Fig 1.
Resting-state connectivity patterns of resting-state networks across the whole group of subjects, overlaid in MNI space.
Network seed coordinates and descriptions are provided in Table 1. Images are displayed in radiologic convention.
Table 1.
Montreal Neurological Institute (MNI) coordinates and abbreviations of resting-state networks.
Fig 2.
Schematic of proposed framework for studying temporal and spectral characteristics of dynamic functional connectivity.
(a) Dynamic functional connectivity (dFC) is first estimated from a model-based state-space approach. (b) dFC estimates are transformed to the spectral domain using Welch’s power spectral density estimate. (c) Temporal and spectral features are extracted from the temporal and spectral domains. (d) Temporal and spectral features are concatenated and (e) used as the feature vector for subject-level prediction.
Table 2.
Temporal and spectral features of dynamic functional connectivity.
Features and abbreviations are shown. dFC, dynamic functional connectivity.
Fig 3.
Pearson correlation heatmap of dFC features for DMN-memory network connectivity.
Red indicates positive correlations; blue indicates negative correlations. Abbreviations: MV, Mean Value; VAR, Variance; ZC, Zero Crossing Rate; PSE, Power Spectral Entropy; PAV, Proportion of Anticorrelated Volumes; SRO, Spectral Rolloff; SCO, Spectral Centroid; SMED, Median Frequency; ALFF-dFC, Amplitude of Low Frequency Fluctuations in dFC; SPR, Spectral Spread; SKW, Spectral Skewness; KURT, Spectral Kurtosis; FLAT, Spectral Flatness; PEAK, Spectral Peak; FLUX, Spectral Flux; CREST, Spectral Crest; SLOPE, Spectral Slope.
Fig 4.
Estimated dynamic functional connectivity, power spectra, and dFC features for connectivity between the DMN and memory network for two sample patients with temporal lobe epilepsy.
(a) original BOLD time-series, (b) estimated dynamic functional connectivity and dFC power spectra, and (c) estimated dFC features. Abbreviations: MV, Mean Value; VAR, Variance; ZC, Zero Crossing Rate; PSE, Power Spectral Entropy; PAV, Proportion of Anticorrelated Volumes; SRO, Spectral Rolloff; SCO, Spectral Centroid; SMED, Median Frequency; ALFF-dFC, Amplitude of Low Frequency Fluctuations in dFC; SPR, Spectral Spread; SKW, Spectral Skewness; KURT, Spectral Kurtosis; FLAT, Spectral Flatness; PEAK, Spectral Peak; FLUX, Spectral Flux; CREST, Spectral Crest; SLOPE, Spectral Slope.
Fig 5.
95% bias-corrected and accelerated (BCa) bootstrap confidence intervals for the difference in means (HC-TLE) for each dFC feature.
Positive values indicate larger values in HC than in TLE. Abbreviations: HC, healthy control; TLE, temporal lobe epilepsy; MV, Mean Value; VAR, Variance; ZC, Zero Crossing Rate; PSE, Power Spectral Entropy; PAV, Proportion of Anticorrelated Volumes; SRO, Spectral Rolloff; SCO, Spectral Centroid; SMED, Median Frequency; ALFF-dFC, Amplitude of Low Frequency Fluctuations in dFC; SPR, Spectral Spread; SKW, Spectral Skewness; KURT, Spectral Kurtosis; FLAT, Spectral Flatness; PEAK, Spectral Peak; FLUX, Spectral Flux; CREST, Spectral Crest; SLOPE, Spectral Slope.
Fig 6.
Predictive performance of dFC features.
Classification accuracy (TLE vs. controls) based on connectivity between DMN and various resting-state networks is shown. The horizontal line indicates the performance of a naïve classifer. The naïve classifer provides a baseline level of performance and classifies all test samples as the most common class in the training set. Static functional connectivity (sFC; cyan), dFC mean (SW-Mean; purple), dFC variance (SW-VAR; gray), dFC mean and variance (SW-MeanVar; green), static functional connectivity with random forests (sFC-RF), and proposed dFC approach (GARCH-RF; red). Mean and 95% CI over thirty replicates are shown.
Fig 7.
Conditional variable importance scores of dFC features.
Mean of Random Forest conditional variable importance scores over 30 replicates, for dFC features between DMN and various resting-state networks. Abbreviations: MV, Mean Value; VAR, Variance; ZC, Zero Crossing Rate; PSE, Power Spectral Entropy; PAV, Proportion of Anticorrelated Volumes; SRO, Spectral Rolloff; SCO, Spectral Centroid; SMED, Median Frequency; ALFF-dFC, Amplitude of Low Frequency Fluctuations in dFC; SPR, Spectral Spread; SKW, Spectral Skewness; KURT, Spectral Kurtosis; FLAT, Spectral Flatness; PEAK, Spectral Peak; FLUX, Spectral Flux; CREST, Spectral Crest; SLOPE, Spectral Slope.
Table 3.
Ranked top five dFC features for discriminating TLE patients from healthy controls for various resting-state networks.
(a) DMN/auditory network,(b) DMN/language network, (c)DMN/memory network, (d) DMN/motor network, (e) DMN/visual network. ALFF-dFC, amplitude of low-frequency fluctuations in dFC; CREST, spectral crest; KURT, spectral kurtosis; MV, mean value; PAV, proportion of anticorrelated volumes; PEAK, dominant frequency; PSE, power spectral entropy; SCO, spectral centroid; SKW, spectral skewness; SLOPE, spectral slope; SPR, spectral spread; VAR, variance.