Feature-selected Graph Spatial Attention Network for Addictive Brain-Networks Identification

TL;DR

This paper introduces a novel graph neural network with feature selection and spatial attention to identify addiction-related brain biomarkers from fMRI data, addressing high dimensionality and noise issues.

Contribution

The study proposes FGSAN, a new model combining spatial attention and Bayesian feature selection for robust brain network analysis in addiction detection.

Findings

FGSAN outperforms existing methods in classifying nicotine addiction.

The model identifies interpretable neural biomarkers linked to addiction.

Enhanced robustness against high-dimensional fMRI data.

Abstract

Functional alterations in the relevant neural circuits occur from drug addiction over a certain period. And these significant alterations are also revealed by analyzing fMRI. However, because of fMRI's high dimensionality and poor signal-to-noise ratio, it is challenging to encode efficient and robust brain regional embeddings for both graph-level identification and region-level biomarkers detection tasks between nicotine addiction (NA) and healthy control (HC) groups. In this work, we represent the fMRI of the rat brain as a graph with biological attributes and propose a novel feature-selected graph spatial attention network(FGSAN) to extract the biomarkers of addiction and identify from these brain networks. Specially, a graph spatial attention encoder is employed to capture the features of spatiotemporal brain networks with spatial information. The method simultaneously adopts a Bayesian feature selection strategy to optimize the model and improve classification task by constraining features. Experiments on an addiction-related neural imaging dataset show that the proposed model can obtain superior performance and detect interpretable biomarkers associated with addiction-relevant neural circuits.