The Journal of Neuroscience, December 8, 2004, ():
Global Forebrain Dynamics Predict Rat Behavioral States and Their Transitions
J. Neurosci. Gervasoni et al.
24: 11137
Supplemental data
Supplementary Information references
Fanselow EE, Nicolelis MA (1999) Behavioral modulation of tactile responses in the rat somatosensory system. J Neurosci 19:7603-7616.
Halasz P (1998) Hierarchy of micro-arousals and the microstructure of sleep. Neurophysiol Clin 28:461-475.
Robert C, Guilpin C, Limoge A (1999) Automated sleep staging systems in rats. J Neurosci Methods 88:111-122.
Timo-Iaria C, Negrao N, Schmidek WR, Hoshino K, Lobato de Menezes CE, Leme da Rocha T (1970) Phases and states of sleep in the rat. Physiol Behav 5:1057-1062.
Winson J (1974) Patterns of hippocampal theta rhythm in the freely moving rat. Electroencephalogr Clin Neurophysiol 36:291-301.
Files in this Data Supplement:
- supplemental material
-
Figure S1.
Diagrams of spontaneous transitions between behavioral states for 120 hours of a rat’s life, determined by the classical classification approach combining the observation of the overt behavior and analysis of LFP spectral features (Timo-Iaria et al., 1970; Winson, 1974; Fanselow and Nicolelis, 1999). Each state is represented by a circle, with area proportional to the amount of time spent in that state (% indicated in italic). Arrows indicate transitions from one state to another. Non-italicized numbers represent the relative occurrence probability of specific transitions. In agreement with previous studies, we found that rats spent ~60% of the day (light period) sleeping, and ~60% of the night (dark period) awake. Some state transitions are highly prevalent (e.g. AE?QW, QW?SWS, QW?WT, SWS?REM and REM?QW), while others are either very rare (SWS?AE, REM?AE and WT?AE) or absent (SWS?WT and AE?REM). REM episodes are nearly always terminated by QW.
- supplemental material
-
Figure S2.
(A) Effect of the size of the smoothing window on the separation of state clusters. We systematically investigated the effects of varying the length of the Hanning smoothing window from 1-50 sec. Linear discriminant analysis (LDA) was used to calculate the misclassification probability when an optimal linear decision boundary was used to discriminate selected pairs of states. The LDA error rates for each selected pairs of states were normalized between [0, 1], with the maximum error (in the unsmoothed case) corresponding to 1 and the minimum error after smoothing corresponding to 0. We found a nearly exponential decay of state classification error as the length of the smoothing window increases. In this exponential decay, a 20-sec window already reduces 80% of the total error reduced by a 50-sec window. Therefore, we chose this 20-sec window as a compromise between classification accuracy and the preservation of temporal resolution. The effective resolution of our method after smoothing with a 20-sec Hanning window is in the order of 5-10 sec, depending on how distinct are the spectral characteristics of the transient fluctuations considered. Even though the smoothing tends to neglect unstable state fluctuations that do not correspond to global state changes, a distinct transient state such as IS (with a typical duration of less than 5 sec) was very well captured by our method, as extensively shown in the paper.
(B) The 2-D state-space captures brief state changes such as microarousals. Spectral trajectories corresponding to micro-arousals, i.e. transient disruptions of SWS episodes lasting 3-10 sec (Halasz, 1998).
- supplemental material
-
Figure S3. Spectral variation of WT among animals.
The average LFP power spectrum for each area and the average pooled coherence spectrum were calculated for all epochs identified as WT state in the three animals with more WT (8.7% of total time in rat 1, 4.4 % in rat 4 and 1.3% in rat 5). The power spectra were normalized to the maximum power at the peak LFP oscillation range (7-12 Hz). Error bars represent the standard error of the mean. In each animal, the average LFP power spectra appear very similar among the forebrain areas during WT, with pooled coherence peaking at the same frequency bands. The general spectra of WT were very similar across animals, with the dominant oscillation at 7-12 Hz and resonant frequencies at 14-18 and 20-28 Hz. However, the relative amplitude at the resonant frequencies was substantially different among animals. There is a positive correlation between the amount of WT in each animal and the relative power at the resonant frequencies. In rat 1 (the animal with the most WT), the 14-18 Hz band reached 0.72 of the primary peak and resonant frequency bands can be seen up to 40 Hz. In the other two rats, the relative amplitude at the higher resonant frequencies was less obvious. This spectral difference at the resonant frequencies among animals accounted for the varying location of the WT cluster in the 2-D state-map.
- supplemental material
-
Figure S4. Area-specific state-space maps and the separation of states.
