Related Experiment Video
Updated: Aug 21, 2026
Deep-Learning Based Multi-Joint Synchronous Tracking for Objective Quantification of Hindlimb Locomotor Kinematics in Rats
Published on: April 3, 2026
Deep Learning-Based Automated Clinical Gait Assessment From Kinematic Data in People With Stroke: Development and
Zhexuan Gu1, Jiaqi Li2,3, Wenbin Zhou4
1Department of Data Science & Artificial Intelligence, Hong Kong Polytechnic University, Hong Kong, China (Hong Kong).
Background:
Gait impairment is a prevalent sequela of stroke. Although observational gait analysis remains a standard clinical practice for assessing neuromotor impairments in people with stroke, it is prone to subjective bias. Consequently, objective assessment is required to inform effective rehabilitation protocols.
Objective:
This study aimed to validate a deep learning framework for automating clinical gait assessment using kinematic data.
Methods:
This study was conducted in a university-affiliated gait analysis laboratory. Kinematic data were collected from 51 individuals with hemiparetic stroke and 18 healthy controls using wearable inertial measurement units, and video recordings were obtained for the Wisconsin Gait Scale (WGS) scoring. WGS scores, rated by 2 experienced physiotherapists, were used to establish the expert reference standard. A statistical graph convolutional network (STAT-GCN), incorporating a STAT-Attention Head module, was developed to predict WGS items 2 to 14 from kinematic data. Model evaluation was performed using a stratified, participant-level 5-fold cross-validation protocol, with all gait cycles from the same participant kept within the same fold. Model performance was assessed at the participant level using exact accuracy, mean absolute error, and weighted κ metrics. A descriptive comparison involving 4 final-year physiotherapy students was also conducted.
Results:
Among participants with stroke, STAT-GCN achieved an overall exact accuracy of 75.42% for WGS items 2 to 14, with a mean absolute error of 0.290 score levels. Across the participant-level 5-fold cross-validation test sets, STAT-GCN achieved a mean exact accuracy of 75.64% (SD 6.58%), whereas the 4 student evaluators achieved mean accuracies ranging from 56.19% (SD 7.03%) to 67.09% (SD 4.03%).
Conclusions:
The proposed STAT-GCN framework demonstrated reasonable preliminary performance for automated WGS item prediction in individuals with stroke. The results suggest that kinematic data-driven models may support more standardized gait assessment, but performance should be interpreted cautiously given the limited sample size, class imbalance across WGS score items, and absence of external validation. Further studies with larger and more clinically diverse stroke cohorts are needed before clinical implementation.
