3.3.1. Visualizations and Mapping
The SNN-based methodology (explained in the methods section above), is constituted of the following steps in this study:
Mapping, modelling, classifying and understanding of EEG data.
Statistical and quantitative analysis on the SNN models to assess the model significance.
Firstly, a brain-inspired 3D SNN model was designed based on the Talairach brain atlas of 1471 neurons (in spiking neural networks, an artificial neuron refers to a computational unit that mimics the behavior of a biological neuron which receives the information, processes it and produces an output). Here, the term neuron is used to represent the center co-ordinate of one cubic centimeter area from the 3D Talairach Atlas. The SNN model input neurons are allocated to the 64 EEG channels to transfer their spike trains into the SNN model.
Then, eight separate 3D models were trained with different EEG data sets related to AM (Figure 3
) and constant (Figure 4
) auditory stimulation across responder group and non-responder groups at baseline (Time 1 [T1]) and post-auditory stimulation (Time 2 [T2]).
As shown in Figure 3
and Figure 4
, at baseline (T1) a difference of connection weights between responders and non-responders was evident. In the AM condition, a slight reduction in overall connectivity was observed between T1 (Figure 3
a, top) and T2 (Figure 3
b, top) for non-responders, and a slight increase in overall connectivity was observed between T1 (Figure 3
a, bottom) and T2 (Figure 3
b, bottom) for responders.
In the constant condition, a reduction in overall connectivity between T1 (Figure 4
a, top) and T2 (Figure 4
b, top) was observed for non-responders, and a larger reduction in overall connectivity between T1 (Figure 4
a, bottom) and T2 (Figure 4
b, bottom) was observed for responders.
To better scrutinize the differences between the SNN models of different stimulations, the connection weights (
) of the difference between two correspondingly trained SNN models (T1, T2) were calculated for each group and subtracted (
). The subtracted connectivity model is depicted in Figure 3
c and Figure 4
c, which shows the involved model brain areas activated in response to both constant and AM auditory stimulations. Figure 5
shows the distributions of weight values across 64 EEG channels for both responder and non-responder groups before (T1) and after (T2) the AM and constant auditory stimulation.
3.3.2. Statistical Analysis of the SNN Models
The numerical information of the trained SNN models can be analyzed to evaluate the models’ statistical significance. For every trained SNN model, an activation level was measured through computing the average value of its connection weights across EEG channels.
Therefore, for every participant, one SNN model was developed at T1 (baseline) and T2 (after auditory stimulation). The average connection weights for each individual SNN model were calculated as a function of group (responder and non-responder) and time (T1, T2).
The connection weights for each EEG channel were then grouped into five sites for each hemisphere with their topographical features (Figure 6
): frontal, temporal, frontocentral, centroparietal and occipitoparietal. It is important to note that these “sites” refer to sections of the modelled data and may not reflect activity of their corresponding anatomical brain regions, since voltages measured on the scalp can have various sources throughout the brain.
A repeated measures ANOVA was performed with Hemisphere (left, right), Site (frontal, temporal, frontocentral, centroparietal, occipitoparietal), Time (T1, T2), and stimulation condition (AM, constant) as within group variables, and Group (responder, non-responder) as a between groups variable. Violations of the assumption of sphericity were corrected using Greenhouse-Geisser corrections. There were significant main effects of Hemisphere [F (1,7) = 153.72, p > 0.001, ηp2 = 0.96] and Site [F (2.24, 15.65) = 224.92, p < 0.001, ηp2 = 0.97] and a significant Hemisphere*Site interaction [F (4,28) = 6.47, p < 0.001, ηp2 = 0.48]. The right hemisphere showed greater mean weights than the left overall and at each site (Occipitoparietal > Centroparietal > Temporal > Frontal > Frontocentral).
Then, separate ANOVAs were conducted for AM and constant stimuli to investigate the effects of auditory stimulation across individuals. This revealed a similar pattern of effects. For AM stimuli, there were significant effects of Hemisphere [F (1,7) = 53.24 p < 0.001, ηp2 = 0.88] and Site [F (4,28) = 123.62, p < 0.001, ηp2 = 0.95], and a significant Hemisphere*Site interaction [F (4,28) = 5.18, p = 0.003, ηp2 = 0.425]. For the constant stimuli, there were significant effects of Hemisphere [F (1,7) = 99.51, p < 0.001, ηp2 = 0.93] and Site [F (1.99,13.96) = 136.42, p < 0.001, ηp2 = 0.95].
Separate ANOVAs were then run on data split further by group. For the responder group in the AM condition, there were significant effects of Hemisphere [F (1,5) = 53.29, p < 0.001, ηp2 = 0.91] and Site [F (4,20) = 91.78, p < 0.001, ηp2 = 0.95] and a significant Hemisphere*Site interaction [F (4,20) = 2.93, p = 0.046, ηp2 = 2.37]. For the non-responder group in the AM condition, there was a significant effect of Site [F (4,8) = 54.73, p <0.001, ηp2 = 0.97] and a significant Hemisphere*Site interaction [F (4,8) = 4.83, p = 0.028, ηp2 = 0.71]. For the responder group in the constant condition, there were significant effects of Hemisphere [F (1,5) = 112.18, p < 0.001, ηp2 = 0.96] and Site [F (4,20) = 102.15, p < 0.001, ηp2 = 0.95]. For the non-responder group in the AM condition, there were significant effects of Hemisphere [F (1,2) = 20.91, p = 0.045, ηp2 = 0.91] and Site [F (4,8) = 76.19, p < 0.001, ηp2 = 0.97].
No effects of group, time or condition reached significance.
Mean responder (n
= 6) and non-responder (n
= 3) group changes were calculated by subtracting the baseline measure from the post-stimulation measure for both the AM and constant stimuli (Figure 7
). For the AM stimulus, responder and non-responder groups showed small changes in opposite directions. For the constant stimulus, changes were larger than for the AM stimulus and both groups showed a decrease, with a larger effect for responders.