Evaluating Augmentation Approaches for Deep Learning-based Major Depressive Disorder Diagnosis with Raw...

Area of Science:

• Neuroscience
• Machine Learning
• Computational Psychiatry

Background:

• Deep learning models are promising for neuropsychiatric disorder diagnosis but are limited by small datasets.
• Electroencephalography (EEG) data augmentation (DA) methods are underexplored for clinical translation in this field.

Purpose of the Study:

• To evaluate the effectiveness of six EEG data augmentation techniques for major depressive disorder diagnosis.
• To introduce a new baseline for comparison, using duplicated training data to control for dataset size bias.
• To analyze the impact of EEG DA on model-learned representations and robustness.

Main Methods:

• A deep learning model was trained for major depressive disorder diagnosis using EEG data.
• Six different EEG data augmentation techniques were applied and evaluated.
• A baseline model trained on duplicated data was used for performance comparison.
• Explainability analyses were conducted to assess the impact on learned representations.

Main Results:

• Most EEG data augmentation methods did not outperform a baseline model trained on duplicated data.
• Channel dropout augmentation significantly improved model performance.
• Certain DA methods enhanced model robustness against frequency and channel perturbations.

Conclusions:

• A robust baseline using duplicated data is crucial for evaluating EEG data augmentation methods.
• Channel dropout is a promising EEG DA technique for major depressive disorder diagnosis.
• Further research into EEG DA methods is needed for effective clinical translation of deep learning models.