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

Walking ability is crucial for maintaining independence and healthy aging. Although joint angle measurement is important for detailed gait assessment, it is rarely performed in clinical practice due to the complexity of motion capture systems. This study investigates individual variability and cross-population generalizability of deep learning-based joint angle estimation from a single inertial measurement unit (IMU) attached to the pelvis. Gait data from three distinct populations were collected: 17 young adults, 20 healthy older adults (aged 65+), and 14 pre-operative patients scheduled for hip replacement surgery due to hip osteoarthritis (also aged 65+). A 1D ResNet-based convolutional neural network was trained to estimate bilateral hip, knee, and ankle joint angles from IMU signals. We systematically compared within-population training (trained and tested on the same population) with cross-population training (trained on combined data from all populations) using nested 5-fold cross-validation. Cross-population training showed population-specific effectiveness: older adults demonstrated consistent improvement, while young adults showed minimal change due to already high baseline performance, and pre-operative patients exhibited highly variable responses. These findings suggest that the effectiveness of cross-population learning depends on within-population gait heterogeneity, with important implications for developing clinically applicable gait analysis systems across diverse patient populations.