Identifying Autism Spectrum Disorder From Resting-State fMRI Using Deep Belief Network

Area of Science:

• Neuroscience
• Computational Psychiatry
• Machine Learning

Background:

• Autism spectrum disorder (ASD) diagnosis relies on clinical observation, lacking objective biomarkers.
• Neuroimaging reveals functional connectivity anomalies in ASD, but computational models struggle with large datasets.
• Early identification of ASD is crucial for effective intervention and treatment.

Purpose of the Study:

• To develop a novel graph-based classification model for improved ASD diagnosis using the Autism Brain Imaging Data Exchange (ABIDE) database.
• To enhance diagnostic accuracy and reduce computational complexity in ASD identification.
• To explore potential ASD subtypes and identify key neural correlation patterns.

Main Methods:

• Utilized a graph-based deep belief network (DBN) model.
• Employed K-nearest neighbors and a restricted path-based depth-first search for feature selection.
• Implemented automatic hyperparameter tuning for DBN optimization.
• Applied data augmentation and oversampling techniques.

Main Results:

• Achieved a 6.4% performance improvement over existing models on the ABIDE database.
• Demonstrated superior diagnostic classification accuracy for ASD.
• Reduced computational complexity and training time through feature reduction.
• Identified remarkable autistic neural correlation patterns.

Conclusions:

• The proposed DBN model offers a more reliable and efficient approach to ASD diagnosis.
• The model's interpretability aids in understanding ASD-related neural mechanisms.
• Further research can explore ASD subtypes using the developed techniques.