Generating Individual Gait Kinetic Patterns Using Machine Learning

In this study, data of 42 healthy individuals walking over a treadmill was used to train and test a neural network that produced individual kinetic patterns of gait cycle as output for a set of atomic features (gender, age, mass, height and gait speed) used as input. The proposed method implements a 3-layer feedforward architecture capable to produce the 3D gait patterns of ankle, knee and hip moment at once, with an average root mean squared error (RMSE) of 7% and average correlation coefficient (\(\rho \)) of 0.94 with respect to the ground truth patterns of the test set. The presented strategy may be used to support individual gait clinical analysis as an alternative to the use of the normal literature pattern that do not take into account the specific characteristics of the patients.