Learning based lower limb joint kinematic estimation using open source IMU data.

Area of Science:

• Biomechanics
• Machine Learning
• Wearable Technology

Background:

• Estimating lower-limb joint kinematics is crucial for biomechanical analysis.
• Inertial Measurement Units (IMUs) offer a portable solution but traditional methods require extensive calibration and data.
• Deep learning models can overcome these limitations but typically demand large datasets.

Purpose of the Study:

• To develop and evaluate a deep learning framework for estimating lower-limb joint kinematics using IMUs.
• To address the data requirements of deep learning models by exploring different training strategies.
• To enhance the generalizability and personalization of IMU-based kinematic estimation.

Main Methods:

• Leveraged an open-source dataset for training and evaluation.
• Developed three distinct training approaches: single-user, multi-user, and multi-user with transfer learning.
• Analyzed optimal IMU placement combinations, focusing on femur and calcaneus sensors.

Main Results:

• Single-user training yielded high accuracy for the individual but lacked generalizability.
• Multi-user training improved generalizability but reduced accuracy due to gait variations.
• Transfer learning significantly enhanced accuracy for new users with minimal data, achieving performance comparable to inverse kinematics.
• Optimal IMU placement was identified on the femur and calcaneus.

Conclusions:

• A deep learning framework with transfer learning offers an effective solution for estimating lower-limb joint kinematics using IMUs.
• This approach overcomes the need for extensive data collection and complex calibration, making it suitable for diverse populations.
• The framework enables personalized gait analysis, improving efficiency and accessibility in clinical and real-world applications.