A Computational Pipeline for Activity Prediction Using Wearable Sensor Data.

Area of Science:

• Biomechanics
• Wearable Technology
• Machine Learning

Background:

• Wearable sensors facilitate real-world ground reaction force (GRF) data collection.
• Accurate classification of activities from GRF data, especially for subject-specific monitoring, remains a significant challenge.

Purpose of the Study:

• To develop and validate a machine learning pipeline for classifying locomotion-related loading behaviors using wearable GRF data.
• To create a dataset of GRF measurements across various walking conditions for research.

Main Methods:

• Collected GRF data from 14 subjects across 18 speed and incline combinations.
• Segmented continuous GRF signals into gait cycles and extracted features using TSFRESH.
• Employed feature selection and Random Forest classification for step-level activity classification.

Main Results:

• Subject-specific models achieved a mean Top-1 accuracy of 0.664 for classifying loading behaviors.
• Top-2 and Top-3 accuracies reached 0.836 and 0.904, respectively.
• Incline-only and speed-only classifications showed high accuracies (0.688 and 0.903).

Conclusions:

• Step-level GRF data can accurately classify locomotion loading conditions.
• The developed pipeline supports subject-specific models for personalized activity monitoring.
• The dataset and pipeline offer a foundation for advancements in wearable biomechanics and musculoskeletal loading analysis.