Connectivity dynamics from wakefulness to sleep

Abstract

Interest in time-resolved connectivity in fMRI has grown rapidly in recent years. The most widely used technique for studying connectivity changes over time utilizes a sliding windows approach. There has been some debate about the utility of shorter versus longer windows, the use of fixed versus adaptive windows, as well as whether observed resting state dynamics during wakefulness may be predominantly due to changes in sleep state and subject head motion. In this work we use an independent component analysis (ICA)-based pipeline applied to concurrent EEG/fMRI data collected during wakefulness and various sleep stages and show: 1) connectivity states obtained from clustering sliding windowed correlations of resting state functional network time courses well classify the sleep states obtained from EEG data, 2) using shorter sliding windows instead of longer non-overlapping windows improves the ability to capture transition dynamics even at windows as short as 30 ​s, 3) motion appears to be mostly associated with one of the states rather than spread across all of them 4) a fixed tapered sliding window approach outperforms an adaptive dynamic conditional correlation approach, and 5) consistent with prior EEG/fMRI work, we identify evidence of multiple states within the wakeful condition which are able to be classified with high accuracy. Classification of wakeful only states suggest the presence of time-varying changes in connectivity in fMRI data beyond sleep state or motion. Results also inform about advantageous technical choices, and the identification of different clusters within wakefulness that are separable suggest further studies in this direction.

Figure 1:

Schematic depicting resting fMRI data processing (adapted from (Damaraju et al., 2014)).

Figure 2:

Sixty-two selected ICNs for further analysis were grouped into 7 modules using previously reported methods (Allen et al., 2012b).

Figure 3:

Cluster centroids from k-means clustering of dFNC window data for window size 30 (A) and the frequency of occurrence of each state in time (B). The standard errors for the frequency of occurrence were computed using 100 bootstrap resamples of the subject dFNC window data. (C) The modularity of the centroids is computed using the Louvain algorithm (“modularity_louvain_und_sign.m” function in the Brain Connectivity toolbox) resulting in three modules (Mod) for states 1,2 and 3 and four modules for states 4 and 5. The top 12 ICs with highest mean within module FC are depicted scaled by mean within module FC. Note that the weights (mean within module connectivity) are lower in states 1 and 5 compared to states 2, 3 and 4.

Figure 4:

Frequency counts of state vector assignments obtained from k-means clustering of dFNC data sorted by hypnogram states. States 1 and 2 primarily occur during wakefulness, dFNC states 4 and 5 during N2 sleep stage and state 5 is predominant during deep sleep (N3 stage).

Figure 5:

Comparison of subject state vectors obtained from k-means clustering of dFNC windows and EEG derived hypnograms for the two subjects with the highest correlation (A and B) and the two subjects with the lowest correlation (C and D). The overall distribution of the correlation between the two state vectors for all 55 subjects is presented in E. As seen, the subjects with low correlation tend to be awake throughout the scan session and the corresponding dFNC state vector transitions within the states 1 and 2 that are prevalent during wakefulness.

Figure 6:

2D Visualization of dFNC data: We selected 2000 random dFNC windows (400 per dFNC state) and projected the multidimensional (1891) data into 2 dimensions using the t-SNE algorithm. The resulting mapping was subsequently color coded by k-means clustering assignment into 5 states (A) and by the subject EEG hypnogram state of that point (B). The data dimension was reduced to 30 principal components and a perplexity value of 35 was used for this projection.

Figure 7:

Comparison of state vector statistics and transition matrices computed from EEG-derived hypnograms and dFNC cluster derived subject state vectors. Hypnogram and dFNC state vector exhibit similar frequency of occurrences (A and C) and mean dwell times (C and D) respectively. The mean state transition matrices for hypnogram vectors (E) and dFNC state vectors (F) inform about the probability of transitioning from a given state i at time t-1 to state j at time t. The probabilities are converted to -log(10) scale, so higher (yellow to white) intensity values mean lower probability to transition. For both modalities, these matrices demonstrate tendency to remain in a given state (diagonal values are lower). The transitions to neighboring states are more likely in both the hypnogram and the dFNC state vectors. While there is chance of transitioning from deep sleep N3 to any other state (W, N1 or N2), the probability of transitioning from wakefulness at time t-1 immediately to deeper sleep stages (N2, N3) at time t is very low suggesting gradual transition from W to N3 stage. Note that dFNC states 1 and 2 are combined for this analysis.

Figure 8:

Classification accuracies using linear SVM for training (A) and test (B) cases of alignment between EEG hypnogram and subject dFNC state vector obtained using a window size of 30 TRs. The alignment is tested for lags −15 to 15 TRs in step of 3 TRs. Each point in the distribution corresponds to the balanced per class accuracy from one of the 11 cross-validation iterations of training data that included data from 50 random subjects and the accuracy from left of test data that included data from 5 remaining subjects. The results using a SVM with RBF kernel for the same data are shown in C and D. Results are consistent for both linear and RBF SVM kernels .

Figure 9:

Linear SVM classification accuracies of subject sleep stage from dFNC estimates obtained using different window sizes for training (A), test (B) data from 11 cross-validation iterations and the classification accuracies obtained from the DCC estimates for the same cross-validation scheme for training and left out test data are presented in the top right (C). The classification accuracies obtained with RBF kernel for the train (D), test(E) dFNC estimates and for DCC estimates (F) are shown.

Figure 10:

Cluster centroids from k-means clustering of window dFNC data of awake only state (State 1). The estimated clusters were observed to have meaningful and distinct structure.