Evaluating Sparse Inertial Measurement Unit Configurations for Inferring Treadmill Running Motion.

Area of Science:

• Biomechanics
• Sports Science
• Wearable Technology

Background:

• Inertial measurement units (IMUs) are crucial for analyzing running performance.
• Limited sensor data (sparsity) restricts the evaluation of digital biomarkers.
• Shallow recurrent decoder networks (SHRED) can reconstruct dense time-series signals from single sensors, showing promise for human mobility analysis.

Purpose of the Study:

• To evaluate the potential of SHRED algorithms for monitoring running performance.
• To train and test subject-specific SHRED models for mapping single IMU input to multiple IMU outputs.
• To investigate the impact of input parameters (sensor location, type, sampling rate, speed) on SHRED inference accuracy.

Main Methods:

• Trained and tested subject-specific SHRED models on nine subjects running on a treadmill.
• Mapped data from one IMU to the remaining three IMUs.
• Varied sensor location, sensor type, sampling rate, and running speed to assess inference error.

Main Results:

• Sensor location and type did not significantly affect SHRED inference accuracy.
• Decreasing sampling rate impacted the accuracy of ankle measurements.
• Inferred ankle acceleration remained below the minimal detectable change threshold (12.0 m/s²).
• SHRED models struggled to accurately infer IMU measurements below this threshold when trained/tested at multiple speeds.

Conclusions:

• SHRED demonstrates potential for reconstructing dense running kinematics and kinetics from limited IMU data.
• The method's accuracy is sensitive to sampling rate, particularly for ankle measurements.
• SHRED can potentially enhance motion analysis by enabling richer datasets with fewer sensors.