Interpretable Supervised Muscle Network Decomposition by Multifactorial ANOVA-ICA

Functional muscular connectivity reflects the underlying muscle coordination and neural control strategies during motor or postural tasks. Dimensionality reduction techniques based on multivariate linear decomposition can identify the fundamental modes of variation from muscle network instances, and enable compact visualizations and human interpretation. Existing muscle network decomposition methods do not explicitly disentangle the influence of experimental factors such as task parameters or subject groups, hindering meaningful interpretation. To address this issue, a multifactorial supervised decomposition technique based on analysis of variance (ANOVA) is introduced and combined with independent component analysis (ICA) to enhance interpretability. The resulting ANOVA-ICA provides a framework for identifying interpretable modes of systematic variation in muscle networks, allowing each mode to be explicitly associated with individual task-/subject-related factors or their combinatorial effects as modeled by ANOVA. The proposed method is tested on intermuscular coherence networks obtained through surface electromyography for postural control during standing and longitudinal running training. Multifactorial ANOVA modeling and ICA both effectively improve the interpretability of the decomposition, relative to other baseline approaches. This study demonstrates the validity of our multifactorial supervised approach to muscle network decomposition, highlighting its potential for applications in areas such as motor neurophysiology and rehabilitation.