Benchmarking Neural Network Personalized Musculoskeletal Hand Models Against Current Personalization Standards Using...

Area of Science:

• Biomechanics
• Computational modeling
• Human motion analysis

Background:

• Physics-based musculoskeletal models are crucial for understanding complex systems like the human hand.
• Generic hand models have limited applicability, necessitating personalization for individual anatomy.
• Existing personalization methods vary in complexity and data requirements, with unclear optimal choices for hand tasks.

Purpose of the Study:

• To apply and evaluate five distinct musculoskeletal personalization methods for hand models.
• To compare the accuracy of simulated muscle activations against measured electromyography (EMG).
• To identify the most effective personalization method for simulating hand tasks.

Main Methods:

• Five personalization methods were applied: scaling, joint moment optimization (NMSM), MRI segmentation, combined MRI segmentation and optimization, and a novel neural network.
• 14 personalized models were created per participant (n=13 healthy adults).
• Models simulated 7 range-of-motion and 8 isometric hand tasks using inverse dynamics and static optimization.

Main Results:

• Personalized models were successfully created using only lateral pinch force data and a neural network, reducing experimental demands.
• No method yielded perfectly anatomically accurate muscle parameters compared to MRI segmentation.
• The neural network method demonstrated significantly lower muscle activation prediction errors than state-of-the-art techniques.

Conclusions:

• A novel neural network approach enables effective personalization of musculoskeletal hand models with reduced data requirements.
• This method offers a promising alternative to current state-of-the-art personalization techniques.
• The reduced barriers may facilitate broader adoption of personalized musculoskeletal models in clinical applications.