Prob-sparse self-attention extraction of time-aligned dynamic functional connectivity for ASD diagnosis
Hongwu Chen1, Fan Feng2, Pengwei Lou3,4
1School Hospital, Shandong Technology and Business University, Yantai, China.
Heliyon
|January 13, 2025
Summary
This study introduces a new framework for diagnosing Autism Spectrum Disorder (ASD) using dynamic functional connectivity (DFC) and advanced AI. The model achieves 81.8% accuracy, identifying key brain region differences in ASD patients.
Area of Science:
- Neuroscience
- Artificial Intelligence
- Medical Diagnostics
Background:
- Dynamic functional connectivity (DFC) shows potential for Autism Spectrum Disorder (ASD) diagnosis.
- Extracting discriminative features from complex DFC matrices is challenging.
Purpose of the Study:
- To propose a novel ASD classification framework, PSA-FCN, utilizing time-aligned DFC and Prob-Sparse Self-Attention.
- To improve the accuracy and reliability of ASD diagnosis through advanced machine learning techniques.
Main Methods:
- Developed a Prob-Sparse Self-Attention mechanism for selective global feature extraction.
- Employed self-attention distillation to capture local patterns and reduce dimensionality.
- Constructed a time-aligned DFC matrix to enhance robustness against temporal sensitivity and mitigate overfitting.
Main Results:
- Achieved a classification accuracy of 81.8% on fMRI data from the ABIDE NYU site.
- Outperformed existing methods in ASD classification.
- Identified significant DFC connection variability in brain regions like Cuneus, Lingual gyrus, and Precuneus in ASD patients.
Conclusions:
- The proposed PSA framework demonstrates significant potential for ASD diagnosis.
- The study highlights critical ASD-related brain region biomarkers, aligning with previous research.
- This approach offers a promising tool for both ASD diagnosis and biomarker discovery.
