To establish a reproducible benchmark for skeleton-based rehabilitation movement-quality classification using the IntelliRehabDS dataset.
Approach:
Dataset Preparation: Removed ambiguous-label files and short sequences, retaining 2,575 sequences from 29 participants, with 2,047 correct and 528 incorrect executions across nine gestures.
Model Evaluation: Evaluated five classical feature-based learners and the ST-SkelNet sequence model through nested subject-wise cross-validation.
Performance Metrics: Reported performance using fold-averaged metrics, pooled out-of-fold confusion matrices, and ROC-AUC analysis.
Key Findings:
The multilayer perceptron achieved the highest accuracy (0.869).
The random forest model had the highest ROC-AUC (0.861).
The support vector machine recorded the highest incorrect-class recall (0.629).
ST-SkelNet achieved 0.846 accuracy, 0.855 ROC-AUC, and 0.612 incorrect-class recall.
Interpretation:
Head-motion and bilateral-asymmetry features provided descriptive separation between correct and incorrect executions in the cohort studied.
Limitations:
The study does not provide a clinical system or state-of-the-art architecture.
Transfer to pediatric populations requires independently collected age-specific data.
Conclusion:
The study establishes a methodological benchmark for assessing rehabilitation movement quality using skeleton data, highlighting the need for rigorous evaluation practices.
