A Causality-Driven Graph Convolutional Network for Postural Abnormality Diagnosis in Parkinsonians.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Abnormal posture is a common, debilitating symptom in Parkinson's disease (PD), increasing fall risk.
• Current PD posture assessment relies on subjective expert judgment, lacking objectivity and consistency.
• Quantitative Susceptibility Mapping (QSM) shows potential for objective PD diagnosis.

Purpose of the Study:

• To develop an automated, QSM-based method for classifying PD patients with and without postural abnormalities.
• To address the challenge of unreliable performance due to non-causal features in PD posture assessment.
• To enhance the stability and reliability of automated PD diagnosis using causal inference.

Main Methods:

• Proposed a causality-driven graph-convolutional-network (GCN) framework with multi-instance learning.
• Implemented an intervention strategy combining non-causal intervenors with causal prediction for enhanced stability.
• Introduced stability and intra-class homogeneity constraints for robust and generalizable feature extraction.

Main Results:

• The proposed method achieved promising performance on a real clinical dataset.
• Extracted features align with previously identified medical markers for PD postural abnormalities.
• Demonstrated a clinically valuable approach for automated, objective, and reliable PD diagnosis.

Conclusions:

• The causality-driven GCN framework offers a reliable method for automated PD postural abnormality detection.
• This approach improves upon subjective clinical assessments, providing objective diagnostic insights.
• The study contributes a robust tool for understanding and diagnosing movement disorders in Parkinson's disease.