Accurate and reliable quantification of human biomechanics has allowed for key insights into a broad spectrum of human health and disease conditions. Biomechanical analysis has been applied to the investigation of the development and progression of osteoarthritis, the design of joint replacements, identification of biomarkers of cognitive decline, diagnosis of traumatic brain injury, design and efficacy of musculoskeletal rehabilitation strategies, and risk of musculoskeletal injury. However, the accurate and reliable measurement of biomechanics requires a dedicated laboratory with advanced instrumentation operated by highly trained individuals, and is exceedingly time consuming, all of which significantly limit its use in routine clinical assessments.
The ability to quickly, easily, and reliably perform accurate, health related, biomechanical measurements outside of the laboratory is an unmet clinical need. To address this need, we have developed a unique artificial intelligence based deep learning system to quantify human biomechanics using off the shelf video camera data. However, traditional deep neural networks simply learn patterns contained within data used during training and do not account for the underlying physics that produce that data. Thus, the objective of this Exploratory Internal Research and Development Project was to directly incorporate a physics-based biomechanical model into the SwRI deep neural network markerless biomechanics system to produce more generalizable, accurate, and biomechanically consistent measurements of human biomechanics resulting in a new SwRI Deep Biomechanics System.
Two tasks were proposed to accomplish the project objective. In each task, we directly incorporated a custom biomechanical model into the markerless biomechanics neural network pipeline. Task 1 implemented a new stacked hourglass (SHG) deep neural network backbone along with a new biomechanics-based network training loss function to predict biomechanically consistent results. Task 2 used the output of the original deep neural network, which outputs joint center locations and body segment orientations, as input to a traditional biomechanical model inverse kinematics analysis procedure.
Tasks 1 and 2 demonstrated significant improvements in markerless biomechanics measurements as well as instances of inconsistent results. The new stacked hourglass network (SHG), when compared to the original network, demonstrated consistently higher levels of accuracy measuring joint center positions, reducing the average RMSE by more than 50%. In contrast, however, this network was consistently underperformed compared to the original network when measuring joint angles. However, as hypothesized, the integration of the biomechanical model loss function during network training significantly improved the joint angle predictions of the SHG network. The results from Task 2, postprocessing the output of the original network using a biomechanical model, reduces noise and removes outliers from the data, significantly improving results as hypothesized. In conclusion, the results of this exploratory project indicate that incorporating biomechanics directly into the neural network has the potential to significantly improve overall performance of the SwRI Markerless Biomechanics System.