Stroke is a leading cause of death and disability worldwide, significantly threatening human health (Feigin et al., 2022; Lanas and Seron, 2021). Hemiplegia, or one-sided paralysis, is a common consequence, severely impairing lower limb motor function. The National Institutes of Health Stroke Scale (NIHSS) (Kasner et al., 1999; Kwah and Diong, 2014; Lyden et al., 1999) is a widely used tool and important guideline for assessing stroke severity, enabling rapid and localized diagnosis. A crucial component, the motor-leg task, requires patients to raise each leg to a 30-degree angle while lying down and hold for 5 seconds (Fig. 1(a)), with physicians assigning scores from 0 to 4 for each leg based on gravity resistance. Accurate scoring objectively reflects lower limb motor function and stroke severity, aiding early symptom identification, risk assessment, and treatment planning (Chen et al., 2023; Han et al., 2024).

Clinically, physicians visually assess leg-raising performance to determine a motor-leg total score (0-8 across 9 levels). However, this method has limitations. It relies on experienced physicians, who may be scarce in underdeveloped regions, and is prone to inter-rater differences and inevitable oversights (Josephson et al., 2006). Given stroke’s rapid progression and the urgency of intervention (Lindley, 2017), timely and accurate assessment is critical, as delays or inaccuracies can worsen outcomes. Therefore, objective and accurate automated clinical-level (i.e., 9-level) scoring and continuous monitoring of the motor-leg task are crucial for stroke early detection, treatment, and rehabilitation.

Existing motor-leg assessment systems typically rely on wearable sensors for binary scoring. For example, Park et al. (2020) applied sensor-based motion data and machine learning for NIHSS-based binary scoring in 15 participants. While these sensors enable precise motion monitoring in specialized settings, their high cost, complex calibration, and contact-based nature hinder large-scale adoption in daily life. In contrast, video-based assessment is low-cost, highly accessible, and contactless, making it a practical solution for automated scoring in daily life, especially in resource-constrained areas. Building on the successful applications of robotic vision in human action recognition and healthcare (Che et al., 2024; Manakitsa et al., 2024), prior studies have adopted similar visual perception techniques to explore video-based assessments for other neurological diseases (Guo et al., 2022; Liu et al., 2023; Lu et al., 2021; Mehta et al., 2021). However, to the best of our knowledge, no existing work has specifically addressed video-based motor-leg assessment for stroke. Thus, this study aims to achieve clinical-level scoring of the motor-leg task using videos captured by consumer-grade cameras.

Video-based clinical-level motor-leg scoring systems face significant challenges with performance instability (Fig. 1(b)). Physicians typically assess leg resistance to gravity by observing relative motion changes between the motor leg and other non-motor body part on the bed. These stable relative motion changes, determining the score, can be considered causal features consistent across samples. However, lying down for the task can easily introduce sample-level personalized interference factors. Patient-specific factors, such as leg length, lying posture, and skeletal scale, combined with environment-specific factors like varying camera positions and shooting angles, create unstable non-causal biases that distort the skeletal morphology. These biases entangle the causal features, ultimately leading to unstable model performance.

To address these challenges, we aim to suppress non-causal biases induced by personalized interference factors. On one hand, we mitigate interference on skeletal representations of the motor leg, ensuring consistent causal feature perception. On the other hand, we reduce interference on the skeletons of non-motor body part, strengthening the robustness of causal information in decision-making. In summary, we propose a causality debiasing graph convolutional network (GCN) for video-based clinical-level motor-leg scoring. This model leverages a GCN to capture motion-related interactions between human joints, thereby extracting essential motion features. On this basis, the model employs a systematic graph causality decoupling and debiasing strategy by randomly swapping causal (or non-causal) components within the same (or different) score classes. This enhances the consistency of unbiased causal components and eliminates biased non-causal interference, ensuring stable causal feature extraction and prediction. The key points include:(1)

A pose estimator is used to extract joint sequences related to the motor-leg task from video data, followed by a spatial-temporal GCN to model these features.

(2)

The proposed intra-class causality enhancement module decouples key causal graph nodes within samples of the same label, generates unbiased samples with the same non-causal graphs but different causal graphs, and enhances their representation similarity, ultimately mitigates interference on the motor leg.

(3)

The inter-class non-causality suppression module is designed to separate biased non-causal graph nodes across different labels, construct biased samples with shared causal but differing non-causal graphs, thereby reducing decision discrepancies and minimizing interference effects on the non-motor body part.

In conclusion, our GCN framework enables stable clinical-level scoring of the motor-leg task in stroke patients using videos captured with consumer-grade cameras. It provides objective NIHSS-based motor-leg total scores, and demonstrates robust performance in scoring of another stroke motor-arm task and an additional Parkinsonian gait task, offering a convenient tool for timely stroke diagnosis, treatment, and remote monitoring, particularly in underserved regions. Notably, the stroke motor-leg and motor-arm datasets used in this study will be publicly available at https://github.com/SJTUBME-QianLab/Stroke_skeleton_dataset. The primary technical contribution is the development of a graph causality decoupling and debiasing scheme, which effectively separates and utilizes causal and non-causal graphs within the GCN, systematically synthesizes unbiased and biased samples, and significantly enhances the model’s ability to represent causal information and resist non-causal interference, thereby overcoming model instability. Specifically:(1)

The intra-class causality enhancement module is designed to enhance the consistency of discriminative causal features within samples of the same label.

(2)

The inter-class non-causality suppression module is proposed to minimize decision differences caused by biased non-causal features across samples of different labels.