Identifiability in connectome based neural mass models

Local dynamic activity within canonical micro-circuits in the brain can be described mathematically by neural mass models with parameters that introduce a variety of oscillatory behavior in local neuron populations. Advances in medical imaging have enabled quantification of the white matter connections that constitute whole brain networks or the "connectome". Recently, connectome-derived coupling terms have been introduced within an array of neural mass models to capture the long-range interactions between local neuronal populations. Although such network-coupled oscillator models are capable of producing steady-state power spectra similar to the brain9s empirical activity, it is unclear if the connectome9s anatomical information is enough to recapitulate the spatial distribution of power spectra across brain regions. Furthermore, these models inherently comprise of hundreds of parameters whose choices have impact on model derived predictions of brain activity. Here we employ a Wilson-Cowan oscillator neural mass model coupled by a structural connectome network to observe the effect of introducing a connectivity and transmission delay to the frequency profile of the brain. We observe that inference of the many parameters of the high dimensional network model produces non-unique results. Parameter optimization of simulated power spectra to better match source localized EEG spectra showed that introducing structural information to neural mass models does not improve model performance. A combinatorial approach to optimizing local and global parameters outperforms other model variations. We demonstrate the inherent identifiability problem in network models that pose challenges for the use of such high dimensional models as diagnostic tools for neurological diseases.