EEG Based Generative Depression Discriminator

TL;DR

This paper tackles objective depression diagnosis from EEG by introducing a Generative Detection Network (GDN) that learns brain activity and regenerates target electrode signals. It employs a two-generator, dual-branch encoding–decoding framework that processes a target electrode’s surroundings, filters signals to the $4$–$14$ Hz band, and decomposes them with a $db6$ wavelet into $cA$ and $cD$ features, which are then mapped to a hidden representation and reconstructed for classification. By comparing the depressed- and control-trained generators’ reconstructions against the original EEG, the method achieves high segment-level accuracy (92.30% on MODMA, 86.73% on HUSM) and provides explainable spatial information via EEG heatmaps, supporting reliable clinical interpretation. The approach demonstrates the potential of physiologically informed generative models to improve objective depression detection and reduce misjudgments, with publicly available datasets and plans to release code.

Abstract

Depression is a very common but serious mood disorder.In this paper, We built a generative detection network(GDN) in accordance with three physiological laws. Our aim is that we expect the neural network to learn the relevant brain activity based on the EEG signal and, at the same time, to regenerate the target electrode signal based on the brain activity. We trained two generators, the first one learns the characteristics of depressed brain activity, and the second one learns the characteristics of control group's brain activity. In the test, a segment of EEG signal was put into the two generators separately, if the relationship between the EEG signal and brain activity conforms to the characteristics of a certain category, then the signal generated by the generator of the corresponding category is more consistent with the original signal. Thus it is possible to determine the category corresponding to a certain segment of EEG signal. We obtained an accuracy of 92.30\% on the MODMA dataset and 86.73\% on the HUSM dataset. Moreover, this model is able to output explainable information, which can be used to help the user to discover possible misjudgments of the network.Our code will be released.

Figures (5)

• Figure 1: Generative Detection Network(GDN)
• Figure 2: These are comparisons of the EEG generated by the generators, where the blue line in each graph is the original EEG signal, the red line is the signal generated by the depression generator, and the yellow line is the signal generated by the control generator. The four graphs in the first row are the EEG signals from the control, and it can be noticed that the output is not that good. In contrast, the four graphs in the second row are EEG signals from depressed patients, and it can be found that the depression generator fits the EEG data of depressed patients better and far outperforms the results generated by the control generator. From this we can determine that the depression generator learns the hidden relation between the depression and its brain activity, but the control generator does not work as well as expected.
• Figure 3: (Fig a) is the distribution on MODMA val dataset, and (Fig b) is on test dataset. By observing the distribution in the validation set, we chose the classified line $n_{0}=81$ and obtained the accuracy on the test set based on the criteria in the validation set.
• Figure 4: This is a display of the final output EEG thermograms, the two pictures in the first row are the EEG topographies output when the depressive prediction is correct. The second row shows the EEG topography when the control is correctly predicted. The first picture in the third row is the EEG topography output when a depressed is misclassified as a normal, and the second is topography when a normal control is misclassified as a depression.
• Figure :