Novel relative relevance score for estimating brain connectivity from fMRI data using an explainable neural network...

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Brain connectivity is complex, involving directional, non-linear, and time-lagged dependencies.
• Deep neural networks (DNNs) excel at pattern recognition but often function as "black-boxes" in neuroscience, hindering result interpretation.
• There is a need for interpretable methods to analyze brain connectivity using DNNs.

Purpose of the Study:

• To develop a novel, explainable neural network (xNN) approach for quantifying brain connectivity.
• To introduce a new metric, the relative relevance score (xNN-RRS), for measuring higher-order connectivity between brain regions.
• To provide an intuitive and transparent method for interpreting DNN weights in the context of brain connectivity.

Main Methods:

• A neural network (NN) based predictor was developed for regression tasks.
• Layer-wise relevance propagation (LRP) was employed to determine the contribution of past data points to predictions, enabling the explanation of DNNs.
• A novel quantitative measure, the relative relevance score (xNN-RRS), was proposed based on NN weights.

Main Results:

• The face validity of the xNN-RRS approach was demonstrated using simulated data, outperforming existing methods.
• Construct validity was confirmed through experiments on resting-state fMRI data.
• The proposed method exhibited superior performance in accuracy and computational complexity compared to state-of-the-art brain connectivity estimation techniques.

Conclusions:

• The developed xNN-RRS method offers a promising, explainable approach for brain connectivity analysis.
• This method provides superior accuracy and computational efficiency over existing techniques.
• The approach is well-suited for clinical applications requiring post-hoc explainability in brain connectivity analysis.