Skeleton-based rehabilitation movement quality classification: a leakage-controlled benchmark on IntelliRehabDS
Frontiers in Medicine · 15.07.2026 · 13
DOI: 10.3389/fmed.2026.1878438 ↗Kurz gefasst
Fachartikel in der Zeitschrift Frontiers in Medicine, online erschienen am 15.07.2026. Verfasst von Zhen Zhu und 2 weiteren Autor:innen. Beteiligte Einrichtung: Department of Orthopedics, Jiujiang Hospital of Traditional Chinese Medicine, Jiujiang, Jiangxi, China. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 22 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Introduction
Automated, objective assessment of rehabilitation movement quality from three-dimensional skeleton recordings can support scalable and low-cost physiotherapy monitoring, but reported performance is often affected by subject leakage, inconsistent class-imbalance handling and non-comparable evaluation protocols.
Methods
This study establishes a leakage-controlled, subject-independent benchmark on the publicly available IntelliRehabDS corpus. After removing 12 ambiguous-label files and 2 sequences shorter than 16 frames, 2,575 sequences from 29 adult and adolescent participants were retained, comprising 2,047 correct and 528 incorrect executions across nine rehabilitation gestures. Sequences were root-centred, torso-scaled, smoothed with a Savitzky–Golay filter, resampled to 64 frames and represented through kinematic features or skeleton tensors. Five classical feature-based learners and the topology-aware ST-SkelNet sequence model were evaluated through nested subject-wise cross-validation. Scaling and SMOTE were fitted only within the relevant training partitions for the classical models, whereas ST-SkelNet used class-weighted binary cross-entropy.
Results
Performance was reported using fold-averaged metrics, pooled out-of-fold confusion matrices and ROC-AUC analysis. No single model dominated: the multilayer perceptron achieved the highest accuracy (0.869), the random forest the highest ROC-AUC (0.861), and the support vector machine the highest incorrect-class recall (0.629). ST-SkelNet achieved 0.846 accuracy, 0.855 ROC-AUC and 0.612 incorrect-class recall, but did not consistently exceed the strongest classical baselines. Head-motion and bilateral-asymmetry features showed descriptive separation between correct and incorrect executions in this adult/adolescent cohort.
Discussion
The contribution is a reproducible methodological benchmark rather than a clinical system or a state-of-the-art architecture. IntelliRehabDS is treated here as a surrogate adult/adolescent dataset for methodological validation and benchmarking, not as infant data. Transfer to paediatric populations, including congenital muscular torticollis, requires independently collected age-specific data and prospective clinical validation.
Schlagwörter
Stichwörter der Autor:innen
Autor:innen und Institutionen
- Department of Orthopedics, Jiujiang Hospital of Traditional Chinese Medicine, Jiujiang, Jiangxi, China.
- School of Rehabilitation Medicine, Gannan Medical University, Ganzhou, Jiangxi, China.
Zitieren
APA
Zhu, Z., Xiang, Y., & Tian, X. (2026). Skeleton-based rehabilitation movement quality classification: a leakage-controlled benchmark on IntelliRehabDS. Frontiers in Medicine, 13. https://doi.org/10.3389/fmed.2026.1878438
Vancouver
Zhu Z, Xiang Y, Tian X. Skeleton-based rehabilitation movement quality classification: a leakage-controlled benchmark on IntelliRehabDS. Frontiers in Medicine. 2026;13. doi:10.3389/fmed.2026.1878438
BibTeX
@article{Zhu2026,
author = {Zhu, Zhen and Xiang, Yuqiong and Tian, Xianzhu},
title = {Skeleton-based rehabilitation movement quality classification: a leakage-controlled benchmark on IntelliRehabDS},
journal = {Frontiers in Medicine},
year = {2026},
volume = {13},
doi = {10.3389/fmed.2026.1878438}
}
EUVIMED