Sparse Adaptive Graph Convolutional Network for Leg Agility Assessment in Parkinson's Disease

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Neurology

Background:

• Motor disorder is a key symptom of Parkinson's disease (PD).
• Current clinical rating scales (MDS-UPDRS) are time-consuming and subjective.
• Telemedicine for PD assessment is crucial, especially post-COVID-19.

Purpose of the Study:

• To develop an automated, objective method for assessing leg agility in Parkinson's disease using video analysis.
• To leverage graph neural networks for fine-grained quantitative assessment of motor symptoms.

Main Methods:

• Utilized a sparse adaptive graph convolutional network (SA-GCN) for skeleton sequence analysis.
• Incorporated temporal context and multi-domain attention modules for comprehensive feature extraction.
• Developed a contactless video-based assessment method.

Main Results:

• The SA-GCN method demonstrated effectiveness and reliability on a dataset of 148 patients and 870 samples.
• The proposed approach outperformed existing state-of-the-art methods in PD motor symptom assessment.
• Achieved fine-grained quantitative assessment of leg agility task.

Conclusions:

• The developed contactless, video-based method offers a reliable and objective tool for automated Parkinson's disease assessment.
• This technology has significant potential for enhancing PD telemedicine and clinical practice.