Self-training EEG discrimination model with weakly supervised sample construction: An age-based perspective on ASD evaluation | JoVE Visualize

Area of Science:

Neuroscience
Artificial Intelligence
Biomedical Engineering

Background:

Deep learning models for Electroencephalography (EEG) excel at brain disorder discrimination but struggle with limited labeled data and individual subject variations, especially for Autism Spectrum Disorders (ASD).
Existing methods face challenges in optimizing EEG models due to data scarcity and the unique characteristics of individual subjects.

Purpose of the Study:

To develop an efficient framework, STEM (Self-Training EEG Model), for optimizing EEG discrimination models in the presence of limited labeled samples and subject individuality.
To enhance the performance of deep learning models for Autism Spectrum Disorder (ASD) detection using EEG data.

Main Methods:

Developed the STEM framework utilizing multi-task learning for model initialization, combining an AutoEncoder with a classifier to learn EEG representations and prediction probabilities.
Implemented a Pseudo-Labeled Sample Construction (PLASC) approach to assign trustworthy pseudo-labels to unlabeled samples, aiding self-training and model optimization.
Employed depth-separable convolutions and BiGRUs within the AutoEncoder for comprehensive EEG representation learning via reconstruction tasks.

Main Results:

The STEM framework achieved superior performance in ASD discrimination using resting-state EEG data from 175 children, with an accuracy of 88.33% and an F1-score of 87.24%.
STEM's multi-task learning outperformed traditional supervised methods when labeled data was scarce.
The PLASC method significantly improved ASD discrimination accuracy (3%-8%) and F1-scores (4%-10%) across different age groups compared to existing methods.

Conclusions:

The STEM framework effectively addresses the limitations of data scarcity and subject individuality in EEG-based deep learning for brain disorder discrimination.
STEM demonstrates a promising approach for enhancing the accuracy and adaptability of models in complex diagnostic scenarios like ASD detection.
The proposed multi-task learning and pseudo-labeling strategies offer a robust solution for improving deep learning model performance with limited labeled EEG data.