Motion-Driven Neural Optimizer for Prophylactic Braces Made by Distributed Microstructures

TL;DR

This paper tackles joint injury prevention by designing personalized prophylactic braces using a motion-driven, differentiable optimization framework. It jointly models biomechanics and physics through a time-variant SIMP-like topology optimization whose design field is represented by a neural network, enabling distributed microstructures to tailor stiffness along the body surface. The approach is validated by fabricating knee and ankle braces and performing physical try-ons and motion assessments, showing preserved sagittal mobility while reducing adverse frontal-plane motions. The method advances toward end-to-end, data-informed, personalized protective wearables, with potential impact on injury prevention and performance in sports and daily activities.

Abstract

Joint injuries, and their long-term consequences, present a substantial global health burden. Wearable prophylactic braces are an attractive potential solution to reduce the incidence of joint injuries by limiting joint movements that are related to injury risk. Given human motion and ground reaction forces, we present a computational framework that enables the design of personalized braces by optimizing the distribution of microstructures and elasticity. As varied brace designs yield different reaction forces that influence kinematics and kinetics analysis outcomes, the optimization process is formulated as a differentiable end-to-end pipeline in which the design domain of microstructure distribution is parameterized onto a neural network. The optimized distribution of microstructures is obtained via a self-learning process to determine the network coefficients according to a carefully designed set of losses and the integrated biomechanical and physical analyses. Since knees and ankles are the most commonly injured joints, we demonstrate the effectiveness of our pipeline by designing, fabricating, and testing prophylactic braces for the knee and ankle to prevent potentially harmful joint movements.

Figures (11)

• Figure 1: Overview of our computational pipeline, which includes four major parts: (a) biomechanical analysis (Sec.\ref{['sec:biomechAnalysis']}), (b) physical analysis (Sec.\ref{['sec:designFramework']}), (c) the neural network (NN) based representation of design function (Sec.\ref{['sec:nerualrepresentation']} and Sec.\ref{['sec:network']}), and (d) microstructure generation (Sec.\ref{['sec:FGM']}), along with a biomechanical data preprocessing (Sec.\ref{['sec:dataProcess']}, marked in black dotted box). From the input motion sequence and paired ground reaction forces (GRFs), we first conduct a preprocessing step to determine the joint angle, the muscle forces, and the design domain as the body surface with the help of a predefined musculoskeletal model. These terms remain unchanged throughout the optimization (as marked in gray blocks). Our topology optimization process, informed by the time-variant forces $F_\text{acting}$ given by the biomechanical analysis (a) and governed by the FEA-based physical analysis (b), iteratively updates the NN's weights (c) as design variables to change the distribution of stiffness on a brace while minimizing the compliance energy $W_c$. In our pipeline, the biomedical analysis and the physical analysis are closely coupled since the reaction force exerted by the braces will further impact the kinematics. The forces generated by a brace at its upper and lower boundaries $F_\text{reaction}$ are applied to the femur and tibia bones, respectively, prompting updates to the joint contact forces for subsequent analysis. (d) The distributed microstructures for realizing the optimized distribution of stiffness can be generated by first computing the distribution of elastic energy $W_e$ and then filling the regions in low and high elastic energies with the well-blended firm (displayed in light gray) and soft microstructures (displayed in light blue).
• Figure 2: (a) Overview of the biomechanical model and the marker layout alignment for kinematics and kinetics analysis, where we customize a detailed knee compartment model biomodelLERNER2015644kneeModelctx344555731410001631 by equipping both knees with three DoFs and adjusting the marker layout (with 96 markers as shown with blue dots) to accommodate more detailed kinematics. (b) From the musculoskeletal model, we extract the outer skin surface by fitting it to the SKEL model ($\beta$, $\mathbf{q}$) keller2023skel, where the shape parameters $\beta$ are obtained from the scanned body as shown in (c) to estimate detailed deformations. The color map measures the shape approximation error between the scanned and updated SKEL bodies.
• Figure 3: We show the kinetic analysis of an example left knee joint during the tennis serve motion, specifically at the landing pose. The biomechanical dynamic analysis can estimate (a) the joint contact forces, $\mathbf{F}_{JCF}$, acting on the tibia bone, (b) the muscle activations and (c) the sum of muscle forces $\sum \mathbf{F}_{mus}$. The acting force (d) can then be computed as Eq. (\ref{['eq:actingForce']}), which serves as the load input for the integrated FEA. Our optimizer considers the entire sequence of a complex motion, such as tennis serving, which includes several critical phases such as deep bending knees, jumping, and landing.
• Figure 4: Different types of microstructures are studied in our research. We provide structures represented by an implicit surface $f_{cell}(\hat{u},\hat{v})$ in the $u,v$-domain, and the corresponding solids obtained by $f_{cell}(\hat{u},\hat{v}) \leq 0.2$ with a user-specified thickness are displayed in blue at the bottom right corner. The function $f_{cell}(\hat{u},\hat{v})$ is formulated as the distance field to the skeletons (highlighted by curves in orange).
• Figure 5: Physical validation of prophylactic brace for the left knee. Left: (d) Motion capture system with twelve Vicon Vero 13 cameras. (a) Motion capture protocol for 13 markers placed on left lower extremity: 2 clusters on femur and tibia -- each with 4 markers (see red boxes), which are used to compute knee angles. Five more calibration markers located on left hip, lateral and medial of left knee and ankle to identify joint and body segments. The participant repeats the tennis serving motion 5 times for (b) unbraced, (c) wearing our optimized brace, (e) a commercial prophylactic brace and (f) an unoptimized brace with uniform 'Voronoi Pattern' microstructure. Right: plot of joint angles for comparison, which are obtained from the average of synchronized curves for each case. Note the peak flexion is used for alignment and the second peak flexion is selected for comparing the angles for both planes at the landing moment, as this passive action can elevate the risk of joint injury, especially if the knee lacks adequate flexibility and remains extended upon landing. Our optimized design yields similar support as the uniform stiff patterns (reduced) and commercial braces while providing more flexibility within the sagittal plane, which is important to allow more bending upon contact with the ground.