Can predictive simulations provide insights for personalizing assistive wearable device design?

1. Problem Statement

Optimizing assistive wearable devices (e.g., exosuits) is critical for their efficacy and user adoption, particularly for personalization. Current state-of-the-art methods face significant limitations:

• Human-in-the-Loop Optimization (HILO): While effective, it is time-consuming, physically demanding for users, and limited to optimizing control parameters rather than mechanical design.
• Inverse Dynamics (Tracking Approaches): These rely on pre-recorded experimental kinematics. They assume human kinematics remain invariant after device attachment, which is often false as humans adapt to external assistance, leading to inaccurate torque and force estimations.
• Predictive Simulations: While promising for exploring design spaces without experimental data, existing approaches often lack systematic optimization frameworks and rigorous validation against real-world data. Furthermore, there is a "simulation-to-reality" gap where simulations may overestimate benefits or fail to capture individual variability.

Core Research Question: Can predictive simulations, integrated into a systematic optimization framework, reliably predict the trends of metabolic cost and muscle activation to guide the personalized design of assistive devices, even if absolute biomechanical values are not perfectly accurate?

2. Methodology

The authors developed and validated a Design Optimization Platform that utilizes a two-level optimization structure to personalize the Biarticular Thigh Exosuit (BATEX), a passive soft exosuit assisting the hip and knee via elastic elements parallel to the Rectus Femoris (RF) and Hamstrings (HAM).

A. Data and Setup

• Dataset: Used a publicly available dataset from 8 healthy male participants walking at 1.3 m/s with the BATEX.
• Conditions: Tested 9 configurations (combinations of 0, 650, and 1100 N/m stiffness for RF and HAM springs) plus a "No Spring" (NS) baseline.
• Measurements: Kinematics (23 markers), Ground Reaction Forces (GRF), EMG (6 muscles: RF, HAM, Vasti, Gluteus Maximus, Soleus, Gastrocnemius), and Metabolic Cost.

B. Predictive Simulation Framework

• Software: SCONE using the Hyfydy engine for forward dynamics.
• Model: A 2D musculoskeletal model (9 DOF, 18 Hill-type muscles) scaled to individual anthropometry. The BATEX mass was distributed as point masses.
• Control: A reflex-based neuromuscular controller with 53 optimizable parameters.
• Optimization Loop:
  • Inner Loop (Simulation): Solves an optimal control problem to generate gait for a specific device configuration. The cost function minimizes metabolic cost, deviation from nominal velocity, and penalties for unrealistic kinematics/kinetics. Crucially, it includes terms to match experimental kinematics and muscle activations for the baseline (NS) condition to personalize the controller.
  • Outer Loop (Design Optimization): Uses a Bayesian optimizer to search the design space (stiffness values $k_{RF}$ and $k_{HAM}$) to minimize the normalized metabolic cost.

C. Validation Strategy

• Baseline Validation: Compared simulated vs. experimental kinematics and muscle activations for the unassisted (NS) condition.
• Assisted Validation: Compared simulated vs. experimental trends in normalized metabolic cost across different stiffness configurations.
• Correlation Analysis: Investigated which baseline simulation accuracies (kinematics/muscles) correlated most strongly with the accuracy of predicting metabolic cost trends in assisted conditions.

3. Key Contributions

1. Development of a Two-Level Optimization Platform: A systematic framework integrating predictive simulations with Bayesian optimization to navigate high-dimensional design spaces for personalized assistive devices.
2. Rigorous Validation: Validated the platform against a comprehensive public dataset (BATEX), assessing not just absolute values but the reliability of trends across different users.
3. Identification of Critical Predictors: Discovered that the model's ability to predict Vasti (VAS) muscle activation in the unassisted baseline is the strongest statistical predictor of its success in predicting metabolic cost trends in assisted conditions.
4. Shift in Design Philosophy: Demonstrated that for design optimization, accurate prediction of performance trends (relative changes) is more critical than perfect biomechanical accuracy (absolute values).

4. Key Results

• Kinematic Accuracy: The model showed high accuracy for hip, knee, and ankle angles (Correlation $\rho > 0.8$, low NRMSE). However, pelvis tilt prediction was poor ($\rho$ as low as 0.18), likely due to the 2D model's simplification of the upper body.
• Muscle Activation:
  • High Accuracy: Vasti (VAS), Soleus (SOL), and Gastrocnemius (GAS) were well-predicted.
  • Low Accuracy: Proximal biarticular muscles (RF, HAM) and Gluteus Maximus (GM) showed high inter-subject variability and lower correlation, particularly in assisted conditions.
• Metabolic Cost Trends:
  • The model successfully captured the general trend of metabolic cost changes across the design space (Average $\rho = 0.56$).
  • Inter-subject Variability: Some participants (e.g., P.5, $\rho=0.78$) showed strong agreement, while others (e.g., P.3, $\rho=-0.28$) showed poor trend prediction.
  • Absolute Values: The model tended to overestimate absolute metabolic costs, but this did not prevent it from identifying beneficial design regions.
• Regression Analysis: A linear regression model identified VAS activation accuracy in the baseline condition as the only statistically significant predictor ($p=0.007$) of metabolic cost trend accuracy in assisted conditions.
• Optimization Outcomes:
  • For participants with high trend prediction accuracy (e.g., P.5), the platform identified optimal stiffness configurations close to experimental bests (e.g., predicted cost 0.91 vs. experimental 0.95).
  • For participants with poor trend prediction (e.g., P.3), the optimizer identified a "false" optimal region (e.g., high RF stiffness) that contradicted experimental findings, highlighting the dependency of the platform on model fidelity.

5. Significance and Conclusion

• Trend over Precision: The study establishes that for assistive device design, a simulation does not need to perfectly replicate absolute biomechanical values to be useful. It must, however, reliably predict the direction and magnitude of performance trends (e.g., which stiffness reduces metabolic cost).
• Personalization Indicator: The accuracy of simulating VAS muscle activation in unassisted walking serves as a "litmus test" for whether a specific user's model is reliable enough for personalized optimization.
• Future Directions: The authors propose that future iterations should incorporate machine learning-enhanced neuromuscular models or muscle synergy approaches to improve the prediction of proximal muscles (RF, HAM, GM) and pelvis kinematics, thereby expanding the platform's reliability to a broader user population.
• Practical Impact: This work validates the feasibility of using "Model-in-the-Loop Optimization" (MILO) to reduce the reliance on extensive, costly human-in-the-loop experiments, paving the way for more efficient, targeted, and personalized development of soft wearable exoskeletons.