Optimizing wearable IMU configurations for running gait analysis: a machine learning-based sensor fusion approach

Area of Science:

• Biomechanics
• Wearable Technology
• Machine Learning

Background:

• High-dimensional sensor networks in wearables face hardware limitations.
• Optimizing sensor placement is crucial for accurate kinematic data.
• Existing methods often focus on activity classification rather than sensor fusion optimization.

Purpose of the Study:

• To determine the feasibility of reducing a 17-sensor network to a "minimal-optimal" subset using machine learning.
• To systematically quantify information redundancy in kinematic chains for sensor fusion.
• To develop a robust, low-cost sensor array design for wearable devices.

Main Methods:

• Applied machine learning (Random Forest regression) to analyze data from 17 IMUs worn by 25 runners.
• Programmatically subsetted sensor signals to simulate minimal configurations.
• Utilized Recursive Feature Elimination (RFE) on time- and frequency-domain features to identify critical sensor attributes.

Main Results:

• A single lumbosacral IMU accurately reconstructed global running parameters (e.g., Cadence, Vertical Oscillation) with R² > 0.95.
• This single-node setup failed to detect gait asymmetry (R² = 0.52).
• A three-sensor configuration (Lumbosacral + Bilateral Ankles) achieved results comparable to the full system (R² > 0.91) for all parameters, resolving asymmetry detection limitations.

Conclusions:

• Validated a machine learning framework for optimizing sensor array design in wearables.
• The proposed three-sensor fusion provides a robust, low-cost blueprint for next-generation devices.
• Optimized sensor placement can achieve high accuracy without complex deep learning models.