Neural Networks Estimate Muscle Force in Dynamic Conditions Better than Hill-type Muscle Models

Abstract

Hill-type muscle models are widely used, even though they do not accurately represent certain muscle mechanics. We explored neural networks to develop new muscle models. We trained neural networks to estimate muscle force from activation, muscle length, and muscle velocity. Training data was recorded using sonomicrometry, electromyography, and a tendon buckle on two muscles of guinea fowl. First, we compared the neural network to a Hill-type muscle model, using the same data for network training and model optimization. Second, we trained neural networks on large datasets, in a more realistic machine learning scenario. We found that neural networks generally yielded higher coefficients of determination and lower errors than Hill-type muscle models. Our neural networks performed better when estimating forces on the muscle used for training, but on another bird, than on a different muscle of the same bird, which could be explained by inaccuracies in activation and force scaling. We extracted forcelength and force-velocity relationships from the trained neural networks and found that both effects were underestimated and that both relationships were not replicated well outside of the training data distribution. We discuss suggested experimental designs to collect suitable training data and conclude that neural networks could provide an accurate alternative to Hill-type muscle models, particularly for modeling dynamic muscle behavior that is prevalent in faster movements, given a suitable training dataset, while scaling of the training data should be comparable between muscles and animals. SummaryNeural networks predict muscle forces more accurately than Hill-type muscle models, particularly under dynamic conditions. However, they struggle to replicate the force-length and force-velocity relationships well.