Accuracy of machine learning techniques for real-time prediction of implanted lower limb mechanics with comprehensive and reduced input parameters.

This study evaluates the accuracy of machine learning techniques for real-time prediction of implanted knee mechanics. A musculoskeletal lower limb model was used to generate joint mechanics for a training dataset of 1500 simulations with varying surgical alignments, loading, and ligament properties. The objective was to determine the minimum input dataset required to estimate implanted biomechanics using three predictive methods: linear-regression, bi-directional long short-term memory (biLSTM), and transformer-based models. Results indicate that the biLSTM model had ∼45% lower nRMSE than the other models with reduced inputs. In the longer-term, this may aid in optimizing implant positioning pre- or intra-operatively.