Kernel Granger Causality Based on Back Propagation Neural Network Fuzzy Inference System on fMRI Data.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Biomedical Engineering

Background:

• Granger causality (GC) is widely used for brain network analysis with fMRI, but conventional linear models struggle with nonlinear dynamics.
• Existing GC methods have limitations in detecting nonlinearity and require stationary time series.

Purpose of the Study:

• To propose a novel Granger causality model, back propagation (BP) based kernel function Granger causality (BP_KFGC), for improved causality detection in brain networks.
• To evaluate BP_KFGC's performance against existing methods for both linear and nonlinear systems.
• To apply BP_KFGC for constructing directed weight networks (DWN) for Alzheimer's disease (AD) classification.

Main Methods:

• Developed BP_KFGC utilizing symplectic geometry for embedding dimension and fuzzy inference for time series prediction.
• BP_KFGC is a multivariate approach applicable to linear/nonlinear systems, independent of vector auto-regression models and stationary assumptions.
• Compared BP_KFGC with linear GC, partial GC, neural network GC, and kernel GC using simulated data with varying nonlinearity.

Main Results:

• BP_KFGC demonstrated superior performance in detecting both linear and nonlinear causalities compared to four other methods.
• Application to Alzheimer's disease (AD) patients and healthy controls (HCs) using DWN and support vector machine classification achieved high accuracy (95.89%), sensitivity (93.31%), and specificity (94.97%).

Conclusions:

• BP_KFGC offers a robust method for analyzing brain network causality, overcoming limitations of traditional linear models.
• The proposed method shows significant potential as an auxiliary tool for the clinical diagnosis of Alzheimer's disease.