Answers are grounded in this paper and cite their sources.
Adult-pretrained models may support paediatric activity recognition, but their performance across datasets and activities remains uncertain. We evaluated harnet10 using thigh accelerometry from 46 typically developing children and 31 children and adolescents with mobility impairments. Training and testing involved separate participants. Source fine tuning achieved participant-equal macro-F1 of 0.904, compared with 0.842 for random initialisation. In the target dataset, five-seed logit ensembles achieved F1 of 0.373 for direct transfer and 0.678 for fine tuning. Fine tuning used five labelled windows per class selected from a fully annotated pool. Across 20 repetitions with independently sampled support sets in each fold, mean F1 was 0.594, compared with 0.372 for matched direct transfer. Each repetition averaged participant-normalised confusion matrices across three optimisation seeds, so the two analyses differed in both support sets and prediction aggregation. Adaptation improved F1 in every repetition and after excluding wheelchair-leading mixed windows. Among 18 participants with wheelchair activity, uncertainty ranking achieved a mean area under the receiver operating characteristic curve of 0.801. The seven-class model predicted no wheelchair windows, giving zero wheelchair recall and F1. Adult pretraining and limited target supervision improved recognition of shared activities in these cohorts. The wheelchair findings highlight the need for activity-specific validation before clinical use.
No takes yet. Share an insight, caveat, or question.
yang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: