Interpretable skeleton-based movement profiles for formative feedback in school physical activity education

Schools are an important setting for developing movement competence and supporting physical activity PA). Teachers provide feedback on how learners move, but individualized observation is difficult to sustain and document in large classes. Existing artificial intelligence (AI) systems typically classify an action or predict one competitive-sport score. These outputs do not identify the specific movement feature that a learner might adjust during practice.

Using three public human action datasets (NTU RGB+D, Penn Action, and a FineGym subset), we benchmarked five representative skeleton-sequence models-recurrent (LSTM, GRU), graph-based (ST-GCN, CTR-GCN), and a Transformer-under a unified setting, and constructed four interpretable movement-profile indicators from joint coordinates and trajectories: movement amplitude, postural stability, left-right coordination, and temporal consistency. Each indicator was normalized against the same-class reference distribution, providing a relative, interpretable description rather than an expert-certified score.

Over five random seeds per dataset, independent-test macro-F1 ranged from 0.76 to 0.97; CTR-GCN obtained the highest mean on every dataset (NTU RGB+D 0.970 ± 0.003). Paired tests supported its advantage over all four comparators on NTU RGB+D and FineGym, whereas on Penn Action only the comparison with LSTM remained significant after correction. In the feature ablation, the full kinematic input increased macro-F1 from 0.925 to 0.936 on 3D skeletons (Holm-adjusted p = 0.022); the corresponding 2D change from 0.893 to 0.905 was not significant after correction (p = 0.079). Per-action profiles and a same-action case analysis showed that the indicators capture interpretable kinematic structure. The resulting percentiles describe relative tendencies within these datasets; a low percentile does not by itself identify an incorrect movement or constitute an age-appropriate standard for children.

Skeleton-only movement profiles offer a privacy-reducing, teacher-in-the-loop approach to formative feedback that could support digital PA education and health promotion in schools. As a computational proof-of-concept evaluated on public datasets, the study does not include health or learning outcomes, age-specific K-12 reference distributions, or authentic-classroom validation. Calibration on curriculum-relevant data from children and validation against teacher or movement-expert ratings are required before the profiles can be interpreted as movement-quality feedback or used to support public-health claims.