Machine Learning Approaches to Provide Spatio-Temporal Characterization of Human Brain Functional Activities

Recently, the interest in pattern recognition approaches to the analysis of clinical neuroimaging data has increased substantially. A crucial advantage of multivariate pattern recognition algorithms in comparison to the traditional univariate approaches is that they provide predictions on the level of individual subjects. It is this multivariate nature of pattern recognition algorithms that results in increased sensitivity over univariate methods and has led to numerous applications in clinical research. Meanwhile, advances in neuroimaging technologies have improved our understanding of brain function in psychiatric and neurological disorders such as mood disorders, drug abuse and addiction, schizophrenia, Alzheimer's disease, traumatic brain injury,-. These promising advances in functional neuroimaging technology and multivariate pattern recognition's applications in neuroimaging data analysis motivated the work presented in this dissertation. Monitoring and evaluating of human brain performance during the execution of functional experiments have revealed evidence regarding distinctive pattern of brain activity between healthy individuals and individuals with brain functional disorders. Except for certain cases, to date, the results of these studies have had minimal clinical impact and despite much interest in the use of brain scans for diagnostic and prognostic purposes, traditional and often ineffective diagnostic and prognostic approaches are the common practice for neurologists and psychiatrists.;There are a few limitations that restrict the clinical translations of identifying functional biomarkers to characterize certain brain functional disorders. Firstly, majority of the related studies are focused on group studies that attempt to signify differences between the groups of subjects and do not provide description at the individual level. Secondly, the common techniques for characterizing functional neuroimaging response at the individual level are traditional single-channel time series feature extraction techniques that do not necessarily fit into the neuroimaging multichannel time series frameworks. Finally, for the more recently developed modalities such as fNIRS very few studies have attempted to identify biomarkers in brain disorders through the data mining and machine learning approaches. Therefore, in this dissertation emphasis was placed on improving, developing, and extracting clinically adaptable neuroimaging features to enable translating the laboratory work into clinical environments. In particular, machine learning algorithms and data mining techniques were utilized to generate spatio-temporal features from the neuroimaging time series and were evaluated for diagnosis of certain brain activity disorders. The presented work in this dissertation offers novel approaches for neuroimaging feature extraction, effective dimensionality reduction, and has applications in non-invasive and early diagnosis of certain brain functional disorders.