Wrist-to-Tibia/Shoe Inertial Measurement Results Translation Using Neural Networks.

Area of Science:

• Biomechanics
• Wearable Technology
• Machine Learning

Background:

• Traditional gait analysis relies on lower-limb inertial sensors, which are impractical for everyday use due to fragility and comfort issues.
• Developing accessible gait evaluation methods is crucial for remote patient monitoring and real-world activity tracking.

Purpose of the Study:

• To propose and evaluate an algorithm for translating inertial signals from a wrist-worn sensor to those typically obtained from lower-limb (tibia or shoe) sensors.
• To assess the efficacy of different neural network architectures in this signal translation task.

Main Methods:

• Utilized inertial measurement unit (IMU) data (acceleration and angular velocity) from wrist-worn sensors.
• Developed and compared four neural network models: Dense autoencoder, Convolutional Neural Network (CNN) autoencoder, CNN-Long Short-Term Memory (LSTM) hybrid, and U-Net.
• Translated wrist-sensor signals to simulate tibia- or shoe-sensor signals.

Main Results:

• The CNN autoencoder and U-Net architectures demonstrated successful application for inertial signal translation.
• Gait parameter estimation using the translated signals achieved results comparable to those obtained using traditional shoe-mounted sensors.

Conclusions:

• Wrist-worn sensors, combined with advanced signal translation algorithms, offer a viable and more practical alternative for gait analysis.
• The developed method holds potential for improving the accessibility and user-friendliness of gait monitoring in everyday environments.