Neural Control Model Personalizationusing an Improved Synergy Optimization Framework

Abstract Muscle synergy models reduce the underdetermined muscle activation search by representingactivations with a small number of time-varying synergy commands and muscle-specificsynergy vector weights. In the original Synergy Optimization (SynO) method, both variableblocks were solved simultaneously using nonlinear optimization, resulting in a computationallydemanding neural control personalization problem. This study presents an updated implementation,termed SynOx, that alternates between nonnegative least-squares updates of synergycommand nodes and synergy vectors while retaining activation bounds, joint moment tracking,and squared-activation minimization. In the benchmark runs, SynOx was over 100 times fasterthan the original SynO implementation while maintaining joint moment matching quality: fitsusing 2 through 6 synergies stopped after 3 to 6 alternating iterations, required 3.11 to 17.56 s,and reached joint moment qualities from 91.90% to 99.68%. Faster neural control personalizationfacilitates construction of personalized neuromusculoskeletal models for neurorehabilitationtreatment design and optimization.