IMU-Based Classification of Locomotion Modes, Transitions, and Gait Phases with Convolutional Recurrent Neural...

Area of Science:

• Biomechanics
• Robotics
• Machine Learning

Background:

• Accurate classification of human locomotion modes is crucial for advanced prosthetic control.
• Current methods often rely on complex sensor setups or time-domain data.
• Frequency domain analysis offers a computationally efficient alternative.

Purpose of the Study:

• To classify seven locomotion modes and transitions using frequency domain data from minimal inertial measurement units (IMUs).
• To evaluate the effectiveness of different deep neural network architectures, specifically combining convolutional and recurrent layers.
• To assess the system's performance on both healthy subjects and an osseointegrated transfemoral amputee.

Main Methods:

• Utilized frequency domain data from one or two IMUs.
• Implemented and compared various deep neural network configurations, including Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks.
• Employed 5-fold cross-validation for performance evaluation.

Main Results:

• A CNN-LSTM model achieved high mean F1-scores: 0.89-0.91 for healthy subjects and 0.92-0.95 for the transfemoral amputee, using one and two IMUs, respectively.
• The system demonstrated robust classification of locomotion modes and transitions.
• Performance was validated across different sensor configurations.

Conclusions:

• Deep learning models effectively classify locomotion modes using minimal frequency-domain IMU data.
• The proposed CNN-LSTM approach shows significant promise for real-time control of transfemoral prostheses.
• This study highlights the potential of frequency-domain analysis and deep learning for wearable sensor applications.