Rationale:
Due to the complex spatiotemporal dependencies between different brain areas, fully comprehending the interplay between structure and function in epilepsy is a challenging and intense area of research. Quantitatively mapping the dynamic interactions between different brain regions using data-driven approaches can provide an inter-region measure of connectivity strength, which can be used to help identify neural activity patterns in epilepsy and non-epilepsy patients. Deep learning and statistical models have enabled performance leaps in the analysis of high-dimensional functional Magnetic Resonance Imaging (fMRI) data, but the advent of new transformer models and deep learning frameworks for network generation allow for new methods of feature selection of brain regions across diverse time scales.
Methods:
Parcellated fMRI Resting scans in control and epilepsy patients were inputted in various network building models. The problem of identifying brain-network topology was decomposed into a variety of feature selection problems. The activity pattern of a set of parcels (target) over time was predicted from the activity patterns of the other parcels (input) using a variety of transformer models, including a Temporal Fusion Transformer (TFT). From the TFT the importance of different input parcels was extracted at different time steps and used as putative regulatory links in an undirected functional connectivity graph adjacency matrix. Hopfield Network models were also used to perform feature selection and model connectivity relationships using both resting fMRI time-series for neurotypical patients and functional connectivity correlations in epilepsy and neurotypical resting fMRI data. Results:
A transformer model was run on 339 resting-state fMRI with max-min normalized activity in neurotypical subjects with brains parcellated into 360 regions according to the Montreal Neurological Institute (MNI) space framework with 1200 time steps. Reductions in feature selection algorithm time were studied using functional connectivities obtained by correlation in resting state epileptic and non-epileptic patients. This model generated an adjacency matrix between brain regions with a loss of 0.0013, measured by prediction of a randomly removed target parcel. To highlight the importance of various features toward the target parcel prediction at various time points, a Temporal Fusion Transformer (TFT) was used. Running on a single CPU the TFT was able to generate feature importances for a single parcel with a loss of 0.005. The time series activity of parcels was also modeled by a Hopfield Network, in which the network nodes were modeled as parcels and patterns of activity were incorporated into the network as associative memories. Running on 300 timesteps across these parcels yielded a model with an accuracy of 88%. Conclusions:
This project uses hopfield networks and time-series-based transformer networks to model resting state fMRI activity with high accuracy. These findings are significant because new methods of analyzing functional connectivity in resting state control and epilepsy patients can help identify novel patterns of activity across patient states.
Funding: None