The large heterogeneity within psychiatric disorders such as schizophrenia and depression have for a long time hampered the generalizability and interpretation of neuroscientific results. Functional magnetic resonance imaging (fMRI) is a non-invasive brain imaging method, where brain activation can be studies either during a specific task or resting state (rs-fMRI). Many studies have proposed fMRI biomarkers for a range of diseases, but currently the use of both diagnostic and prognostic fMRI biomarkers is still seldomly used in clinic due to the poor generalizability between studies. Recently, there has been an increased focus to redefine and identify subtypes and/or symptoms of psychiatric disorders in terms of their biological systems, by using multi-site fMRI datasets (that have been made publicly available) and machine learning (ML). This now enables the field to study complex and subtle disease symptoms, such as anhedonia (reduced ability to experience pain), and to generate new data driven biomarkers that can be used both in the clinic and for drug development
During my PhD I will systematically compare different feature extraction methods to investigate fMRI brain networks, and their ability to classify both disease diagnosis and symptoms. First, I will compare the novel unsupervised decomposition method, Multi-Subject Archetypal Analysis (MSAA) to a broad range of feature extraction methods using task-based fMRI from a single site. Then I will use resting state fMRI from a large multi-site fMRI database to: i) investigate if MSAA can extract stable resting state networks using multi-site fMRI data, ii) the networks ability to classify subjects according the both disease diagnosis and symptoms, and iii) use unsupervised method to determine data-driven biological sub-groups in the data, that are not constrained to the given diagnosis or clinical rating of symptom severity.
Throughout the analysis, I aim to minimize subjectivity and bias in the data analysis (e.g. by implementing a data driven approach for preprocessing), determine ways to reduce between-site variability using harmonization methods with and without travelling subjects, and investigate different ways to translate brain network classifiers between datasets.