Deep-Fusion of Scalogram and Spatio-Temporal EEG Features with Attention Mechanism for Autism Spectrum Disorder...

Area of Science:

• Neuroscience
• Biomedical Engineering

Background:

• Autism Spectrum Disorder (ASD) is a neurodevelopmental condition impacting social interaction, communication, and behavior.
• Early diagnosis of ASD is crucial for effective intervention and improved quality of life.
• Current diagnostic methods can be subjective and time-consuming.

Purpose of the Study:

• To develop and validate a novel deep learning model for objective Autism Spectrum Disorder detection.
• To explore the efficacy of integrating resting-state electroencephalography (EEG) and scalogram features for ASD identification.
• To enhance diagnostic accuracy through advanced feature fusion and attention mechanisms.

Main Methods:

• A novel deep learning pipeline was developed, combining spatiotemporal EEG features with image-based scalogram features.
• Feature fusion was processed using an EfficientNet-based model incorporating a Convolutional Block Attention Mechanism.
• The attention module utilized channel and spatial attention to refine feature extraction for improved ASD classification.

Main Results:

• The proposed model achieved an accuracy of 84% in distinguishing Autism Spectrum Disorder from healthy controls.
• Performance metrics included an F1-score of 82%, recall of 82%, precision of 82%, and an AUC-ROC of 0.89.
• These results demonstrate the significant potential of EEG-based analysis for ASD detection.

Conclusions:

• Resting-state EEG offers a non-invasive, cost-effective, and objective method for early Autism Spectrum Disorder detection.
• The integration of EEG, scalograms, and deep learning with attention mechanisms enhances diagnostic accuracy.
• EEG signatures show promise as reliable biomarkers for ASD, supporting clinical screening tools.