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

Wearable activity-recognition models may lose reliability when transferred across participants, sensor placements, sampling rates, acquisition protocols, and activity definitions. This study evaluates a compact unsupervised target-aware feature-selection strategy for cross-dataset wearable activity monitoring under a pre-specified MHEALTH → PAMAP2 transfer protocol. Five-second wearable-signal windows were represented using 270 handcrafted inertial and magnetic descriptors. Source-domain mutual-information relevance was then combined with unlabeled source–target descriptor stability to select a compact set of 48 features, after which lightweight classifiers were trained using MHEALTH labels only. The original activities were harmonized into three coarse states: sedentary, active, and vigorous, while harmonized PAMAP2 target labels were withheld from model development and used only for final evaluation. Under leave-one-subject-out evaluation on MHEALTH, the Random Forest classifier achieved 0.9985 accuracy and 0.9984 macro-F1. In the primary cross-dataset experiment, the pre-specified stability-aware configuration achieved 0.9042 accuracy and 0.8753 macro-F1 on PAMAP2, compared with 0.8888 accuracy and 0.8436 macro-F1 for the full 270-feature representation and 0.8734 accuracy and 0.7940 macro-F1 for source-mutual-information-only selection. With all operating values frozen, reverse PAMAP2 → MHEALTH transfer also ranked Stable-48 highest (accuracy 0.5811; macro-F1 0.4400), ahead of Source-MI-48 (0.5720; 0.4320) and Full-270 (0.5445; 0.4257), although Holm-adjusted participant-level contrasts were not significant. The compact configuration also achieved higher complete-window accuracy and macro-F1 than the tested shallow target-feature transformations and feature-level MLP adaptation baselines within the same handcrafted descriptor space. These findings indicate that jointly considering source-domain relevance and unlabeled cross-domain descriptor stability can improve compact cross-dataset wearable activity monitoring. The conclusions remain specific to the evaluated protocols and activity taxonomies and do not establish general superiority over raw-signal deep-learning methods or clinical readiness.