Dynamic functional connections analysis with spectral learning for brain disorder detection.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Dynamic functional connections (dFCs) offer insights into neural activity and brain disorders.
• Existing methods for dFCs analysis often use shallow temporal features, limiting their ability to capture complex temporal patterns.

Purpose of the Study:

• To propose a novel method, dynamic functional connections analysis with spectral learning (dCSL), for effectively exploring inherent temporal patterns of dFCs.
• To enhance the detection of brain disorders by leveraging advanced temporal pattern analysis.

Main Methods:

• dCSL employs a sliding window technique for dFCs estimation.
• A spectral kernel mapping, combining Fourier transform and non-stationary kernels, is constructed.
• This mapping is integrated into a deep kernel network for higher-order temporal pattern analysis via spectral learning.

Main Results:

• The proposed dCSL method achieved a 5% accuracy improvement over general sequence analysis methods.
• dCSL demonstrated a 1.3% accuracy improvement over state-of-the-art dFCs analysis methods.
• Discriminative brain regions for Autism Spectrum Disorder (ASD) detection were identified, aligning with clinical findings.

Conclusions:

• dCSL effectively captures long-range relationships and higher-order temporal patterns in dFCs.
• The method shows significant potential for improving brain disorder detection and understanding.
• Findings support the clinical relevance of identified brain regions in ASD.