Solving musculoskeletal biomechanics with machine learning

Abstract

Deep learning is a relatively new computational technique for the description of the musculoskeletal dynamics. The experimental relationships of muscle geometry in different postures are the high-dimensional spatial transformations that can be approximated by relatively simple functions, which opens the opportunity for machine learning applications. In this study, we challenged general machine learning algorithms with the problem of approximating the posture-dependent moment arm and muscle length relationships of the human arm and hand muscles. We used two types of algorithms, light gradient boosting machine (LGB) and fully connected artificial neural network (ANN) solving the wrapping kinematics of 33 muscles spanning up to six degrees of freedom (DOF) each for the arm and hand model with 18 DOFs. The input-output training and testing datasets were generated by our previous phenomenological model based on the autogenerated polynomial structures (Sobinov et al., 2019). Both models achieved a similar level of errors: ANN model errors were 0.08{+/-}0.05% for muscle lengths and 0.53{+/-}0.29% for moment arms, and LGB model made similar errors--0.18{+/-}0.06% and 0.13{+/-}0.07%, respectively. LGB model reached the training goal with only 10^3 samples, while ANN required 10^6 samples; however, LGB models were about 39 slower than ANN models in the evaluation. The sufficient performance of developed models demonstrates the future applicability of machine learning for musculoskeletal transformations in a variety of applications, such as in advanced powered prosthetics. Author SummaryThe accurate decoding of arm and hand motor intent from biological signals remains a key challenge. Solving this task with machine learning requires vast posture- and task-dependent data for identifying structural and functional parameters within dynamic musculoskeletal relationships. This problem is related to the curse of dimensionality where the processing complexity grows exponentially with the number of degrees of freedom described by the model. Here, we developed a tool based on artificial neural networks (ANN) to solve the kinematic transformation from posture to muscle path length and muscle moment arms. We used an accurate model of posture-dependent muscle moment arms and length to train and test the ability of ANN to solve this high-dimensional and computationally intense transformation and compare it to the boosted decision tree approach. We demonstrated that model-driven training is an efficient method to handle the encoding of high-dimensional musculoskeletal relationships. Adding muscles to the transformation, which increases the input-output complexity, does not reduce the prediction accuracy and does not require the increase in the number of elements within the network demonstrating the viability of this approach for applications using musculoskeletal biomechanics.