Dynamics of functional brain connectivity in schizophrenia: machine learning models for diagnosis and prognosis
Document TypePhD thesis
Access StatusThis item is embargoed and will be available on 2021-11-07.
© 2019 Akhil Raja Karazhma Kottaram
To date, most studies of resting-state functional connectivity implicitly assume that connections remain unchanged over time. However, recent studies suggest that functional connectivity across a range of species exhibits time-varying behaviour. Understanding the time-varying properties of functional connectivity appears particularly beneficial in studying disorders such as schizophrenia. This thesis scrutinises the potential applications of dynamic functional connectivity (dFC) in aiding both the diagnosis and prognosis of schizophrenia. While previous dFC studies focussed on assessing variability of connection strengths over time, we propose a model that can simultaneously capture dynamics in both time and space. Temporal sliding windows were used to map dynamics over time. Spatial variability was accounted by a modified seed-based connectivity approach that allowed different network regions to vary their spatial layout, by expanding or contracting over time according to their connectivity profile. Connectivity measures based on the proposed method were then compared to those derived from traditional static and temporally dynamic connectivity, in predicting the diagnostic status in schizophrenia using support vector machine-based classifier models. Prediction accuracies exceeding 91% were obtained with our method, while previous methods yielded significantly lower accuracies. This suggests that the proposed method provides a better characterisation of connectivity dynamics and extracts novel disease-specific information that can potentially yield new insights into the pathophysiology of schizophrenia. Compared to healthy individuals, schizophrenia patients exhibited both temporally and spatially diminished, but more variable functional connectivity across different resting-state networks. Further, dynamic interactions among different resting-state networks were characterised using a hidden Markov model (HMM). Fluctuations in fMRI activity within 14 canonical networks, derived from both healthy individuals and schizophrenia patients, were concatenated and then quantized into 12 states using the HMM. We observed that patients spent significantly greater amounts of time in states characterised by low default-mode network (DMN) activation and heightened activity within different sensory networks. It was also found that patients lacked the ability to effectively up/downregulate the activity within the DMN. Furthermore, measures of dynamics derived from the model associated significantly with positive symptoms of schizophrenia and provided high predictive diagnostic accuracy (~85%). Finally, we examined the prognostic predictive power of dFC measures. Specifically, we tested if measures derived from dynamic connectivity among the DMN regions aid in classifying patients into worsening or improving in symptom severity after a year. Classifiers trained on DMN connectivity dynamics yielded 75-80% accuracies in predicting prognostic status in all the three types of scores considered (positive, negative and overall symptom severity). Importantly, dynamic connectivity measures were found to be better predictors than other, previously proposed variables such as cortical thickness, grey matter volume, clinical and behavioural measures and static connectivity. Together, the analyses presented in this thesis validate the utility of dynamic functional connectivity in characterising schizophrenia pathology and in aiding the adoption of more evidence-based treatment options.
Keywordsneuroimaging; functional connectivity; image processing; machine learning; schizophrenia; neuroscience
- Click on "Export Reference in RIS Format" and choose "open with... Endnote".
- Click on "Export Reference in RIS Format". Login to Refworks, go to References => Import References