A Computational Pipeline for Activity Prediction Using Wearable Sensor Data

Wearable sensors enable collection of ground reaction force (GRF) data in real-world settings, but translating these data into meaningful activity classifications remains challenging, particularly for subject-specific monitoring.

We present a dataset of wearable GRF measurements from 14 subjects walking across 18 combinations of speed and incline, along with a machine learning pipeline for step-level classification of loading behaviors. Continuous GRF signals were segmented into individual gait cycles and transformed into features using TSFRESH, followed by feature selection and Random Forest classification. Subject-specific models achieved a mean Top-1 accuracy of 0.664 (SD = 0.053), exceeding chance performance (0.056), with Top-2 and Top-3 accuracies of 0.836 and 0.904, respectively. Accuracy remained similar for incline-only classification (0.688 ± 0.030) but increased for speed-only classification (0.903 ± 0.097).

These results demonstrate that step-level GRF data can support accurate classification of locomotion-related loading conditions and enable the development of subject-specific models for monitoring individual activity. The dataset and pipeline provide a foundation for future work in wearable biomechanics and personalized analysis of musculoskeletal loading.