Improving Multichannel Raw Electroencephalography-based Diagnosis of Major Depressive Disorder by Pretraining Deep Learning Models with Single Channel Sleep Stage Data *

Figures

• 

  Figure 1.. Model Architecture. Models A through D have 2 sections that are separated by an alpha dropout layer (alpha): (i) feature extraction, which repeats 4 times, and (ii) classification. The grey inset within section (ii) repeats twice. The 4 convolutional (conv1d) layers have 5, 10, 10, and 15 filters, respectively (kernel size = 10, 10, 10, 5). They are followed by max pooling layers (pool size = 2, stride = 2). Section (ii) has 3 dense layers (64, 32, and 2 nodes) with interspersed alpha dropout layers (alpha) with rates of 0.5. The alpha layer between (i) and (ii) also has a rate of 0.5. Yellow circles containing an “E”, “B”, or “S” correspond to ELU activations, batch normalization, and softmax activations. All conv1d and dense layers had max norm kernel constraints with a max value of 1. In Model B the channel dropout layer was placed between “Input” and Section (i).

• 

  Figure 2.. Spectral Importance. In panels from left to right are spectral importance for Models A, C, and D. The y-axes are aligned and show the percent change in BACC, and the x-axes indicate each frequency band. Note that δ, followed by γ activity was important across models, though the pretrained models also prioritized β activity.

• 

  Figure 3.. Spatial Importance. The top, middle, and bottom panels show spatial importance for Models A, C, and D, respectively. Channels are displayed on the x-axis, and the percent change in BACC is shown on the y-axis. Note how F7, Cz, and P4 are important across models. Model D was more like Model A than Model C in F8, Pz, and Fz importance, and Model D was unique in relying upon F4 and O2.