An explainable artificial intelligence‑driven motor rehabilitation assessment system based on perception‑cognition decoupling: design and validation
10.3969/j.issn.1006-9771.2026.09.013
- VernacularTitle:基于感知-认知解耦的可解释人工智能运动康复评估系统设计与验证
- Author:
Xuxia GUO
1
;
Hengrui YU
1
;
Yucheng LU
2
;
Sichuang YANG
3
;
Yu LONG
3
;
Jiawei LI
1
;
Lin CHEN
3
;
Feiyu NONG
4
;
Yaobin LONG
4
Author Information
1. School of Physical Education, Guangxi University, Nanning, Guangxi 530004, China
2. School of Civil Engineering and Architecture, Guangxi University, Nanning, Guangxi 530004, China
3. School of Mechanical Engineering, Guangxi University, Nanning, Guangxi 530004, China
4. The Second Affiliated Hospital of Guangxi Medical University, Nanning, Guangxi 530007, China
- Publication Type:Journal Article
- Keywords:
gait analysis;
explainable artificial intelligence;
large language model;
X-Gait system;
3D human pose estimation;
sports rehabilitation
- From:
Chinese Journal of Rehabilitation Theory and Practice
2026;32(9):1108-1116
- CountryChina
- Language:Chinese
-
Abstract:
ObjectiveTo propose an X-Gait motor rehabilitation assessment system based on a perception-cognition decoupling architecture to jointly achieve high-accuracy gait recognition and clinical interpretability for low clinical credibility and opaque decision-making workflows in existing intelligent gait assessment models. MethodsIn the perception layer, YOLOv11-Pose and PoseFormerV2 were integrated for 2D-to-3D human pose reconstruction. A long short-term memory network was adopted for temporal classification of normal, abnormal and pathological gaits. The layer-wise relevance propagation algorithm was used to identify discriminative joints and automatically quantify biomechanical indices including trunk forward inclination angle, stance phase percentage and stride symmetry. In the cognition layer, rehabilitation-oriented prompt engineering and prior knowledge graphs were combined to drive large language models, converting quantitative biomechanical measurements into clinically reasonable structured reports. Eight subjects were enrolled for bedside clinical validation to evaluate classification performance and clinical consistency. ResultsThe overall gait classification accuracy of X-Gait reached 94.1%. The 2D pose detection achieved 100% valid detection rate with an average confidence of 0.819. The mean per-joint position error of 3D pose reconstruction was 105.97 mm. In clinical validation of eight cases, outputs generated by explainable artificial intelligence combined with large language models showed high consistency with manual assessments from experienced rehabilitation physicians. ConclusionThe proposed X-Gait system based on perception-cognition decoupling realizes high-precision gait identification and clinically interpretable outputs.
