Fuzzy Synchronization Likelihood Graph in Deep Neural Networks for Human Motion Time Series Analysis.

Area of Science:

• Multivariate time series analysis
• Biological data analysis
• Graph-based machine learning

Background:

• Variable interactivity is key in biological time series, but graph construction methods are often costly or ignore data complexities.
• Existing methods struggle with computational expense, training needs, and handling nonlinearities and nonstationarity.

Purpose of the Study:

• To propose a novel method for constructing graphs to represent variable interactions in biological multivariate time series.
• To apply this method to automated rehabilitation exercise evaluation using human joint motion data.
• To enhance the explainability of decision-making processes in complex biological systems.

Main Methods:

• Utilized the Fuzzy Synchronization Likelihood (FSL) criterion for graph construction, focusing on qualitative similarity and variable dependency.
• Applied the FSL-constructed graphs to human joint motion data from rehabilitation exercises.
• Extended a Deep Mixture Density Neural Network (DMDN) with a convolutional layer to process FSL graphs, creating the FSL Graph-based Deep Neural Network (FSLGDN).

Main Results:

• The FSLGDN model outperformed approaches using linear correlations and human anatomy for graph construction.
• Task-based analysis of joint motion interactions proved more beneficial than anatomy-based graphs.
• The FSLGDN model captured more information from nonstationary motion data compared to linear correlation methods.

Conclusions:

• The proposed FSL graph construction offers a more intuitive and explainable approach to analyzing variable dependencies in multivariate time series.
• The FSLGDN method provides a powerful tool for automated rehabilitation exercise evaluation, offering deeper insights into joint kinematics.
• This approach enhances decision-making processes by providing a clearer representation of feature dependencies.