A graph-based approach to study muscle coordination during walking in patients affected by Parkinson’s disease

Muscle synergies are a popular tool for assessing the modular organization of the central nervous system during several motor tasks. However, extracting muscle synergies has encountered issues such as the ability to compare their results when different sEMG signal pre-processing steps are applied or when using various methods to establish the number of them. The aim of this thesis is to evaluate the motor control strategies during walking through the use of graph theory and machine learning in order to overcome the method- ological limitations inherent in the theory of muscle synergies. The method is tested to quantitatively assess changes in muscle coordination behaviour between a group of 20 healthy subjects and a group of 20 Parkinson’s Disease patients. The second category is evaluated at baseline (before surgery, T0), 3 months (T1) and 12 months (T2) after Deep Brain Stimulation surgery. Cross-correlation and spectral coherence matrices are calculated from the provided sEMG signals for each pair among the 12 trunk and lower limb muscles. These matrices are thresholded to obtain the adjacency matrices by which the networks are generated. Each node of the graph corresponds to a muscle, connected to others with a different weight. Three clustering methods (K-Means, Hierarchical and Louvain) are used to group the muscles based on distances between nodes and modularity of the graph. Around 5 clusters are found for all subjects, with a higher modularity in Controls subjects (correlation method: 0.46 ± 0.03, coherence method: 0.33 ± 0.04) than in PDs T0 subjects (correlation method: 0.45 ± 0.03, coherence method: 0.28 ± 0.07). Greater variability is observed in Parkinsonian patients across the three time points compared to healthy subjects. Moreover, clusters organization appears to be in line with the results of classical muscle synergy extraction. Through statistical tests based on multivariate analysis of variance, the Louvain method was identified as the most efficient clustering method. Furthermore, the graph method is able to differentiate information among the four subject categories behaviour (1-way MANOVA: correlation: p=0.002, coherence: p=0.017) and also between the ’Controls’ and ’PDs T0’ categories (Tukey’s HSD test: correlation: p=0.013, coherence: p=0.011). In addiction, results offer insights into the definition of new parameters that, due to their correlation with the clinical scores assigned to PDs patients, may be useful in characterizing the severity and progress of the disease from a biomechanical point of view. The proposed methods intend to open a different perspective on the modular organization of the CNS during locomotion, paving the way for future developments in applications involving also other categories of subjects, such as post-stroke or orthopedic patients, and potentially extending to involve upper limbs as well.