Sensitivity analysis guided improvement of an electromyogram-driven lumped parameter musculoskeletal hand model

EMG-driven neuromusculoskeletal models have been used to study many impairments and hold great potential to facilitate human-machine interactions for rehabilitation. A challenge to successful clinical application is the need to optimize the model parameters to produce accurate kinematic predictions. In order to identify the key parameters, we used Monte-Carlo simulations to evaluate the sensitivities of wrist and metacarpophalangeal (MCP) flexion/extension prediction accuracies for an EMG-driven, lumped-parameter musculoskeletal model. Four muscles were modeled with 22 total optimizable parameters. Model predictions from EMG were compared with measured joint angles from 11 able-bodied subjects. While sensitivities varied by muscle, we determined muscle moment arms, maximum isometric force, and tendon slack length were highly influential, while passive stiffness and optimal fiber length were less influential. Removing the two least influential parameters from each muscle reduced the optimization search space from 22 to 14 parameters without significantly impacting prediction correlation (wrist: 0.90 ± 0.05 vs 0.90 ± 0.05, p = 0.96; MCP: 0.74 ± 0.20 vs 0.70 ± 0.23, p = 0.51) and normalized root mean square error (wrist: 0.18 ± 0.03 vs 0.19 ± 0.03, p = 0.16; MCP: 0.18 ± 0.06 vs 0.19 ± 0.06, p = 0.60). Additionally, we showed that wrist kinematic predictions were insensitive to parameters of the modeled MCP muscles. This allowed us to develop a novel optimization strategy that more reliably identified the optimal set of parameters for each subject (27.3 ± 19.5%) compared to the baseline optimization strategy (6.4 ± 8.1%; p = 0.004). This study demonstrated how sensitivity analyses can be used to guide model refinement and inform novel and improved optimization strategies, facilitating implementation of musculoskeletal models for clinical applications.