A Causality-Aware Graph Convolutional Network Framework for Rigidity Assessment in Parkinsonians | JoVE Visualize

Area of Science:

Neuroimaging and Computational Neuroscience
Biomedical Engineering and Medical Informatics

Background:

Parkinson's disease (PD) rigidity is a debilitating motor symptom impacting quality of life.
Current rigidity assessment relies on subjective, neurologist-dependent rating scales.
Quantitative Susceptibility Mapping (QSM) shows promise for PD diagnosis, but automated rigidity assessment faces challenges from noise and distribution shifts.

Purpose of the Study:

To develop a robust and automated framework for assessing Parkinson's disease rigidity using QSM.
To address performance instability in automated assessment caused by confounding factors like noise and distribution shifts.
To ensure causality-informed and stable model decisions for reliable rigidity evaluation.

Main Methods:

Proposed a causality-aware graph convolutional network (GCN) framework integrating causal feature selection and causal invariance.
Developed a GCN model with causal feature selection at node, structure, and representation levels, learning a causal diagram to extract truly-causal information.
Implemented a non-causal perturbation strategy with an invariance constraint to ensure assessment stability across different data distributions and avoid spurious correlations.

Main Results:

The proposed causality-aware GCN framework demonstrated superior performance in automated PD rigidity assessment compared to existing methods.
Extensive experiments validated the method's superiority, with selected brain regions showing direct clinical relevance to PD rigidity.
The framework's extensibility was confirmed on related tasks, including PD bradykinesia and Alzheimer's disease mental state assessment.

Conclusions:

The developed causality-aware GCN framework offers a clinically potential tool for automated and stable assessment of Parkinson's disease rigidity.
This approach enhances diagnostic reliability by mitigating confounding factors and focusing on causal relationships within neuroimaging data.
The framework's adaptability suggests broader applications in neurological disorder assessment.