Neural network based approximation of muscle and joint contact forces during jumping and landing

Abstract

Musculoskeletal models have been used to estimate the muscle and joint contact forces expressed during movement. One limitation of this approach, however, is that such models are computationally demanding, which limits the possibility of using them for real-time feedback. One solution to this problem is to train a neural network to approximate the performance of the model, and then to use the neural network to give real-time feedback. In this study, neural networks were trained to approximate the FreeBody musculoskeletal model for jumping and landing tasks. The neural networks were better able to approximate jumping than landing, which was probably a result of the greater variability in the landing data set used in this study. In addition, a neural network that was based on a reduced set of inputs was also trained to approximate the outputs of FreeBody during a landing task. These results demonstrate the feasibility of using neural networks to approximate the results of musculoskeletal models in order to provide real-time feedback. In addition, these neural networks could be based upon a reduced set of kinematic variables taken from a 2-dimensional video record, making the implementation of mobile applications a possibility.

Real-Time Prediction of Joint Forces by Motion Capture and Machine Learning

“…Artificial neural networks are commonly used to predict lower limb joint angles and moments [40][41][42][43][44], ground reaction forces [45], joint forces or impulses [46][47][48][49] and contact pressures [50][51][52]. On the other hand, support vector machines have also demonstrated promising prediction performance in various regression biomechanical problems, such as electromyography (EMG)-based prediction of lumbosacral joint loads [53] and optical marker-based prediction of lower limb joint angles and moments [54,55], standing out for their substantial generalization ability to unseen datasets [56].…”

Machine Learning for Optical Motion Capture-Driven Musculoskeletal Modelling from Inertial Motion Capture Data

“…This can be used for translational research that uses MSK model outputs on IMC data obtained in real-world settings, such as supermarkets and manual material handling [21,22,[74][75][76], as a training corpus for an ensemble of ML models which can be used to facilitate clinical diagnosis rapidly. In addition to IMC data, recent work on markerless systems [77], such as twodimensional recordings [35], to obtain mocap quality data provides another source of training data while being used without specialised equipment.…”

“…In particular, using supervised ML models to circumvent biomechanical models that are computationally expensive has become commonplace. Numerous studies have reported ML applications for generating in vivo insights (e.g., joint and muscle loading) from either OMC or IMC inputs [28][29][30][31][32][33][34][35][36][37][38] or markerless methods [27]. Among ML approaches, deep learning models, including Convolutional Neural Networks (CNN) and, increasingly, Recurrent Neural Networks (RNN), have become popular in lowerextremity biomechanical analyses [26].…”