Modelling and Mining Brain Network Data for Diagnosis of Neurodegenerative Diseases

Connectomes are brain networks represented as a graph with the vertices being the regions of the brain and weighted edges representing strength of connections between the regions inferred from brain imaging techniques such a Functional MRI (fMRI). An intense research activity is to use the connectomes to identify markers for brain disorders, especially neuro-degenerative diseases such as Autism Spectrum Disorder (ASD) by studying the differences in the connectomes of healthy subjects and patients. This paper presents a novel data model for the connectome data and analyzes its efficacy in the classification of ASD and Typically Developing (TD) (healthy) connectomes. The proposed data modelling begins by clustering the vertices (brain regions) using the Graph Spectral Clustering into fixed number of clusters, the number of clusters chosen as four based on the empirical evidence. The resulting clustering is used to map the vertices into a binary matrix which is then converted into a binary row vector to form a vector space model of the connectome data that is employed to classify the connectome data. The developed model is first validated using Human Connectome Protocol (HCP) FMRI derived connectome data of 812 healthy patents. Binary data models of the UCLA Autism dataset with fMRI and DTI scans of 42 ASD and 37 TD subjects are generated and employed for their classification. Different classification algorithms are trained, tested and their performance evaluated using the resulting dataset. Cross Validation (CV) estimates identified the best performance (83% recall and 83% precision) for DTI data and (73% recall and 89% precision) for the fMRI data achieved using Logistic Regression.