Evaluation of machine learning techniques for real-time prediction of implanted lower limb mechanics

Accurate prediction of knee biomechanics during total knee replacement (TKR) surgery is crucial for optimal outcomes. This study investigates the application of machine learning (ML) techniques for real-time prediction of knee joint mechanics.

A validated finite element (FE) model of the lower limb was used to generate a dataset of knee joint kinematics, kinetics, and contact mechanics. The models were trained on joint alignment data, ligament information, and external boundary conditions. Several predictive algorithms were explored, including linear regression (LRM), multilayer perceptron (MLP), bi-directional long short-term memory (biLSTM), convolutional neural network (CNN), and transformer-based approaches. The performance of these models was evaluated using average normalized root mean squared error (nRMSE).

The biLSTM model achieved the highest accuracy, with a significantly lower nRMSE compared to other models. Compared to traditional FE or rigid body dynamics models, these predictive models offered significantly faster prediction speeds, enabling near-instantaneous insights into the TKR system's performance. The small size of the predictive models makes them suitable for deployment on edge devices potentially used in operating rooms.

These findings suggest that real-time biomechanical prediction using biLSTM models has the potential to provide valuable feedback for surgeons during TKR surgery. Applications of this work could be applied to provide pre-operative guidance on optimal target implant alignment or given the real-time prediction ability of these models, could also be used intra-operatively after integration of patient-specific intra-op kinematic and soft-tissue information.