Individual Variability in Deep Learning-Based Joint Angle Estimation from a Single IMU: A Cross-Population Study.

Area of Science:

• Biomechanics
• Machine Learning
• Gerontology

Background:

• Joint angle measurement is vital for gait analysis but complex motion capture limits clinical use.
• Wearable inertial measurement units (IMUs) offer a simpler alternative for gait assessment.
• Deep learning models show promise for estimating joint kinematics from IMU data.

Purpose of the Study:

• To investigate the generalizability of deep learning models for joint angle estimation using a single pelvic IMU.
• To compare within-population versus cross-population training strategies for gait analysis.
• To assess the impact of population-specific gait characteristics on model performance.

Main Methods:

• Collected gait data from young adults, healthy older adults, and hip osteoarthritis patients.
• Trained a 1D ResNet convolutional neural network to estimate hip, knee, and ankle joint angles from IMU signals.
• Employed nested 5-fold cross-validation to compare within-population and cross-population training approaches.

Main Results:

• Cross-population training significantly improved gait analysis for older adults.
• Young adults showed minimal improvement due to high baseline performance.
• Pre-operative patients exhibited highly variable responses to cross-population training, highlighting gait heterogeneity.

Conclusions:

• The effectiveness of cross-population learning in gait analysis is influenced by within-population gait variability.
• Findings have implications for developing robust, clinically applicable gait analysis systems for diverse populations.
• Single IMU-based deep learning offers a scalable approach to objective gait assessment.