Deep neural networks for estimating knee joint kinematics from inertial measurement units

Inertial measurement units (IMUs) are one of the most practical alternatives to optoelectronic systems for human motion analysis. Their low price, flexibility, and ease of use make them highly attractive, as they promise to facilitate out-of-clinic patient monitoring and therapy for a wide range of diseases. The reliance of traditional sensor fusion algorithms on magnetic readings, however, drastically limits real-world applications. In an attempt to address this limitation, we developed machine learning models to predict knee joint kinematics from thigh and shank accelerations and angular velocities, without any reliance on magnetic data.