Diffusion-Augmented Spatiotemporal Graph Convolution for Clinical Gait and Motor Function Assessment.

Area of Science:

• Biomedical Engineering
• Computer Science
• Pediatric Neurology

Background:

• Accurate gross motor function assessment in cerebral palsy (CP) is crucial but challenged by data limitations and patient variability.
• Existing deep learning models for CP gait analysis lack generalizability and fairness.

Purpose of the Study:

• To develop a unified generative-diagnostic pipeline for automated clinical gait assessment in children with CP.
• To enhance the accuracy and fairness of gross motor function prediction using skeleton data.

Main Methods:

• Integration of a Conditional Skeleton Diffusion Model (CSDM) for generating synthetic gait data and a Biomechanically-Aware Spatio-Temporal Graph Convolutional Network (BA-STGCN) for analysis.
• CSDM generates 2D skeleton sequences conditioned on clinical variables (GMFCS, GDI) and anthropometrics, preserving biomechanical fidelity.
• BA-STGCN utilizes a symmetry-based loss and multi-task learning for joint GMFCS classification and GDI regression.

Main Results:

• Achieved 85.7% accuracy in Gross Motor Function Classification System (GMFCS) level prediction with balanced precision and recall.
• Reduced mean absolute error in Gait Deviation Index (GDI) prediction to 4.6.
• Significantly improved the recognition of severe CP phenotypes.

Conclusions:

• The proposed conditional skeleton diffusion and biomechanically informed graph learning pipeline offers a scalable, interpretable, and privacy-preserving solution for automated CP gait assessment.
• This approach addresses data scarcity and heterogeneity, paving the way for improved clinical decision-making in pediatric CP management.