Unlocking Gait Analysis Beyond the Gait Lab: High-Fidelity Replication of Knee Kinematics Using Inertial Motion Units and a Convolutional Neural Network.

Background: Gait analysis using three-dimensional motion capture systems (3D motion capture) provides a combination of kinematic and kinetic measurements for quantifying and characterizing the motion and loads, respectively, of lower extremity joints during human movement. However, their high cost and limited accessibility impact their utility. Wearable inertial motion sensors offer a cost-effective alternative to measure simple temporospatial variables, but more complex kinematic variables require machine learning interfaces. We hypothesize that kinematic measures about the knee collected using motion capture can be replicated by coupling raw data collected from inertial measurement units (IMUs) to machine learning algorithms. Methods: Data from 40 healthy participants performing fixed walking, stair climbing, and sit-to-stand tasks were collected using both 3D motion capture and IMUs. Sequence to sequence convolutional neural networks were trained to map IMU data to three motion capture kinematic outputs: right knee angle, right knee angular velocity, and right hip angle. Model performance was assessed using mean absolute error. Results: The convolutional neural network models exhibited high accuracy in replicating motion capture-derived kinematic variables. Mean absolute error values for right knee angle ranged from 4.30 ± 1.55 to 5.79 ± 2.93 degrees, for right knee angular velocity from 7.82 ± 3.01 to 22.16 ± 9.52 degrees per second, and for right hip angle from 4.82 ± 2.29 to 8.63 ± 4.73 degrees. Task-specific variations in accuracy were observed. Conclusions: The findings highlight the potential of leveraging raw data from wearable inertial sensors and machine learning algorithms to reproduce gait lab-quality kinematic data outside the laboratory settings for the study of knee function following joint injury, surgery, or the progression of joint disease.