Spring-loaded inverted pendulum modeling improves neural network estimation of ground reaction forces.

Inertial-measurement-unit (IMU)-based wearable gait-monitoring systems provide kinematic information but kinetic information, such as ground reaction force (GRF) are often needed to assess gait symmetry and joint loading. Recent studies have reported methods for predicting GRFs from IMU measurement data by using artificial neural networks (ANNs). To obtain reliable predictions, the ANN requires a large number of measurement inputs at the cost of wearable convenience. Recognizing that the dynamic relationship between the center of mass (CoM) and GRF can be well represented by using spring mechanics, in this study we propose two GRF prediction methods based on the implementation of walking dynamics in a neural network. Method 1 takes inputs to the network that were CoM kinematics data and Method 2 employs forces approximated from CoM kinematics by applying spring mechanics. The gait data of seven young healthy subjects were collected at various walking speeds. Leave-one-subject-out cross-validation was performed with normalized root mean square error and r as quantitative measures of prediction performance. The vertical and anteroposterior (AP) GRFs obtained using both methods agreed well with the experimental data, but Method 2 yielded improved predictions of AP GRF compared to Method 1 (p = 0.005). These results imply that knowledge of the dynamic characteristics of walking, combined with a neural network, could enhance the efficiency and accuracy of GRF prediction and help resolve the tradeoff between information richness and wearable convenience of wearable technologies.