A Gaussian Process Model of Human Electrocorticographic Data

Abstract

We present a model-based method for inferring full-brain neural activity at millimeter-scale spatial resolutions and millisecond-scale temporal resolutions using standard human intracranial recordings. Our approach makes the simplifying assumptions that different people’s brains exhibit similar correlational structure, and that activity and correlation patterns vary smoothly over space. One can then ask, for an arbitrary individual’s brain: given recordings from a limited set of locations in that individual’s brain, along with the observed spatial correlations learned from other people’s recordings, how much can be inferred about ongoing activity at other locations throughout that individual’s brain? We show that our approach generalizes across people and tasks, thereby providing a person- and task-general means of inferring high spatiotemporal resolution full-brain neural dynamics from standard low-density intracranial recordings.

Distributed Subnetworks of Depression Defined by Direct Intracranial Neurophysiology

“…Using leave-one-patient out validation of the correlational model, we found that the distribution of correlations (mean r = 0.38) was similar to the prior reconstruction accuracies (Owen et al, 2020) and centered well above shuffled correlational models (mean r = 0.00) suggesting the algorithm estimates activity patterns substantially better than chance. The distribution of patient level fisher transformed correlation coefficients was significantly different than 0 (t = 13.94, p = 1.04e −25 , Figure 2F).…”

“…For the model building step, we used a functional connectivity imputation technique, called SuperEEG ( Owen et al, 2020 ) to map continuous iEEG recordings from different patients into a common neural space ( Figure 2 ). This method provided an important advance over previous iEEG studies ( Kirkby et al, 2018 ; Sani et al, 2018 ; Scangos et al, 2019a ) that were limited to region-based analyses conducted in small samples due to heterogeneous electrode placement.…”

“…This method provided an important advance over previous iEEG studies ( Kirkby et al, 2018 ; Sani et al, 2018 ; Scangos et al, 2019a ) that were limited to region-based analyses conducted in small samples due to heterogeneous electrode placement. To generate this model, pre-processed iEEG signals were chunked into 60 s non-overlapping blocks and filtered for putative epileptiform activity or artifacts using kurtosis, a measure of infrequent extreme peaked deviations ( Akbarian and Erfanian, 2018 ; Owen et al, 2020 ). We then randomly sampled the 60 s intervals across daytime hours (8am–10pm) and concatenated them into 2-h blocks, each representative of naturalistic activity.…”

“…We then used Gaussian process regression based on the population-level model and individual time series for each subject to reconstruct whole-brain local field potentials for each subject. Evaluation of the SuperEEG algorithm has been performed previously on two large independent iEEG datasets using leave-one-out cross-validation ( Owen et al, 2020 ). Reconstruction accuracy was measured by calculating the correlation between the true and reconstructed signals for each held-out electrode from the held-out patients.…”

“…We hypothesized that we could apply a novel computational approach to a large unique dataset of multi-region, multi-day iEEG recordings in 54 participants to uncover distributed cortico-subcortical networks in depression. To tackle inconsistent network sampling across individuals, we utilized a method called SuperEEG ( Owen et al, 2020 ) that uses the correlational structure of brain activity across the population to create a model of multiregional iEEG activity for each individual despite heterogeneous electrode placement. This model provided a highly detailed representation of brain activity across space and time and allowed us to chart out the inherent organization of the brain into functional networks.…”