Effective treatment of affective disorders is often challenged by incorrect or delayed diagnosis. Patients with bipolar disorder (BD) typically receive between 1 and 4 prior diagnoses of other mental health disorders preceding correct diagnosis, commonly with a delay from illness onset of 10 years, highlighting the difficulty of accurate diagnosis in a clinical setting. Most commonly, bipolar disorder is misdiagnosed as unipolar major depressive disorder (MDD), with between 25 and 50\% of people diagnosed with depression being later re-classified as bipolar. The current problem of misdiagnosis is compounded by the lack of laboratory based quantitative testing for affective disorders available in clinical situations.
The overarching objectives of this thesis are to demonstrate the utility of machine learning algorithms to neuroimaging data in order to provide a computational platform for classification, diagnosis and prediction of mood disorders. The thesis hence aims to unify three main themes - namely the clinical psychiatry of affective (mood) disorders, analytic principles of neuroimaging data (such as network-based platforms) and machine learning.
I first combine network theory and machine learning to address the issue of diagnostic classification using resting state functional magnetic resonance imaging (fMRI). I find that key metrics describing functional organization of the brain can be harnessed to yield accurate diagnostic boundaries in the data set I examine. In particular, the organization of ongoing brain activity into communities provides key information which distinguishes MDD from healthy controls.
I next investigate a core group of brain regions which modulate network efficiency at a global scale, known as the rich club. Although this has previously been examined in structural brain networks, I apply it to resting state functional images. Furthermore, I investigate the robustness of the functional rich club following the correction of systematic artifacts in fMRI scans. Here I conclude a densely connected rich club is present in resting state fMRI across using all processing styles. However, the composition of members changes depending on key choices in the preprocessing of data. Further validation of the rich club should assess the impact of using different scanners or brain atlases.
From chapter 4 onward, I shift the focus from depression to bipolar disorder. Here I start by assessing the functional connectivity of the inferior frontal gyrus, a region which is down regulated during an emotional recognition task in first degree relatives of bipolar subjects. Using a technique capable of identifying subtle alterations in functional networks I identified a community of connected brain regions with substantially reduced functional connectivity in bipolar subjects. I used machine learning to distinguish people with bipolar from first degree relatives and a healthy cohort. This was most effective between bipolar subjects and control subjects, while identifying the first degree relatives from either of the other groups was substantially less successful. Interestingly, the most salient information, selected as features in the classification overlapped strongly with the regions identified using the classic between group contrast.
Following this I broaden the analysis of bipolar disorder to a whole brain analysis and move from functional to structural connectivity. Using probabilistic tractography derived from diffusion weighted images, I first identify the structural rich club. The rich club was found to be more densely connected in first degree relatives, with no substantial differences between either bipolar or control subjects. The rich club is then used to train classification algorithms, allowing me to assess the applicability of this structural backbone in providing salient features which may be used in the identification of bipolar disorder. The next logical step would be to combine functional and structural information in the same subjects in an effort to improve the predictive power of the classifiers.
In sum, I have assessed the applicability of network measures of brain organization towards quantitative classification of affective disorders. Findings described throughout this thesis add to the growing body of knowledge surrounding subtle neurobiological differences in affective disorders. This study lays the foundation for an approach which would ideally be extended in future studies to larger cohorts, multiple disorders and cross-imaging platform translation.