Individual Human Brain Areas Can Be Identified from Their Characteristic Spectral Activation Fingerprints

Abstract

The human brain can be parcellated into diverse anatomical areas. We investigated whether rhythmic brain activity in these areas is characteristic and can be used for automatic classification. To this end, resting-state MEG data of 22 healthy adults was analysed. Power spectra of 1-s long data segments for atlas-defined brain areas were clustered into spectral profiles ("fingerprints"), using k-means and Gaussian mixture (GM) modelling. We demonstrate that individual areas can be identified from these spectral profiles with high accuracy. Our results suggest that each brain area engages in different spectral modes that are characteristic for individual areas. Clustering of brain areas according to similarity of spectral profiles reveals well-known brain networks. Furthermore, we demonstrate task-specific modulations of auditory spectral profiles during auditory processing. These findings have important implications for the classification of regional spectral activity and allow for novel approaches in neuroimaging and neurostimulation in health and disease.

Fig 1. Analysis pipeline.

After general preprocessing of the MEG data, the ~7-min continuous data were first segmented into 1-s segments. Artefactual segments and channels were rejected, and complex Fourier spectra were computed for each segment. Single trials (refers to 1-s segments) were then projected into source space by using previously computed linear constraint minimum variance (LCMV) weights. Data were spectrally normalised by dividing the spectrum of each segment and voxel by the average power spectrum across all segments and voxels per participant (ratio normalisation). Voxels were grouped according to the Automated Anatomical Labeling (AAL) atlas and activity averaged for each area. k-means and Gaussian mixture (GM) modelling algorithms were applied to reduce the dimensionality of trials for each participant (1st-level models, subject level) into ten distinct clusters. These ten clusters were clustered across all 22 participants, again using k-means and GM modelling algorithms (2nd-level models, group level). For 2nd-level models, the optimal number of clusters per area was computed, using a Silhouette criterion. Second-level GM models are referred to as spectral profiles or fingerprints.

Fig 2. Spectral profiles of 16 example areas seen a) laterally and b) from the midsagittal plane.

Clustered power spectra in source space represent normalised power, i.e., spectral power in comparison to the whole brain. Legends show the corresponding duration of each pattern (i.e., the percentage of trials in which each spectrum was present on average during recording). Shaded error bars illustrate the standard error of the mean across participants. The lines are colour-coded for the respective frequency bands (red: delta, green: theta, blue: alpha, orange: beta, grey: gamma). Inlets show average power spectra for respective areas without normalisation (dotted lines) and with ratio normalisation (continuous lines). Note that frequency on the x-axis is scaled logarithmically, and that data at 50 Hz (line noise) is interpolated in the plots.

Fig 3. Classification procedure and results.

a) Illustration of classification procedure. Second-level GM training models were created for all brain areas with data from half of the sample (group-level, right panel). Area-specific first-level GM test data from the remaining sample (subject-level, left panel) were then tested against each area-specific GM training model. The fit of each test set to each GM training model (determined through negative log-likelihood) was ranked. This was repeated 120 times, and the position of the correct area was averaged across iterations. For the illustrated example, the mean rank equalled 1.6 across 120 iterations. b) Distribution of mean ranks in classification. (left) Histogram of mean ranks across all 115 brain areas. Bin width is one. (middle) Topography of mean ranks (colour-coded from 1, blue, to 4, yellow; bin width is one). (right) Linear regression revealed a dependency of mean rank per area and the radius r from the centre of the brain. The mean rank increases the further an area is away from the centre of the brain. In other words, areas closer to the centre of the brain are easier to classify than more superficial areas. Data underlying this plot can be found in S1 Data.

Fig 4. Area networks according to similarity analysis.

(left) Topography of similar areas. (right) Slice view of the brain, including the cerebellum, with colour-coded similarity. A hierarchical clustering analysis (using negative log-likelihood between areas as distance measure) revealed several groups of areas that resemble previously identified brain networks. The frontal network (light blue) consists of bilateral medial and dorsolateral superior frontal gyrus, middle frontal gyrus, triangular inferior frontal gyrus, and right opercular inferior frontal gyrus. The sensorimotor network (orange) consists of bilateral pre- and postcentral gyri, paracentral lobule, and supplementary motor area. Two visual networks were found, a medial visual network (green) consisting of bilateral calcarine, cuneus, and precuneus and a parieto-occipital visual network (red) consisting of bilateral superior, middle, and inferior occipital gyri, superior parietal gyri, and supramarginal and angular gyri.

Fig 5. Distribution of number of clusters.

(left) Histogram of cluster numbers across all 115 brain areas. Bin width is one. (middle) Topography of cluster numbers (colour-coded from 1, blue, to 9, yellow; bin width is one). (right) Linear regression reveals a dependency of number of clusters per area and the radius r from the centre of the brain. The number of clusters decreases with increasing radius. Data underlying this plot can be found in S2 Data.

Fig 6. Comparison of spectral profiles in (a) left and (b) right primary auditory cortices during rest and listening.

Upper panels for left and right PAC show spectral profiles including all modes for the rest (left) and listening (right) conditions. The x-axis is limited to 30 Hz. The amplitudes of spectra with comparable peak frequencies in both conditions were subjected to an independent sample t test across participants (lower panels). Bars illustrate amplitude means across individual spectra that contributed to group clusters. Error bars illustrate standard error of the mean across participants. Asterisks denote significance between rest and listening: * p < .05, ** p < .01, *** p < .001, n.s. not significant. Data underlying this plot can be found in S3 Data.