Machine Learning for Musculoskeletal Modeling of Upper Extremity

• Published in: IEEE sensors journal Vol. 22; no. 19; pp. 18684 - 18697
• Main Authors: Sharma, Rahul, Dasgupta, Abhishek, Cheng, Runbei, Mishra, Challenger, Nagaraja, Vikranth H.
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.10.2022; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Analytical models; Biological system modeling; Biomechanical analysis; Biomechanics; Computational modeling; Data models; Decision analysis; Decision making; Extremities; inertial motion capture (IMC); Kinematics; linear models; Load modeling; Machine learning; machine learning (ML); Modelling; motion analysis; Motion capture; musculoskeletal (MSK) modeling; Neural networks; neural networks (NNs); Sensors; Training; upper extremity
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2022.3197461

We propose a novel machine learning (ML)-driven methodology to estimate biomechanical variables of interest traditionally obtained from upper-extremity musculoskeletal (MSK) modeling. MSK models facilitate personalized modeling, perform "what-if" analyses, and potentially enhance clinical decision-making. In certain settings, MSK models are driven by inertial motion capture (IMC) data. IMC systems are portable, user-friendly, and relatively affordable as well as provide additional biomechanical information. However, MSK models can be computationally expensive, often require extensive data, and can be prohibitively slow in making real-time predictions. Our ML method-involving a rigorous hyperparameters search-predicts kinematic and kinetic biomechanical information associated with human upper-extremity movements solely using IMC input data, thereby bypassing MSK models. The scaled cadaver-based MSK model was based on the Dutch Shoulder Model and the spine model implemented in the AnyBody Managed Model Repository. We employ neural networks (NNs), which are trained on IMC data obtained from five nondisabled subjects in subject-exposed (SE) settings (at least a trial from all subjects is used in training) and subject-naive (SN) settings (no information from test subjects is used in training). We compare the predictions of our ML model with that of an MSK model and find an excellent agreement in SE settings (average Pearson's <inline-formula> <tex-math notation="LaTeX">{r}={0.96} </tex-math></inline-formula>, <inline-formula> <tex-math notation="LaTeX">\text {normalized RMSE (NRMSE)} ={0.23} </tex-math></inline-formula>) and strong correspondence in SN settings (average <inline-formula> <tex-math notation="LaTeX">{r}={0.89} </tex-math></inline-formula>, <inline-formula> <tex-math notation="LaTeX">\text {NRMSE} ={0.45} </tex-math></inline-formula>). The linear model performed poorly for SN settings, which motivated a more complex ML model. Our findings suggest that an ML-based approach is highly viable for estimating upper-extremity biomechanical information solely from IMC data.