Sensor-Based Gait Parameter Extraction With Deep Convolutional Neural Networks.

Area of Science:

• Biomechanical Engineering
• Machine Learning
• Wearable Technology

Background:

• Objective gait impairment scoring commonly relies on measuring stride-related biomechanical parameters.
• Current double-integration methods using inertial sensors have clinical limitations due to underlying assumptions.
• A need exists for integration-free, data-driven methods for mobile gait analysis.

Purpose of the Study:

• To present a novel method using deep convolutional neural networks (CNNs) to extract spatio-temporal stride parameters from wearable sensor data.
• To enable integration-free and data-driven mobile gait analysis.
• To compare a combined CNN approach with an ensemble CNN approach for parameter estimation.

Main Methods:

• Developed a deep learning framework to translate wearable sensor data into context-related expert features.
• Compared two CNN modeling approaches: a combined network and an ensemble network.
• Validated the method on a clinically relevant, publicly available benchmark dataset.

Main Results:

• The ensemble CNN approach outperformed the combined network for parameter estimation.
• Accurate estimations of stride length, width, and foot angle were achieved.
• Precise estimations of stride, swing, stance, heel, and toe contact times were obtained, comparable to or exceeding state-of-the-art methods.

Conclusions:

• The proposed deep learning methodology offers a viable alternative to assumption-driven double-integration methods.
• This approach enables mobile assessment of spatio-temporal stride parameters in critical clinical situations, such as spastic gait impairments.