Classifying muscle parameters with artificial neural networks and simulated lateral pinch data

Area of Science:

• Biomechanics
• Computational modeling
• Machine learning

Background:

• Hill-type muscle models are essential for simulating human movement but have parameters difficult to measure in vivo.
• Dynamometric data contains encoded Hill-type muscle parameters, yet a generalizable estimation approach is lacking.

Purpose of the Study:

• To use musculoskeletal models and artificial neural networks (ANNs) to classify maximum isometric force, a Hill-type muscle parameter.
• To estimate this parameter from simulated lateral pinch force, a measurable dynamometric dataset.
• To compare the accuracy of feedforward and long short-term memory (LSTM) neural networks, assessing if dynamic behavior improves classification.

Main Methods:

• Generated four forward dynamics datasets with increasing complexity by adjusting muscle parameters.
• Simulated lateral pinch force to evaluate the effect of varying thumb muscle maximum isometric force.
• Employed feedforward and LSTM neural networks to classify muscle parameter groups based solely on lateral pinch force data.

Main Results:

• Both ANNs achieved >80% accuracy when classifying datasets varying only the flexor pollicis longus and/or abductor pollicis longus.
• Model accuracy decreased below 30% when including muscles with redundant functions.
• While both ANNs outperformed random guessing, the LSTM model did not consistently outperform the feedforward model.

Conclusions:

• ANNs offer a cost-effective, data-driven method for approximating Hill-type muscle-tendon parameters from accessible dynamometric data.
• Classifying muscle parameters becomes more challenging with muscles exhibiting redundant functions or minimal impact on force production.