Multimodal autism detection: Deep hybrid model with improved feature level fusion

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Social communication deficits are hallmarks of autism spectrum disorder (ASD).
• Traditional autism diagnosis relies on subjective behavioral observations, necessitating more accurate methods.
• Existing machine learning approaches for autism detection are complex and time-consuming.

Purpose of the Study:

• To develop a novel deep learning model for accurate autism detection.
• To overcome limitations of traditional diagnostic methods and current computer-aided systems.
• To improve prediction accuracy in autism diagnosis.

Main Methods:

• A multi-modal deep learning approach integrating EEG and facial image data.
• Preprocessing techniques including Gabor filtering for facial images and Wiener filtering for EEG.
• Feature extraction using Common Spatial Pattern (CSP), Singular Spectrum Entropy, Active Appearance Model, and GLCM, followed by feature-level fusion.
• A hybrid detection model combining Convolutional Neural Networks (CNN) and Bidirectional Gated Recurrent Units (Bi-GRU).

Main Results:

• The proposed MADDHM model achieved high accuracy: 91.03% for EEG and 91.67% for facial analysis.
• Outperformed other models including SVM, DNN, Bi-GRU, LSTM, and CNN in accuracy.
• Demonstrated superior performance in autism detection compared to existing methods.

Conclusions:

• The developed methodology shows significant potential for early autism detection.
• This advancement represents a crucial step towards timely interventions for individuals with ASD.
• The study highlights the efficacy of deep learning in improving autism diagnostic accuracy.