A Neural Network Approach to Physiological Joint Loading Profile Generation in the Kansas Knee Simulator

Abstract

Replicating complex physiological joint loads in an in-vitro simulator allows for research that can be used to improve the outcome of prosthetic design, characterize knee laxity, and evaluate the mechanics of injury. Producing physiological in-vivo loads in an in-vitro simulator can prove challenging due to the complex nature of in-vitro simulators. The goal of this research was to develop a method that can accurately determine the actuator loading profiles necessary to replicate in-vivo joint loads in the Kansas Knee Simulator. Previously, an Adams model was used for profile generation, but was dependent on the fine tuning of small parameters of the system. A neural network was chosen as the basis for a new method to circumvent this issue as neural networks do not rely on information about the system to produce results. Through an iterative process the neural network’s method, inputs, outputs, and necessary training data were determined. A custom built instrumented tibia was implanted into a cadaveric knee and training data was collected and used to train the neural network for a physiological walk cycle. After training, a profile was produced with an RMS error of 31.7 lbs. in the superior-inferior (S-I) direction and 3.8 lbs. in the anterior-posterior (A-P) direction. A second cadaveric knee was used to verify the profile, resulting in an RMS error of 32.4 lbs. in S-I and 12.0 lbs. in A-P. These results are promising, as variation of walk cycles measured in-vivo produced an average RMS difference of 38.3 lbs. in S-I and 7.1 lbs. in A-P. This method shows promise for use as profile generation technique, but additional research is needed to collect the training data to create more profiles in the future.