators, D) Izhikevich columns.
Figure 8: Model orders for EEG signals with increasing interaction delay
Model orders estimated with BIC and AIC for the time-series simulated with increasing interaction delay after EEG forwarding with the generative models: A) AR models, B) KIe,i sets, C) Kuramoto oscillators, D) Izhikevich columns. The lagged observations for these interaction delays with 250 HZ sampling rate are [1, 5, 10, 15, 20, 25].
Figure 9: GGC DOI for BOLD signals with increasing coupling strengths
GGC DOI in the 0.01-0.1 Hz band after BOLD forward modeling the time series from the generative models with varying coupling strengths: A) AR models, B) KIe,i sets, C) Kuramoto oscillators, D, E and F) Izhikevich columns.
Figure 9: GGC DOI for BOLD signals with increasing interaction delays
GGC DOI in the 0.01-0.1 Hz band after BOLD forward modeling the time series from the generative models with varying interaction delays: A) AR models, B) KIe,i sets, C) Kuramoto oscillators, D) Izhikevich columns.
