Feed forward artificial neural network to predict contact force at medial knee joint: Application to gait modification

Knee contact force (KCF) is one of the most meaningful parameters to evaluate function of the knee joint. However in vivo measurement of KCF is not always straight forward. Inverse dynamics analysis, as one of the most frequently used computational techniques to calculate KCF, has its own limitations. The purpose of this study was to develop a feed forward artificial neural network (FFANN) to predict the medial condyle KCF corresponding to two different gait modifications known as medial thrust and trunk sway. Four patients implanted with unilateral knee sensor-based prostheses were obtained from the literature. The network was trained based on pre-rehabilitation gait patterns and was recruited to predict the medial KCF associated with rehabilitation patterns. Generalization ability of the proposed network was tested within three different levels including intra subject (level 1), inter condition (level 2) and inter subject (level 3). FFANN predictions were validated against in vivo measurements.Results showed subject-specific neural network could predict KCF to a certain high level of accuracy (medial thrust  : NRMSE¯=10.6%, ρ¯=0.96; trunk sway  : NRMSE¯=9.6%, ρ¯=0.96) based on the ground reaction forces (GRFs) and some independent marker trajectories (level 1) which suggested that not all of the markers are necessary for knee force calculation. Moreover at level 2, a generic FFANN could predict the medial knee force based on electromyography (EMG) signals and GRFs (medial thrust  :NRMSE¯=11.2%, ρ¯=0.96; trunk sway  :NRMSE¯=10.5%, ρ¯=0.95) which released the necessity of motion capture and subject specific scaling of a musculoskeletal model. At level 3, neural network could predict the general pattern and features of KCF for a new subject that was not used in the network training (medial thrust  : NRMSE¯=12.6%, ρ¯=0.95; trunk sway  : NRMSE¯=13.3%, ρ¯=0.94).In conclusion, FFANN could predict the medial knee joint loading corresponding to two different knee rehabilitations based on pre-rehabilitation gait patterns. Compared to the inverse dynamics method, artificial intelligence represents a much easier and faster method; together they can be combined to calculate joint loading involving fewer markers and speed up the calculations.