Physics-Informed Learning Framework for Lower Limb Kinematic Prediction With Sparse Sensors and Its Application in...

Area of Science:

• Biomechanics
• Rehabilitation Engineering
• Machine Learning

Background:

• Gait analysis is crucial for diagnosing diseases and assessing rehabilitation progress, especially in stroke survivors who exhibit altered lower limb kinematics.
• Current precise motion capture systems are lab-bound, while wearable inertial measurement units (IMUs) require multiple sensors, hindering daily life monitoring.
• Accurate lower limb joint angle measurement is vital for evaluating stroke patient functional status and guiding rehabilitation.

Purpose of the Study:

• To develop a physics-informed learning framework using a temporal convolutional network (TCN) for accurate lower limb kinematics prediction.
• To significantly reduce the number of required IMUs for gait analysis to just two.
• To validate the framework's effectiveness in healthy subjects and chronic stroke patients.

Main Methods:

• A physics-informed learning framework integrating geometric physical constraints into a temporal convolutional network (TCN).
• Utilized two IMUs for lower limb kinematics prediction, incorporating human gait modeling and IMU-derived constraints.
• Validated the framework on six healthy subjects and seventeen chronic stroke patients.

Main Results:

• The proposed TCN framework accurately predicted lower limb kinematics using only two IMUs.
• Achieved comparable accuracy to systems using four or more IMUs.
• Demonstrated the framework's effectiveness in both healthy individuals and stroke survivors.

Conclusions:

• The physics-informed TCN framework offers a practical solution for lower limb kinematics assessment using minimal wearable sensors.
• This approach has significant potential for unobtrusive gait monitoring and personalized rehabilitation for stroke patients.
• Reduced sensor count enhances convenience and applicability in real-world rehabilitation settings.