Benchmarking neural network personalized musculoskeletal hand models against current personalization standards using experimental magnetic resonance imaging and fine-wire electromyography.

Physics-based musculoskeletal models are often used to understand complex systems, like the human hand. Because generic hand models typically represent less than 1% of the population, models are often personalized to match individual people. Multiple personalization approaches exist. Yet, it is unknown which methods are best for studying hand tasks. Here we i) apply five musculoskeletal personalization methods to hand models and ii) evaluate the methods by comparing simulated muscle activations to measured electromyography. The five personalization methods [scaling, joint moment optimization using Neuromusculoskeletal Modeling Pipeline (NMSM), magnetic resonance imaging (MRI) segmentation, MRI segmentation combined with joint moment optimization, and a novel neural network] were used to create 14 personalized models for each participant (n=13 healthy adults). The models were then used to simulate seven range of motion and eight isometric hand tasks in the inverse direction using static optimization. In contrast to current personalization methods that require extensive experimental data and processing, we demonstrated musculoskeletal hand models could be personalized using only lateral pinch force data and a custom neural network. None of the personalization methods resulted in anatomically accurate muscle parameter sets, when compared to the anatomical standard of MRI segmentation. However, the neural network method led to significantly lower errors when predicting muscle activations compared to current state-of-the-art methods for personalizing musculoskeletal model parameters without personalizing associated neural control models.. By reducing the experimental data and time barriers required to create personalized hand models, neural networks may allow for wider use of musculoskeletal models in clinical settings.