Leveraging Pretrained Vision Transformers for classifying Alcohol Use Disorder using Raw Resting-State EEG.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Psychiatry

Background:

• Alcohol Use Disorder (AUD) is a widespread neuropsychiatric condition impacting millions, yet lacks objective diagnostic biomarkers.
• Current diagnostic methods for AUD are limited, highlighting the need for novel neurophysiological tools.

Purpose of the Study:

• To investigate the efficacy of deep learning, specifically the EEGViT model, in classifying individuals with AUD using raw resting-state electroencephalogram (EEG) data.
• To explore the potential of transformer-based models for psychiatric classification and the development of EEG-based diagnostic tools.

Main Methods:

• Utilized a large dataset from the Collaborative Study on the Genetics of Alcoholism (COGA), including 5,402 EEG recordings from 2,710 participants.
• Applied demographic matching and undersampling to manage confounding factors and class imbalance, preserving raw EEG features.
• Employed EEGViT, a hybrid deep learning architecture, for end-to-end classification of AUD, CUD, and OUD, with analyses stratified by sex and age.

Main Results:

• The AUD deep learning model achieved an overall classification accuracy of approximately 56%, with variations between sexes (54% males, 58% females).
• Models for Cannabis Use Disorder (CUD) and Opioid Use Disorder (OUD) showed higher accuracies around 63%.
• Temporal analysis revealed improved model performance in later EEG recording intervals, suggesting dynamic neural patterns.

Conclusions:

• Transformer-based deep learning models demonstrate promise for classifying AUD using raw EEG data, despite current modest accuracy.
• These findings provide a foundational step towards developing objective, EEG-based diagnostic tools for AUD and other substance use disorders.
• Further research and model refinement are warranted to enhance accuracy and clinical utility for psychiatric diagnosis.