Neuroimaging based predictions and subtyping in Schizophrenia Spectrum Disorders

Schizophrenia is a complex neuropsychiatric syndrome with a high internal heterogeneity (interindividual variations in neurobiological, genetic and phenotypic profile). Currently, the causes and neurobiology of schizophrenia are not fully understood, and there is a large unmet medical need, since many patients do not respond adequately to available treatments and have poor long term outcomes. Even though various mechanistically plausible biomarkers for schizophrenia have been suggested, none of these are so far used clinically. Having reliable and objective biomarkers is important to better understand the disorder, to support clinical decision and to assist the development of new treatments. Early studies show that patients with schizophrenia have widespread impairments in neural communication, which can be measured using functional magnetic resonance imaging (fMRI). Abnormal brain activation has been linked to different aspects of the disorder, but firm conclusions are yet to be made, since the results vary across studies. These variations are often attributed to the high internal heterogeneity and limited sample sizes in most studies. In the last decade, the use of data-driven methods and large multi-site datasets, have led to several advancements, which is promising for future research progress. However, overall the field is still at a stage of identifying solutions to methodological challenges, rather than developing specific biomarkers for clinical practice (yet). The goal of this PhD project was to address a part of the methodological puzzle, by exploring different ways to use machine learning in the search for robust and reproducible fMRI biomarkers. Throughout the analyses, we have used supervised machine learning to enable clinical predictions and unsupervised machine learning for feature extraction (decomposition methods) and disease subtyping. The work has been organized into four studies as summarized below. In an attempt to search for early risk prediction biomarkers, the goal of Study 1 was to classify healthy participants with schizotypal traits according to their degree of social anhedonia. We developed a classification framework with a broad selection of feature extraction methods to determine which of these could drive the classification. We found significant predictions when using both temporal and spatial network features, and achieved the highest performances, when using features from the two data-driven decomposition methods: independent component analysis (ICA) and multi-subject archetypal analysis (MSAA). Throughout our analyses, we discovered how much the final results depended on the parameters within the analysis pipeline. Thus, for the remaining studies, we focused our analysis to increase the robustness and reproducibility, e.g. by using multi-site data which had been made publicly available through data-sharing initiatives. This enabled us to train the models on a more heterogenous multi-site discovery dataset, and to test the generalizability of our findings on external data. In Study 2 the goal was to classify patients with schizophrenia using multi-site data. We adjusted the prediction framework to handle multi-site predictions and furthermore aimed to make each step as data-driven and robust as possible. For the decomposition methods, we also investigated different ways of using transfer learning to bridge feature extraction between datasets. Using spatial network features from both the decomposition methods and parcellation-based connectivity analysis we found high and reproducible classification performances that generalized to the external data. The highest performances were obtained when using ensemble decision models, which supports earlier findings that schizophrenia affects a wide range of brain networks. In Study 3, we used the same features to predict the symptom severity (measured using the Positive and Negative Syndrome Scale (PANSS)) and three PANSS subscales in an attempt to address the internal heterogeneity of schizophrenia. We used Gaussian process regression (GPR), which is a non-parametric Bayesian approach to regression that provides an uncertainty estimate for the predictions. Here, we only found moderate prediction performances, which resembled a positive trend around the mean PANSS score, and which generally did not reproduce on the external data. These findings indicate that the study could be underpowered or that the betweensite differences are too large compared to the signal of interest. Another possible explanation could be the internal consistency of the PANSS scales, or that the used datatype (resting state connectivity) or applied methods are not the right path forward. Finally, in Study 4, the goal was to search for data-driven disease subtypes using a multiple co-clustering (MCC) method that is based on Bayesian mixture models. Since the subtyping field is still at an exploratory stage, we dedicated a large part of our investigation to study the stability of the MCC method. We found that the clustering solutions were highly dependent on changes in the dataset, Nevertheless, we found subtypes with significant diagnosis association that reproduced on the external data. To conclude, we see our work providing important methodological contributions towards using machine learning and multi-site data in the search for robust and reproducible fMRI biomarkers. All our analyses were performed on fMRI data either from individuals at risk of developing psychosis (Study 1) or from patients with schizophrenia (Studies 2−4), however the developed methods can be directly used to study other clinical populations.