Musculoskeletal models permit the determination of internal forces acting during dynamic movement, which is clinically useful, but traditional methods may suffer from slowness and a need for extensive input data. Recently, there has been interest in the use of supervised learning to build approximate models for computationally demanding processes, with benefits in speed and flexibility. Here, we use a deep neural networks to learn the mapping from movement space to muscle space. Trained on a set of kinematic, kinetic and electromyographic measurements from 156 subjects during gait, the network’s predictions of internal force magnitudes show good concordance with those derived by musculoskeletal modelling. In a separate set of experiments, training on data from the most widely known benchmarks of modelling performance, the international Grand Challenge competitions, generates predictions that better those of the winning submissions in four of the six competitions. Computational speedup facilitates incorporation into a lab-based system permitting real-time estimation of forces, and interrogation of the trained neural networks provides novel insights into population-level relationships between kinematic and kinetic factors.

BackgroundBy altering muscular activation patterns, internal forces acting on the human body during dynamic activity may be manipulated. The magnitude of one of these forces, the medial knee joint reaction force (JRF), is associated with disease progression in patients with early osteoarthritis (OA), suggesting utility in its targeted reduction. Increased activation of gluteus medius has been suggested as a means to achieve this.MethodsMotion capture equipment and force plate transducers were used to obtain kinematic and kinetic data for 15 healthy subjects during level walking, with and without the application of functional electrical stimulation (FES) to gluteus medius. Musculoskeletal modelling was employed to determine the medial knee JRF during stance phase for each trial. A further computer simulation of increased gluteus medius activation was performed using data from normal walking trials by a manipulation of modelling parameters. Relationships between changes in the medial knee JRF, kinematics and ground reaction force were evaluated.ResultsIn simulations of increased gluteus medius activity, the total impulse of the medial knee JRF was reduced by 4.2 % (p = 0.003) compared to control. With real-world application of FES to the muscle, the magnitude of this reduction increased to 12.5 % (p < 0.001), with significant inter-subject variation. Across subjects, the magnitude of reduction correlated strongly with kinematic (p < 0.001) and kinetic (p < 0.001) correlates of gluteus medius activity.ConclusionsThe results support a major role for gluteus medius in the protection of the knee for patients with OA, establishing the muscle’s central importance to effective therapeutic regimes. FES may be used to achieve increased activation in order to mitigate distal internal loads, and much of the benefit of this increase can be attributed to resulting changes in kinematic parameters and the ground reaction force. The utility of interventions targeting gluteus medius can be assessed in a relatively straightforward way by determination of the magnitude of reduction in pelvic drop, an easily accessed marker of aberrant loading at the knee.

In the NeurIPS 2018 Artificial Intelligence for Prosthetics challenge, participants were tasked with building a controller for a musculoskeletal model with a goal of matching a given time-varying velocity vector. Top participants were invited to describe their algorithms. In this work, we describe the challenge and present thirteen solutions that used deep reinforcement learning approaches. Many solutions use similar relaxations and heuristics, such as reward shaping, frame skipping, discretization of the action space, symmetry, and policy blending. However, each team implemented different modifications of the known algorithms by, for example, dividing the task into subtasks, learning low-level control, or by incorporating expert knowledge and using imitation learning.

Linear scaling of generic shoulder models leads to substantial errors in model predictions. Customisation of shoulder modelling through magnetic resonance imaging (MRI) improves modelling outcomes, but model development is time and technology intensive. This study aims to validate 10 MRI-based shoulder models, identify the best combinations of anthropometric parameters for model scaling, and quantify the improvement in model predictions of glenohumeral loading through anthropometric scaling from this anatomical atlas. The shoulder anatomy was modelled using a validated musculoskeletal model (UKNSM). Ten subject-specific models were developed through manual digitisation of model parameters from high-resolution MRI. Kinematic data of 16 functional daily activities were collected using a 10-camera optical motion capture system. Subject-specific model predictions were validated with measured muscle activations. The MRI-based shoulder models show good agreement with measured muscle activations. A tenfold cross-validation using the validated personalised shoulder models demonstrates that linear scaling of anthropometric datasets with the most similar ratio of body height to shoulder width and from the same gender (p < 0.04) yields best modelling outcomes in glenohumeral loading. The improvement in model reliability is significant (p < 0.02) when compared to the linearly scaled-generic UKNSM. This study may facilitate the clinical application of musculoskeletal shoulder modelling to aid surgical decision-making.

No abstract