Total knee arthroplasty (TKA) is an effective treatment for end-stage knee osteoarthritis (OA). However, dissatisfaction persists in one in five patients. Musculoskeletal model-based approaches have the potential to predict implant alignment that restores pre-diseased knee functioning and improves patient satisfaction, but their computational complexity limits clinical application. This study aimed to accelerate the estimation of optimal implant alignment using a neural network-based surrogate model, which provides an assessment of the kinematics and ligament strains associated with a candidate alignment. A dataset of eight knee OA patients, including implant positions and the deviations in kinematics and ligament strains between implanted and pre-diseased models, was used to train a multi-task multilayer perceptron (MLP). One MLP was trained per patient to predict implant position quality as quantified by kinematics and ligament strains. A gradient-based optimization method was then applied to the output of the MLP to identify optimal implant positions. Strong correlations were observed between predicted and computed values across all patients. The gradient-based optimization revealed either a single or multiple optimal positions per patient, depending on the geometry of the loss landscape. This reflects the model's ability to capture regions of biomechanical equivalence, as also observed in the musculoskeletal simulations. This proof-of-principle study was designed for patient-specific modeling and therefore focuses on individual optimization rather than generalization to new patients. This study is a critical step toward using biomechanical modeling for the determination of patient-specific optimal implant alignment during TKA in a clinical setting.
Klooster et al. (Sat,) studied this question.