Key points are not available for this paper at this time.
Objective: This study applies machine learning (ML) techniques to address this hardware limitation by determining the feasibility of reducing a high-dimensional 17-sensor network to a "minimal-optimal" subset without compromising measurement accuracy. Unlike previous studies focusing on activity classification, we systematically quantify the information redundancy in kinematic chains to optimize sensor fusion architectures. Methods: Twenty-five recreational runners performed treadmill protocols at three speeds (8, 10, and 12 km/h) while wearing a gold-standard Xsens MVN system (17 IMUs). Raw accelerometer and gyroscope signals were programmatically subsetted to simulate minimal configurations. A Random Forest (RF) regression model was selected after benchmarking against baseline Linear Regression and deep learning (LSTM) models. A comprehensive vector of time-and frequency-domain features was extracted via sliding windows, and Recursive Feature Elimination (RFE) was applied to identify the most critical signal attributes. Results: ). Performance remained robust across all running speeds, with only a marginal accuracy drop at 12 km/h. Conclusion: This study validates a machine learning framework for optimizing sensor array design. The proposed three-sensor fusion offers a robust, low-cost architectural blueprint for next-generation wearable devices, proving that complex deep learning is not always required when sensor placement is biomechanically optimized.
Yuan et al. (Wed,) studied this question.