Predicting lower limb joint kinematics using wearable motion sensors

• Published in: Gait & posture 2008-07, Vol.28 (1), p.120-126
• Main Authors: Findlow, A, Goulermas, J.Y, Nester, C, Howard, D, Kenney, L.P.J
• Format: Article
• Language: English
• Subjects: Algorithms; Ankle Joint - physiology; Biomechanical Phenomena; Biomedical Engineering - instrumentation; Gait - physiology; Gait analysis; Hip Joint - physiology; Kinemtaic prediction; Knee Joint - physiology; Motion; Motion sensors; Neural network; Orthopedics
• ISSN: 0966-6362; 1879-2219
• DOI: 10.1016/j.gaitpost.2007.11.001

Abstract The aim of this study was to estimate sagittal plane ankle, knee and hip gait kinematics using 3D angular velocity and linear acceleration data from motion sensors on the foot and shank. We explored the accuracy of intra-subject predictions (i.e., where training and testing uses trials from the same subject) and inter-subject (where testing uses subjects different from the ones used for training) predictions, and the effect of loss of sensor data on prediction accuracy. Hip, knee and ankle kinematic data were collected using reflective markers. Simultaneously, foot and shank angular velocity and linear acceleration data were collected using small integrated accelerometers/gyroscope units. A generalised regression networks algorithm was used to predict the former from the latter. The best results were from intra-subject predictions, with very high correlations (0.93–0.99) and low mean absolute deviation (≦2.3°) between measured kinematic joint angles and predicted angles. The inter-subject case produced poorer correlations (0.70–0.89) and larger absolute differences between measured and predicted angles, ranging from 4.91° (left ankle) to 9.06° (right hip). The angular velocity data added little to the accuracy of predictions and there was also minimal benefit to using sensor data from the shank. Thus, a wearable system based only on footwear mounted sensors and a simpler sensor set providing only acceleration data shows potential. Whilst predictions were generally stable when sensor data was lost, it remains to be seen whether the generalised regression networks algorithm is robust for other activities such as stair climbing.