Leveraging pretrained vision transformers for classifying alcohol use disorder using raw resting-state EEG

Area of Science:

• Neuroscience and Computational Psychiatry
• Artificial Intelligence in Healthcare

Background:

• Alcohol Use Disorder (AUD) affects millions, lacking objective diagnostic biomarkers.
• Resting-state electroencephalogram (EEG) offers a rich source of neural data.
• Deep learning, particularly transformer architectures, shows promise in analyzing complex biological signals.

Purpose of the Study:

• To investigate the efficacy of a deep learning model (EEGViT) for classifying AUD using raw EEG.
• To evaluate the model's performance on independent cohorts with Cannabis Use Disorder (CUD) and Opioid Use Disorder (OUD).
• To explore the potential of transformer-based models for psychiatric diagnosis from EEG data.

Main Methods:

• Utilized EEGViT, a hybrid deep learning architecture combining convolutional patch embedding and a Vision Transformer (ViT).
• Applied the model to 5402 raw resting-state EEG recordings from the Collaborative Study on the Genetics of Alcoholism (COGA).
• Matched and stratified analysis groups by age and sex; evaluated independently on CUD and OUD datasets.

Main Results:

• The AUD model achieved approximately 57% classification accuracy overall (55% males, 59% females).
• Independent evaluations showed ~64% accuracy for CUD and ~63% for OUD.
• Temporal analysis revealed improved accuracy in later EEG recording segments.

Conclusions:

• Transformer-based deep learning models can be applied to psychiatric classification using raw EEG data.
• Findings provide preliminary evidence for EEG-based biomarkers in AUD, CUD, and OUD.
• This study establishes a foundation for future research into advanced EEG analysis for psychiatric disorders.