Article Figures & Data
Figures
- Figure 1.
A–C, Individual pial surface and electrode reconstructions. One representative electrode was selected from each of the following regions: early visual cortex (blue), middle fusiform gyrus (orange), pars triangularis (yellow), pars opercularis (purple), ventral sensorimotor cortex (green), and auditory cortex (red). D, Picture naming stimuli: coherent (left) and scrambled (right).
- Figure 2.
A, Graphical representation of an ARHMM with autoregressive (AR) order 2: latent states z and observations x evolve according to state transition matrix Φ, AR coefficients A, and process covariance Q. B, Illustration of the ARHMM latent state space model. C, Simulated time series of data points emitted by three latent states (top) with inferred state probabilities (bottom).
- Figure 3.
Comparison of state clustering for simulated data. A, Underlying data with varying state durations, and mean-dependent variance with associated state sequence and network states on the right. B, Ground truth: rasters of discrete latent states over time and trials and graphs of network interactions for each latent state, measured by PDC. C, Estimates inferred from data with four different SNRs (0.3, 1.5, 3, and 30). Estimates are obtained from mean-subtracted time series. Color represents the discrete states, and confidence (responsibility) is encoded by brightness and increases with SNR.
- Figure 4.
GBA at salient brain regions following picture presentation. A, Trial-averaged GBA, relative to baseline. Each articulation onset is indicated by a vertical blue line below (mean 1.2 s), and the visual stimulus is presented during the black interval. B, Density plot of trial-wise GBA, z-scored.
- Figure 5.
Model selection. BIC (A) and held-out log-likelihood (B) as a function of number of states. C, BIC as a function of AR model order (number of time delays).
- Figure 6.
State sequences in brain activity. A, Time dependence of most probable states from left and right hemispheric brain activity for three subjects. For each of six brain regions, we selected the most active electrode, fit the ARHMM to the broadband γ power on these six electrodes, and then estimated the sequence of latent brain states that best explains the observed activity. The top shows states as a trial-by-trial raster, and the bottom shows the fraction of trials on which each state was most probable. B, Interactions between brain regions during the corresponding named network states, plotted as in Figure 3C. Black arrows indicate state transition probabilities according to the inferred state transition matrix. C, Dissimilarity between brain states between and within subjects. Dissimilarity is measured as the difference between integrated PDC magnitude, according to Equation 7. Extended Data Figure 6-1 shows the same analysis, using not the most active electrodes in each region but instead the multi-electrode activity patterns with greatest variance within each region.
- Figure 7.
Duration of the active states; (A) visual processing, (B) language processing, and (C) articulation, as well as (D) the termination of the “language” state, compared to reaction times. Two patients’ left hemispheres (blue and orange) are plotted. The Pearson correlation coefficient r and p value for the null hypothesis of uncorrelated values are shown.
- Figure 8.
Multi-subject integration (anatomic grouping) of the two left-hemispheric recordings for coherent (top left) and scrambled (bottom left) stimulus condition. Differences in information flow between both stimulus conditions are shown on the right, where colored and gray arrows denote excess activity in the coherent and scrambled condition, respectively.
- Figure 9.
Comparison of AMVAR and ARHMM estimates of activity in Broca’s area during articulation. The ARHMM shows stronger total interactions than AMVAR analysis, especially for Broca’s area. A, Total incoming and outgoing activity (red and blue, respectively) for Broca’s pars opercularis (pOp) and pars triangularis (pTr) during articulation, according to AMVAR and ARHMM models. B, Connectivity graphs for AMVAR (top) and ARHMM (bottom), shown for two left-hemispheric patients for the language processing (red) and articulation state (green).
Extended Data
Extended Data 1
Code for the ARHMM. Download Extended Data 1, ZIP file.
Extended Data Figure 6-1
State sequences in brain activity, plotted as in Figure 6, but for meta-electrodes created by selecting the principal component for subsets of electrodes in each brain region. States estimated from meta-electrodes produced crisper states (A) but were less similar across subjects (B, C) than observed when selecting single electrodes with strongest signals (Fig. 6). Download Figure 6-1, EPS file.
In this issue
