Subtype Identification of Parkinson’s Disease Using Sparse Canonical Correlation and Clustering Analysis of Multimodal Neuroimaging

Abstract

Parkinson’s disease (PD) is a progressive neurodegenerative disorder with heterogeneity, which indicates that there are subtypes within PD. Identification of subtypes in PD is important because it may provide a better understanding of PD and improved therapy planning. Our aim was to find and characterize the subtypes of PD using multimodal neuroimaging. We computed structural neuroimaging and structural connectivity information from 193 patients. The structural connectivity information was computed through connectivity analysis derived from tractography of diffusion tensor imaging. A three-way sparse canonical correlation analysis was applied to reduce the dimension of three modalities into three latent variables. A clustering analysis with four clusters using the resulting latent variables was conducted. We regarded each cluster as subtypes of PD and showed that each subtype had distinct patterns of correlation with important known clinical scores in PD. The clinical scores were unified Parkinson’s disease rating scale, mini-mental state examination, and standardized uptake value of putamen calculated using positron-emission tomography. The distinct correlation patterns of subtypes supported the existence of subtypes in PD and showed that the subtypes could be effectively identified by clustering a few features obtained with dimensionality reduction.