Automated motor-leg scoring in stroke via a stable graph causality debiasing model

Abstract

Difficulty in resisting gravity is a common leg motor impairment in stroke patients, significantly impacting daily life. Automated clinical-level quantification of motor-leg videos based on the National Institutes of Health Stroke Scale is crucial for consistent and timely stroke diagnosis and assessment. However, real-world applications are challenged by interference impacting motion representation and decision-making, leading to performance instability. To address this, we propose a causality debiasing graph convolutional network. This model systematically reduces interference in both motor and non-motor body parts, extracting causal representations from human skeletons to ensure reliable decision-making. Specifically, an intra-class causality enhancement module is first proposed to resolve instability in motor-leg representations. This involves separating skeletal graphs with the same score, generating unbiased samples with similar discriminative features, and improving causal consistency. Subsequently, an inter-class non-causality suppression module is designed to handle biases in non-motor body parts. By decoupling skeletal graphs with different scores, this module constructs biased samples and enhances decision stability despite non-causal factors. Extensive validation on the clinical video dataset highlights the strong performance of our method for motor-leg scoring, achieving an impressive correlation above 0.82 with clinical scores, while independent testing at two additional hospitals further reinforces its stability. Furthermore, performance on another motor-arm scoring task and an additional Parkinsonian gait assessment task also successfully confirmed the method’s reliability. Even when faced with potential real-world interferences, our approach consistently shows substantial value, offering both clinical significance and credibility. In summary, this work provides new insights for daily stroke assessment and telemedicine, with significant potential for widespread clinical adoption.