Abstract:
This thesis employs hybrid neural networks to process electroencephalography (EEG)
data to provide an accurate diagnosis of major depressive disorder (MDD). Since EEG Data is
of spatiotemporal nature, a hybrid neural network can learn both sequential and spatial
dependencies from the data to gain a deeper understanding. This thesis also employs domian
adversarial training (DAT) to improve the generalizability of the hybrid neural network using
a domain classifier and gradient reversal layers. This thesis also devises a framework to explore
the deep learning features extracted through the hybrid neural network to translate them, and
gain insights into how MDD manifests itself in EEG data. Finally, the thesis utilizes the deep
learned features extracted for the classification experiment to predict various clinical scales
like PHQ-9, PSQI, GAD-7 and others. The classification experiment yielded a classification
accuracy of 98.30% and a test loss of 0.0965 validated through 10-fold cross validation.
Visualizing the features using t-SNE shows the evolution of the features from the head of the
network having low separation to the tail of the network with separation ramping up as we go
deeper into the layers. The visualization also shows that although DAT increases the
generalizability of the model, this comes at a cost on the classification accuracy. The
Regression scores for PHQ-9, GAD-7 and PSQI indicate that hybrid neural networks can learn
features informative enough to predict clinical scores data which was not a part of the training
process. Furthermore, the experiment shows the earliest layers of the model do not produce
features that have a strong positive or negative correlation with our non-linear hand-crafted
features. The correlation then begins to increase as we go deeper and then starts to decrease, as
the increase in dimensionality causes the features to no longer be correlated with the non-linear
hand-crafted features.
