Diagnosis of Autism Disorder Based on Deep Network Trained by Augmented EEG Signals

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Autism spectrum disorder (ASD) diagnosis relies on behavioral assessments, which can be subjective.
• Analyzing electroencephalogram (EEG) signals offers a potential objective biomarker for ASD.
• Deep learning models, particularly 2D-DCNNs, show promise for EEG signal classification but require substantial data.

Purpose of the Study:

• To develop and evaluate a deep learning-based method for ASD detection using EEG signals.
• To address the challenge of limited data in neurological studies through data augmentation.
• To investigate the efficacy of a novel EEG-as-an-image data augmentation technique.

Main Methods:

• EEG signals from individuals with ASD and healthy controls were transformed into image representations.
• A two-dimensional Deep Convolutional Neural Network (2D-DCNN) was employed for image classification.
• A data augmentation technique, termed 'channel combination,' was adapted from image processing for EEG data.

Main Results:

• The channel combination augmentation method, when used with 2D-DCNN, achieved an average accuracy of 88.29% for classifying short EEG signals from healthy individuals and those with ASD.
• 100% accuracy was obtained in distinguishing between ASD and epilepsy using this method.
• Utilizing long EEG signals and a decision on joined windows, 100% accuracy was achieved in detecting ASD subjects.

Conclusions:

• The proposed method of converting EEG signals to images and applying channel combination augmentation is effective for ASD detection.
• Deep learning models, enhanced by appropriate data augmentation, can significantly improve the accuracy of ASD diagnosis from EEG data.
• This approach holds potential for a more objective and efficient diagnostic tool for autism spectrum disorder.