The Impact of Parameter Uncertainty on Patient-Specific Musculoskeletal Modelling for Ligament-Balanced TKA

A common cause of knee pain is arthritis, where Osteoarthritis (OA) is the most prevalent with an estimated lifetime risk for symptomatic OA of 46.8% in women and 39.8% in men. The goal of a total knee arthroplasty (TKA) is to relieve pain and restore function to a knee with end stage OA. The failures that result in revision are in 47.4% of the cases due to joint stiffness, joint instability or implant loosening. The failure modes that lead to revision are often related to a sub-optimal patient-specific implant position (e.g. post-operative stiffness or instability). Preoperative planning can support the surgeon in determining the optimal patient-specific implant position. Currently, most preoperative planning processes solely account for bone geometry when determining an implant position that is consistent with a mechanically aligned TKA. The joint anatomy as well as several studies describe that not accounting for balanced ligaments when determining the ideal implant position is the cause of different failure types and the high patient dissatisfaction. To account for ligament balancing in pre-operative planning for TKA, a computationally efficient and reliable knee model has to be added to the pre-operative planning process. These models however require several inputs that are uncertain. To apply a computational knee model in clinical practice, a thorough uncertainty quantification is required to assess the effect of patient-specific uncertainties (e.g. ligament stiffness, attachment sites) as well as surgical precision. To convert the uncertainty in the implant position parameters as well as the ligament properties to uncertainty in tibio-femoral (TF) kinematics and ligament strain in a computationally efficient manner, a surrogate model (artificial neural network) approximating the computational knee model was developed. To account for the patient-specific bone and cartilage geometry, they were added as an additional input to the surrogate model. The developed probabilistic, patient-specific knee model can estimate uncertainty in TF kinematics and ligament strain in 16 s. This allows for application of the probabilistic model in pre-operative planning or intra-operatively to aid with surgical navigation for TKA. However, some aspects need to be accounted for in order to establish clinical application. Uncertainty quantification studying the effect of the surgical precision shows that robot assisted surgery is not precise enough for realization of the planned implant position with 90% probability of a ligament balanced outcome. Our results show that to further increase surgical technique precision the focus should be on six critical implant position parameters, namely anterior/posterior and proximal/distal translation, and varus/valgus and internal/external rotation of the femoral component as well as proximal/distal translation and varus/valgus rotation of the tibial component. Uncertainty quantification studying the effect of the ligament properties shows that a success probability of only 12 % on a ligament balanced outcome can be reached without extra measurement of the ligament properties. Our study shows that the reference strain and attachment sites are strongly correlated and most influential for the planned implant position. The first step to mitigate the large uncertainty is to increase knee model robustness wrt the reference strain and attachment sites. If this measure proves insufficient, research should invest in in vivo measurement techniques of the critical properties either through direct measurement or motion tracking.