Spatiotemporal dynamics of human high gamma discriminate naturalistic behavioral states

Figures

• 

  Fig 1. Simulated electrodes visualized for two different behavioral states. (A) Two simulated electrodes are shown to visualize different multi-variate signal components that could differentiate between two behavioral states. There is a difference between the electrodes’ mean and variance when comparing states 1 and 2. Two spatiotemporal features are present at different rates in both states. There is a slow component where the electrodes are positively correlated and a fast component where the electrodes are negatively correlated. In state 1, the slow component is more prominent, while in state 2 the fast component is more prominent. (B) The values of the electrodes across both states are shown in a scatter plot. This method of visualizing electrodes could provide valuable details about the differences between the two states spatially, but it lacks any information regarding the temporal dynamics at play.

• 

  Fig 2. ECoG grid and sEEG shanks locations. (A) shows the coverage of the ECoG grid on the cortex for Subject 1 as well as the corresponding channel numbers. (B) shows the location of where the sEEG shanks penetrated the cortex for Subject 2 as well as the shank alphabetical labels.

• 

  Fig 3. Decoding performance using first and second-order statistics. The decoding performances of Subjects 1 and 2 are shown in (A) and (B), respectively. All five different spatial and temporal components in the neural activity are able to discriminate between abstract behavioral states. The slow-varying mean and variance achieves the highest decoding performance. The covariance matrix is also able to separate behavioral states without leveraging single channel variances, as the variances are all normalized to one prior to classification. Therefore, inter-electrode correlations contribute to behavioral state separability. Finally, both the high gamma band amplitude estimate’s slow and fast components contain information regarding the behavioral state.

• 

  Fig 4. Power spectral density estimates of binned high gamma amplitude estimate. Sample electrodes are shown for Subjects 1 and 2 in (A) and (B), respectively. The power spectral densities in both sample electrodes show that different behavioral contexts exhibit varying temporal dynamics at different timescales. For example, in (A), “Dialogue” and “Rest” have more power in comparison to “Electronics” and “TV” for temporal dynamics slower than 3 seconds but vice versa for dynamics faster than 3 seconds. Most of the signal variance is concentrated in the slower components for both subjects. For Subject 1, the example electrode is channel 37 and for Subject 2 the example electrode is the bipolar pair E’9-E’10. For Subject 2, a bipolar pair with a larger number corresponds to a bipolar pair that is closer to the cortical surface. See Fig 2 for the anatomical locations of the ECoG grid and sEEG shanks. The standard error of the power spectral density estimate as calculated prior to averaging the periodograms in the power spectral density calculation. It is shown in the shaded regions.

• 

  Fig 5. Root mean square error vs. latent state dimensionality determined from FA, GPFA, and reduced GPFA. (A) and (B) show the RMSE using the leave-one-electrode-out method for Subjects 1 and 2, respectively. For both subjects, the error decays as a function of state dimensionality, indicating the presence of complex spatial components in the neural activity. Additionally, for both subjects, the RMSE in FA is higher than GPFA for all dimensions, and the difference in RMSE increases as more latent states are learned. The orthogonal factors determined using reduced GPFA further reduces the error and shows that there exists a smaller set of spatial components that could reconstruct the neural activity as well as its higher dimension non-orthogonal counterparts.

• 

  Fig 6. Number of latent factors vs. decoding accuracy. The decoding accuracy as a function of the number of factors generated by GPFA used (by incremental addition from slowest to fastest factors and vice versa) is shown in (A) and (B) for Subjects 1 and 2, respectively. Using a quadratic discriminant analysis as the classification algorithm, the separability of 30-second epochs increases as a function of the number of factors used. Additionally, for both subjects, there is additive information in both fast and slow factors. As a control, we used an linear discriminant analysis (LDA) to show that separability is due to differences in the latent neural trajectories’ power rather than their mean (which would indicate a behavioral state-specific shift in variance or mean in the data). Standard errors of the decoding accuracies across folds are shown in the shaded region. Finite chance performance is also shown in the black line.

• 

  Fig 7. Decoding accuracy of single factors. The decoding accuracy of single factors generated by the GPFA algorithm is shown for Subjects 1 and 2 in (A) and (B), respectively. The factors are sorted from slowest to fastest. The parameter tau is fitted using the EM algorithm for GPFA and is expressed in milliseconds. We show that there is relevant information in latent factors spanning different timescales. Additionally, several factors are extracted from the data that do not contain any information regarding behavioral states. The accuracies shown are the averages of the single factor accuracies sorted from slowest to fastest across the cross validation folds. Finite chance performance is also shown in the black line.

• 

  Fig 8. Visualizations of latent factors extracted from GPFA. We show the values of the factor weights extracted from GPFA for Subject 1 and 2 in (A) and (B), respectively. The latent factors’ values show that there is synchronized activity corresponding to different areas of the brain. Additionally, some factors have both positive and negative values assigned to different channels, indicating that some factors exhibit synchronized activation and deactivation of certain brain regions. Refer to Fig 2 for the anatomical locations of the ECoG electrodes for Subject 1 and the sEEG shanks for Subject 2. Note that a bipolar pair with a larger number is closer to the cortical surface.

• 

  Fig 9. Sample latent neural trajectories with variable temporal dynamics. For Subject 1, we show the fastest and slowest latent factors extracted from the data. The factors’ spatial weights are shown in the cortex in (A) and (C). Sample trajectories of the fastest and slowest factor of a single epoch are shown in (B) and (D) overlaid with the channel’s value that contributes most to the corresponding factor. The factor in (B) is inverted in order to match the sign of the channel for visualization purposes. In (B), the latent factors track many high frequency peaks in the single channel amplitude. Additionally, (D) shows that the slow latent factor tracks the slow moving component of the channel amplitude.