Enhancing Electroencephalogram-Based Prediction of Posttraumatic Stress Disorder Treatment Response Using Data...

Area of Science:

• Neuroscience
• Machine Learning
• Computational Psychiatry

Background:

• Predicting treatment response in posttraumatic stress disorder (PTSD) is challenging due to limited patient data.
• Electroencephalogram (EEG) data offers potential biomarkers for neuropsychiatric conditions but often suffers from small sample sizes.
• Deep learning models require substantial data for effective training and generalization.

Purpose of the Study:

• To enhance the prediction of treatment response in PTSD patients using electroencephalogram (EEG) data.
• To apply a variational autoencoder (VAE)-based data augmentation (DA) technique to increase the volume and diversity of EEG data.
• To evaluate the impact of VAE-based DA on the performance of deep neural network (DNN) classifiers for treatment response prediction.

Main Methods:

• Collected EEG spectrograms from patients diagnosed with PTSD.
• Pretrained a VAE model on original EEG spectrograms to generate synthetic augmented data.
• Trained a DNN classifier using both original and VAE-augmented EEG data, comparing performance against models trained solely on original data using Area Under the Receiver Operating Characteristic Curve (AUC).

Main Results:

• The DNN model trained with VAE-augmented EEG data achieved an AUC of 0.85.
• This represents a significant improvement of 0.11 in AUC compared to the model trained without data augmentation.
• The VAE-based DA approach demonstrated enhanced classification performance and improved model generalization.

Conclusions:

• Variational autoencoder-based data augmentation is an effective strategy for overcoming data limitations in clinical EEG studies for PTSD.
• This method significantly improves the performance of deep neural network models in predicting treatment response.
• The VAE-DA approach offers a promising avenue for future EEG-based research in neuropsychiatry, particularly with small datasets.