We investigated the distribution of the information about global brain states among various forebrain areas. State-space maps were generated using LFP from each single area in each animal. Principal Component Analysis (PCA) was used to combine amplitude ratios obtained from different LFP (left column). The resulting area-specific state-space maps were plotted for each animal and color coded for behavioral states. Forty-eight hours of data were plotted for each animal (sub-sampling one-third of data for clarity). The results indicate that single-area maps provide as much information as multiple-area pooled maps with regard to state classification. Taken together, these results indicate that indeed there is a high degree of redundancy among the forebrain areas studied, with the possible exception of cortex. In three out of five animals, cortex-specific maps present a certain degree of overlap between clusters.
- supplemental material
-
Figure S5.
(A) Enhanced maps for an automatic classification of global brain states. Enhanced maps for automatic state detection were generated by dividing the point density of 2-D maps (density plot) by the square of the average spectral change speed at each bin (speed plot). The spectral speed refers to the values obtained by dividing the distance between two consecutive dots in the 2-D state-space by the time that separates them, i.e. 1 sec. One hundred linearly-spaced contours covering the whole range of the new map were calculated, and 3 (4 for Rat 1) sets of mutually excluding concentric contours corresponding to the main clusters were identified (Contour map). Within each cluster-specific set of contours, the 95% most-inclusive contour was chosen as the initial boundary of each cluster, resulting in non-overlapping state-specific limits. We then took advantage of the temporal continuity of the spectral trajectories to account for transient within-state fluctuations of power spectrum: all short trajectories (less than 20 sec) leaving and re-entering a given cluster boundary without touching the boundaries of other clusters were assigned to that cluster. Thus, points outside the initial cluster boundaries could also be assigned to the major states, depending on their temporal continuity with adjacent clusters. Data points not coded as any of the major states (15-20%) were labeled as ‘state transitions’.
(B) Quantitative comparison of the state classification provided by the automatic and the visually-coded methods in 3 rats. The parameters compared were (i) accuracy (Acc.), (ii) sensitivity (Sen) and (iii) specificity (Spe), respectively defined as (i) the second-by-second agreement between the two methods, (ii) the probability that visually-coded states were correctly identified by the algorithm, and (iii) the probability that epochs not visually-coded as a given state were correctly not labeled as that state by the algorithm. These three parameters were calculated with or without the transition points (inside and outside parenthesis, respectively). (*) Values not available; rats 2 and 3 presented a total duration of WT of less than 0.1% and 0.3% respectively. The middle column displays for each rat, spectrograms of 2000 sec-long neural data segments over color-coded bars that indicate automatic and visually-coded state classification. Automatic and visually-coded classifications show a very high degree of agreement during global brain states, but differ around state transitions. State boundaries resulting from visually-coded classification often do not match spectral boundaries (inset in spectrogram of Rat 2). The color-coded maps (right panels) show a high topographic correspondence between automatic state-maps and those generated by visual coding. Forty-eight hours of data were plotted (sub-sampling one-third of data for clarity) for the three animals.
The general improvement of automatic state classification after the exclusion of transition points underscores the fact that human-assisted coding imposes a discrete classification scheme on a dataset that is by nature continuous. Accurate automatic classification simply does not provide state identification around state transitions (white breaks in the auto-generated hypnogram), which are better understood by way of trajectory analysis (see Fig. 6). In contrast, human-assisted classification always assigns one of two successive states to each data point. Since these discrete state boundaries often fail to match the underlying sharp spectral boundaries, human-assisted classification likely introduces false positives. Furthermore, agreement between expert observers usually falls between 80-90% (Robert et al., 1999). Thus, the automatic state identification provided by the robust cluster topography of 2-D state-maps constitutes a more conservative, objective and accurate method for state detection than that provided by human visual coding.
- supplemental material
-
Figure S6. Pooled coherence well captures the state-dependent patterns of coupling between forebrain areas. (A) LFP pooled coherence spectrogram showing the variations of coherence across all areas during the wake-sleep cycle, aligned with behavioral state identification (visual coding). (B) Pairwise coherence spectrograms. Two interesting differences between the pooled and the pairwise coherence plots. Notice that theta coherence was high throughout QW and REM episodes in all brain area pairs, but low when calculated across the ensemble.
- supplemental material
-
Figure S7. Transitions between states as trajectories connecting distinct clusters
Major state transitions were identified and plotted in the 2-D state-space using parametric trajectory analyses. The average and mode of the duration are indicated (±s.e.m.) for the animal. The right column shows the mean ± s.e.m. and the mode ± s.e.m. of the duration of transitions for all animals (n=5). Notice all trajectories between SW and REM course through the IS region and present the longest duration.
- supplemental material
-
Figure S8. Pooled coherence spectra during state transitions
The coherence spectra for the major global state transitions are consistent across the five animals recorded in this study. The same transitions shown in Fig. 7 are plotted for all remaining animals.
