Researchers tested the strategy under a fixed MHEALTH-to-PAMAP2 transfer protocol. Five-second windows of wearable-signal data were converted into 270 handcrafted descriptors derived from inertial and magnetic measurements. Mutual information estimated each feature’s relevance to activity labels in the source dataset, while unlabeled source–target stability measured how consistently those descriptors behaved across the two domains. The two criteria were used jointly to select a “Stable-48” feature set. Lightweight classifiers were then trained using MHEALTH labels alone, meaning that PAMAP2 labels were withheld until final evaluation.

To make the activity categories comparable, the researchers harmonized the original labels into three broad states: sedentary, active, and vigorous. Within the source dataset, leave-one-subject-out evaluation with a Random Forest produced nearly perfect results: 0.9985 accuracy and 0.9984 macro-F1. The cross-dataset results were more demanding. Stable-48 achieved 0.9042 accuracy and 0.8753 macro-F1 on PAMAP2, compared with 0.8888 accuracy and 0.8436 macro-F1 for all 270 descriptors. Selecting 48 features using source-domain mutual information alone yielded lower scores of 0.8734 accuracy and 0.7940 macro-F1, indicating that the stability component contributed beyond source-label relevance in this experiment.

The researchers also froze all operating values and reversed the transfer direction, moving from PAMAP2 to MHEALTH. Stable-48 again ranked highest, with 0.5811 accuracy and 0.4400 macro-F1. The source-only 48-feature method reached 0.5720 accuracy and 0.4320 macro-F1, while the full 270-feature representation achieved 0.5445 accuracy and 0.4257 macro-F1. However, participant-level comparisons in the reverse direction were not significant after Holm adjustment, so the small numerical advantage should not be interpreted as definitive evidence of superiority. The compact method also performed better than tested shallow target-feature transformations and feature-level multilayer-perceptron adaptation within the same handcrafted descriptor space.

These results suggest that a relatively simple domain-aware selection procedure can make wearable activity models more portable while reducing their feature burden. The gains were not universal: performance fell substantially in the reverse transfer, underscoring the difficulty of transferring recognition systems when datasets differ in participants, sensor setups, acquisition conditions, and activity definitions. The study does not establish that the method surpasses raw-signal deep-learning approaches, and its conclusions are limited to the evaluated transfer protocols and three-state activity taxonomy. It also does not demonstrate clinical readiness. Even so, the findings offer a practical direction for cross-dataset monitoring: a model can be made more compact by selecting descriptors that are both informative in the source domain and stable in the unlabeled target data.

Subject of Research: Evaluation of compact target-aware feature selection for cross-dataset wearable activity recognition.

Article Title: Compact target-aware stability-based feature selection for cross-dataset wearable activity monitoring

Article References: Ayouni, S., Sliti, M., Abdulaziz Alghanem, A., Alhayan, F., & Maddeh, M. (2026). Compact target-aware stability-based feature selection for cross-dataset wearable activity monitoring. PLOS One, 21(10), e0359300. https://doi.org/10.1371/journal.pone.0359300

Image Credits: AI Generated

DOI: 10.1371/journal.pone.0359300

Keywords: wearable sensors, activity recognition, feature selection, domain adaptation, cross-dataset transfer, inertial signals, magnetic signals, machine learning, MHEALTH, PAMAP2, Compact, target-aware