Assessment of a multi-measure functional connectivity approach

• Autor(a) principal: Fernandes, Miguel Claudino Leão Garrett
• Data de Publicação: 2017
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10362/28548
• topic: fMRI; classification; functional connectivity matrices; SVM; feature selection; mutual information; Domínio/Área Científica::Engenharia e Tecnologia

Efforts to find differences in brain activity patterns of subjects with neurological and psychiatric disorders that could help in their diagnosis and prognosis have been increasing in recent years and promise to revolutionise clinical practice and our understanding of such illnesses in the future. Resting-state functional magnetic resonance imaging (rsfMRI) data has been increasingly used to evaluate said activity and to characterize the connectivity between distinct brain regions, commonly organized in functional connectivity (FC) matrices. Here, machine learning methods were used to assess the extent to which multiple FC matrices, each determined with a different statistical method, could change classification performance relative to when only one matrix is used, as is common practice. Used statistical methods include correlation, coherence, mutual information, transfer entropy and non-linear correlation, as implemented in the MULAN toolbox. Classification was made using random forests and support vector machine (SVM) classifiers. Besides the previously mentioned objective, this study had three other goals: to individually investigate which of these statistical methods yielded better classification performances, to confirm the importance of the blood-oxygen-level-dependent (BOLD) signal in the frequency range 0.009-0.08 Hz for FC based classifications as well as to assess the impact of feature selection in SVM classifiers. Publicly available rs-fMRI data from the Addiction Connectome Preprocessed Initiative (ACPI) and the ADHD-200 databases was used to perform classification of controls vs subjects with Attention-Deficit/Hyperactivity Disorder (ADHD). Maximum accuracy and macro-averaged f-measure values of 0.744 and 0.677 were respectively achieved in the ACPI dataset and of 0.678 and 0.648 in the ADHD-200 dataset. Results show that combining matrices could significantly improve classification accuracy and macro-averaged f-measure if feature selection is made. Also, the results of this study suggest that mutual information methods might play an important role in FC based classifications, at least when classifying subjects with ADHD.