Towards the Development of a Deep Learning Framework Using Adaptive and Non-Adaptive Time-Frequency Features for...

Area of Science:

• Neuroscience
• Computational Psychiatry
• Medical Imaging

Background:

• Predicting individual response to depression therapies like SSRIs and rTMS is challenging, with current success rates around 50%.
• Treatment selection often relies on trial-and-error due to the lack of predictive biomarkers.
• Developing a method to predict therapy outcomes pre-treatment is crucial for personalized medicine.

Purpose of the Study:

• To present a computer-aided decision (CAD) framework for predicting depression therapy outcomes.
• To utilize pre-treatment electroencephalogram (EEG) signals with advanced time-frequency representations and convolutional neural networks (CNNs).
• To determine the optimal time-frequency representation and CNN architecture for therapy-specific outcome prediction.

Main Methods:

• EEG signals from 30 SSRI and 46 rTMS patients were analyzed.
• Continuous Wavelet Transform (CWT) and Variational Mode Decomposition (VMD) were used for time-frequency image generation.
• Four pretrained CNNs (ResNet-18, MobileNet-V3, EfficientNet-B0, TinyViT-Hybrid) were fine-tuned and evaluated using 6-fold cross-validation.

Main Results:

• CWT representations excelled in SSRI outcome prediction (ResNet-18: 99.43% accuracy), while VMD representations were superior for rTMS (ResNet-18: 98.77%).
• ResNet-18 and TinyViT-Hybrid architectures outperformed others.
• Subject-independent evaluation achieved 82.50% (SSRI) and 83.53% (rTMS) accuracy, with therapy-specific channel dominance observed.

Conclusions:

• The choice of time-frequency representation is critical and therapy-specific for predicting depression treatment outcomes.
• Effective prediction can be achieved using well-designed spectral images and simpler CNN architectures, without complex layers.
• This CAD framework offers a promising approach for personalized depression therapy selection.