Learning Interpretable Brain Functional Connectivity via Self-Supervised Triplet Network With Depth-Wise Attention.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Conventional brain functional connectivity analysis faces limitations with deterministic models and empirical analysis.
• Deep learning methods often prioritize state classification over interpretable connectivity characteristics.

Purpose of the Study:

• To propose a self-supervised triplet network with depth-wise attention (TripletNet-DA) for generating interpretable functional connectivity.
• To enhance the capability of deep learning models in capturing functional connectivity dynamics.

Main Methods:

• Utilized channel-wise transformations for temporal data augmentation to create correlated and uncorrelated sample pairs for self-supervised training.
• Employed a convolution network channel encoder and a similarity estimator to extract deep features and generate functional connectivity representations.
• Applied Triplet loss with an anchor-negative similarity penalty to minimize similarities of uncorrelated sample pairs, enhancing learning.

Main Results:

• TripletNet-DA demonstrated superior performance in Autism Spectrum Disorder (ASD) discrimination and Major Depressive Disorder (MDD) classification compared to state-of-the-art methods.
• Achieved high accuracy rates for ASD (up to 98.32%) and MDD (up to 91.80%) classification using specific EEG frequency bands.
• Identified significant functional connectivity differences between ASD and typically developing individuals, aligning with empirical findings.

Conclusions:

• TripletNet-DA offers a powerful approach for generating interpretable functional connectivity from EEG data.
• The model shows potential as a biomarker for clinical analysis in neurological and psychiatric disorders like ASD and MDD.
• This method advances the integration of deep learning with neuroscience for understanding brain dynamics.