Related Experiment Video
Updated: Sep 23, 2026
A Human-machine-interface Integrating Low-cost Sensors with a Neuromuscular Electrical Stimulation System for Post-stroke Balance Rehabilitation
Published on: April 12, 2016
Toward Automation of Motor Assessments in Stroke: Proof-of-Concept Model to Interpret Balance and Gait in Patients
Yee Mah1,2, Vassilios Tahtis2, Abigail Beddard2
1School of Biomedical Engineering and Imaging Sciences, King's College London, London, United Kingdom.
Objective:
To evaluate the feasibility of automated action recognition and movement characterization during the Performance-Oriented Mobility Assessment (POMA) using markerless three-dimensional pose estimation and to explore whether data-driven whole-body movement analysis differentiates healthy adults from stroke survivors.
Design:
Observational, cross-sectional proof-of-concept study.
Setting:
Single-center study conducted at a UK National Health Service Hospital stroke unit.
Participants:
Fifty-six healthy adults and 17 stroke survivors able to complete the POMA assessment (N=73).
Intervention:
Participants completed the POMA under therapist supervision while recorded using 2 synchronized RGB cameras. Major joints were extracted to generate skeletal representations, which were analyzed using machine-learning methods to automatically detect distinct POMA actions.
Main Outcome Measures:
Primary outcomes were overall action recognition accuracy and F1 scores for individual POMA actions. Secondary outcomes included qualitative and quantitative differences in movement patterns between healthy participants and participants with stroke derived from whole-body kinematic embeddings.
Results:
The action recognition model achieved a mean classification accuracy of 78.9%. Performance was highest for sitting, standing, and turning, with F1 scores of 0.95, 0.82, and 0.80, respectively. Low-dimensional embeddings of movement trajectories revealed distinct clustering patterns and greater dispersion among participants stroke, consistent with impaired balance and gait control.
Conclusions:
Markerless pose estimation combined with machine-learning analysis can accurately recognize actions and detect clinically meaningful differences in movement during a standardized balance and gait assessment. Although exploratory, these findings illustrate the potential of artificial intelligence to characterize movement patterns and identify kinematic features distinguishing patient groups. This approach offers a pathway toward scalable, objective, and automated motor assessment, with the potential to identify patients at risk of functional decline and support targeted clinical interventions.
