Abstract
Introduction:
Primary stability is crucial for long-term fixation of cementless tibial trays. Micromotion less than 50 μm is associated with stable bone ingrowth and greater than 150 μm causes the formation of fibrous tissue around the implant [1, 2]. Finite element (FE) analysis of complete activities of daily living (ADL's) have been used to assess primary stability, but these are computationally expensive. There is an increasing need to account for both patient and surgical variability when assessing the performance of total joint replacement. As a consequence, an implant should be evaluated over a spectrum of load cases. An alternative approach to running multiple FE models, is to perform a series of analyses and train a surrogate model which can then be used to predict micromotion in a fraction of the time. Surrogate models have been used to predict single metrics, such as peak micromotion. The aim of this work is to train a surrogate model capable of predicting micromotion over the entire bone-implant interface.
Methods:
A FE model of an implanted proximal tibia was analysed [3] (Fig. 1). A statistical model of knee kinetics, incorporating subject-specific variability in all 6-DOF joint loads [4], was used to randomly generate loading profiles for 50 gait cycles. A Latin Hypercube (LH) sampling method was applied to sample 6-DOF loads of the new population throughout the gait cycle. Kinetic data was sampled at 10, 50 and 100 instances and FE predictions of micromotion were calculated and used to train a surrogate model capable of describing micromotion over the entire bone-implant interface. The surrogate model was tested for an unseen gait cycle and the resulting micromotions were compared with FE predictions.
Results and discussion:
Accuracy of the surrogate model increased with increasing sample size in the training set; with a LH sample of 10, 50 and 100 trials, the surrogate model predicted micromotion at the bone-implant interface during gait with RMS accuracy of 61, 44 and 33 μm, respectively (Fig. 2). Similar range in micromotion was measured in FE and surrogate models; although the surrogate model tended to over-predict micromotion early in the gait cycle (Fig. 2). There was good agreement in location and magnitude of micromotion at the interface surface through out the gait cycle (Fig. 3). Although encouraging, further work is required to optimize the number and distribution of the training samples to minimize the error in the surrogate model. Analysis time for the FE model was 15 hours, compared to 30 seconds for the surrogate model. The results suggest that surrogate models have significant potential to rapidly predict micromotion over the entire bone-implant interface, allowing for a greater range in loading conditions to be explored than would be possible through conventional methods.
