SpectraVis

Figure 4.3. A screenshot of the SpectraVis interface.

There are three main parts of the user interface: the network view, the individual sensor pair view, and the controls. The controls alter the network view. Using the controls, the user can change the subject or metric used for the network, display the network at other time-frequency bins, or change the spatial layout of the network, or “play” the network forward in time — a movie of how the network changes. A dynamic legend which updates based on the data and the edge statistic used is below the controls. Users can click on nodes or edges of the network in the network view to get a more detailed view in the individual sensor pair view, displaying all the time-frequency bins of the spectra and coherences (or other associative measure such as

correlation) corresponding to the sensor pair selected. Mousing over a time-frequency bin in the individual sensor pair view will display the network at that time-frequency bin in the network

view. The individual sensor pair view also shows the change over time of the edge statistic and spectra at the current frequency of the network view (bottom).

Functional network analysis is a growing area of neuroscience research, driven in part by technological improvements allowing us to record from more sensors simultaneously. However, as researchers record from more sensors, network analyses can become unwieldy and hard to interpret, because the number of possible network connections scales quadratically with the number of sensors (e.g. electrodes). Further, we expect neural processes to form dynamic networks that vary over time, frequency, and spatial scales (e.g. within and between brain regions), adding numerous dimensions to network analyses.

SpectraVis is an interactive visualization aimed at enhancing exploratory analysis of networks by allowing the user to efficiently: (1) compare task-related functional networks over time and frequency, (2) compare individual and

associative measures on all sensor pairs (e.g. spectra, coherences), and (3) compare different measures of association (e.g. correlation vs. coherence, binary vs. weighted networks). The different views of SpectraVis are dynamically linked, highlighting relationships between the metrics in response to user interaction.

Figure 4.3 shows a typical view of SpectraVis. The network view shows the anatomical location of the sensors (circles with sensor number) and edges (lines) weighted by the edge statistic (color of the line, measure of association between the sensors). In this example, the edges are binary, representing

significant changes in local field potential coherence between Speech — subjects reading aloud the words of a famous speech or nursery rhyme — and Silence at a particular frequency (10 Hz) and time (187.5 ms after speech onset). The network

has dense connectivity within and between primary motor and primary somatosensory cortices (M1 and S1). Users can compare between binary edge statistics (see Figure 4.4, middle and right image), which categorically declare associations between sensors, and weighted edge statistics (see Figure 4.4, left image), which use continuous measures such as the raw coherence difference and z-scored coherence difference, specified via the edge statistic dropdown.

The controls can also be used to examine the evolution of the network over time using either the time slider, which can be dragged to a time of interest, or the play button, which will automatically advance the time slider. Examining the network over time can potentially reveal differences in network structure that could result from an experimental cue or event. SpectraVis enables quick comparison between all time points. The user can also compare networks at different frequency bands (for example, comparing a 10 Hz alpha band network to a 20 Hz beta band network) using the frequency slider. This is important because different frequency bands may have different functional roles (Ainsworth et al., 2012; Engel and Fries, 2010; Kopell et al., 2010; Palva and Palva, 2007).

One difficulty of analyzing networks is interpreting the edges between sensors, particularly if the network is a weighted network and there are many sensors. There often is not enough room in a visualization to display all the edges without much overlap. To solve this problem, we use two strategies: layout and filtering.

The network layout toggle controls where nodes are positioned (and consequently the edges between nodes). The anatomical layout places the nodes according to the anatomical position defined in the JSON files (Figure 4.4, left and middle image). For example, in an ECoG dataset the network nodes often correspond to electrodes, which can be displayed over an image of the brain to give the end-user a sense for the sulci and gyri underlying each electrode. In some cases the anatomical locations of the nodes may be less important than their position induced by the network topology: for example, a node with strong connections to other nodes may serve as a hub. To help with this kind of

interpretation, SpectraVis offers a topological layout that models the nodes and edges as a physical system to limit the number of overlapping edges (Figure 4.4, right image). Node numerical labels are preserved across choices of layout.

Figure 4.4 Different layouts for understanding associative networks.

Left layout draws edges between all nodes. The color of the edges corresponds to a relative weight

— in this case a difference of coherence between speech and no-speech conditions in the experiment. Middle layout only shows the edges that correspond to statistically significant changes in coherence (binary network). This network is easier to understand (compare to the left weighted network) because there are fewer edges, although it is subject to the choice of statistical thresholding — there might be meaningful changes that are not shown due to the choice of threshold or vice versa. A topological layout (right layout) removes the constraint of spatial anatomical location and attempts to position the nodes so that connected nodes are closer together and overlapping of edges is minimal, so the user can better see which edges and nodes are well-connected (hubs).

The second strategy to make edges more visible allows the user to isolate subsets of the network by filtering edges and sensors, thus reducing the number of comparisons needed in any one view. Using the edge filter dropdown, the user can isolate the edges between sensors that reside within the same brain area (e.g. only auditory cortex - auditory cortex sensor pairs) or between sensors that have non-matching brain areas (e.g. only auditory cortex – motor cortex sensor pairs).

Once the desired network view has been obtained, users can get further detail by clicking on a pair of sensors. This loads a sensor view (dotted box) which depicts the relationship (spectra and coherences) between a selected pair of sensors (circled in black, network view, sensors 85 and 90) at all times and frequencies. Here, the edge between M1 (sensor 90) and S1 (sensor 85)

represents a 10 Hz increase in speech coherence relative to silence. The increase co-occurs with higher frequency beta (15-25 Hz) power suppression on the M1 sensor. The user can investigate the relationship between the sensor pair and the network view by mousing over a time-frequency bin in the sensor view, which correspondingly updates the network view to the time-frequency bin under the cursor.