How improper dataset split hinders model generalizability: a systematic comparison in Human activity recognition and...

Human Activity Recognition (HAR) and exercise assessment models are increasingly used in healthcare to support clinical evaluation, rehabilitation, and remote monitoring. However, their real-world applicability critically depends on the ability to generalize across unseen subjects, whose movement patterns may differ substantially due to inter-individual variability. Despite this, many studies adopt random noncross-subject (NCS) data splits, where samples from the same individual appear in both training and test sets, potentially leading to overly optimistic and clinically misleading performance estimates.

We investigate (i) how NCS and cross-subject (CS) splits affect performance estimation across machine learning and deep learning models under tasks of increasing complexity, (ii) how data splitting and differences between training and test sets contribute to predictive variance and stability.

Experiments were performed using a large-scale HAR benchmark dataset (NTU RGB+D 120) and a rehabilitation-specific dataset (IntelliRehabDS). A total of 12 machine learning and deep learning models were trained across both tasks, and their performance was estimated and compared using a simulation-based approach. Predictive variance decomposition, via Generalized Linear Mixed-Effects models, was applied to link the split strategy and differences in training and test instances to model output stability.

NCS splits consistently overestimated model performances, with discrepancies increasing alongside task and model complexity. DL architectures, in particular, showed markedly higher NCS performance compared to CS splits, generally with statistical significance. Variance decomposition revealed that greater subject difference between training and test sets often enhances predictive instability, while CS splitting reduces variance by promoting more generalizable representations.

Improper dataset splits can mislead model evaluation, exaggerate generalization capabilities, and undermine clinical trust. Our study provides empirical evidence for computer vision-based rehabilitation models and offers methodological guidance for robust evaluation practices, supporting reproducible and trustworthy AI deployment in rehabilitation and broader healthcare applications.