Abstract
Many brain disorders - such as Alzheimer's disease, Parkinson's disease, schizophrenia and autism - are heterogeneous, that is, they may have several subtypes. Traditionally, clinicians have identified subtypes, such as subtypes of psychosis, using clinical criteria. Neuroimaging has the potential to detect subtypes based on objective biomarker-based criteria; however, there are no studies that evaluate the application of combining unsupervised machine learning and anatomical connectivity analysis to accomplish this goal. We propose a computational framework to detect subtypes based on anatomical connectivity computed from diffusion tensor imaging data, in a data-driven and fully automated way. The proposed method exhibits excellent performance on simulated data. We also applied this approach to a real-world dataset: the Nathan Kline Institute data set. The Nathan Kline Institute study consists of 137 normal adult subjects (mean age 41 years (standard deviation 18), male/female 85/52). We examined the association between detected subtypes and the impulsive behavior scale. We found that a subtype characterized by lower connectivity scores was associated with a higher positive urgency score; positive urgency is a vulnerability marker for drug addiction. The top-ranked connections characterizing subtypes involve several brain regions, including the anterior cingulate gyrus, median cingulate gyrus, thalamus, superior frontal gyrus (medial), middle frontal gyrus (orbital part), inferior frontal gyrus (triangular part), superior frontal gyrus, precuneus and putamen. The proposed framework is extendable, and can be used to detect subtypes from other features, including clinical and genomic biomarkers.