Automated motor-leg scoring in stroke via a stable graph causality debiasing model
Rui Guo1, Xinyue Li1, Miaomiao Xu2
1Shanghai JiaoTong Affiliated Sixth People's Hospital, School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai 200030, China.
Medical Image Analysis
|May 27, 2025
Summary
Stroke patients often struggle with leg weakness. A new causality debiasing graph convolutional network accurately assesses leg motor function from videos, improving stroke diagnosis and rehabilitation.
Area of Science:
- Biomedical Engineering
- Artificial Intelligence in Medicine
- Rehabilitation Technology
Background:
- Leg motor impairment due to stroke significantly affects patient mobility and daily living.
- Automated assessment of motor function using video analysis is vital for consistent stroke diagnosis and monitoring.
- Real-world video data presents interference challenges, impacting the accuracy and stability of automated motor function scoring.
Purpose of the Study:
- To develop a robust automated system for clinical-level quantification of leg motor impairment in stroke patients using video analysis.
- To address performance instability caused by interference in motion representation and decision-making processes.
- To enhance the reliability and clinical adoption of automated stroke assessment tools.
Main Methods:
- Proposed a novel causality debiasing graph convolutional network (CDGCN) to systematically reduce interference from both motor and non-motor body parts.
- Introduced an intra-class causality enhancement module to stabilize motor-leg representations by separating and refining skeletal graphs.
- Implemented an inter-class non-causality suppression module to mitigate biases from non-causal factors in non-motor body parts.
Main Results:
- Achieved a high correlation (above 0.82) between the CDGCN's motor-leg scores and established clinical scores on a validation dataset.
- Demonstrated consistent performance and stability through independent testing at two additional hospitals.
- Validated the method's reliability on motor-arm scoring and Parkinsonian gait assessment tasks, confirming its generalizability.
Conclusions:
- The proposed CDGCN effectively extracts causal representations from human skeletons, ensuring reliable decision-making for motor function assessment.
- The method shows significant value in real-world scenarios, offering substantial clinical significance and credibility for stroke assessment and telemedicine.
- This approach holds considerable potential for widespread clinical adoption, advancing stroke diagnosis, rehabilitation, and remote patient monitoring.
