Publication
Exploration and application of machine learning algorithms to functional connectivity data
| Summary: | Methods for the study of the functional connectivity in the brain have seen several developments over the last years, however not yet in a fully realized manner. Machine learning and complex network analysis are two promising techniques that together can help the process of better exploring functional connectivity for future clinical applications. Machine learning and pattern recognition algorithms are helpful for mining vast amounts of neural data with increasing precision of measures and also for detecting signals from an overwhelming noise component (Lemm, Blankertz, Dickhaus, & Müller, 2011). Complex network analysis, a subset of graph theory, is an approach that allows the quantitative assessment of network properties such as functional segregation, integration, resilience, and centrality (Rubinov & Sporns, 2010). These properties can be fed into classification algorithms as features. This is a new and complex approach that has no standard procedures defined, so the aim of this work is to explore the use of fMRI-derived complex network measures combined with machine learning algorithms in a clinical dataset. In order to do so, a set of classifiers is implemented on a feature dataset built with brain regional volumes and topological network measures that, in turn, were constructed based on functional connectivity data extracted from a resting-state functional MRI study. The set of classifiers includes the nearest neighbor, support vector machine, linear discriminant analysis and decision tree methods. A set of feature selection methods was also implemented before the classification tasks. Every possible combination of feature selection methods and classifiers was implemented and the performance was evaluated by a cross-validation procedure. Although the results achieved weren’t exceptionally good, the present work generated knowledge on how to implement this recent approach and allowed the conclusion that, for most cases, feature selection improves the performance of the classifier. The results also showed that the decision tree algorithm produces relatively good results without being associated with a feature selection method and that the SVM classifier, together with RFE feature selection method, produced results on the same level as other work done with a similar approach. |
|---|---|
| Main Authors: | Veloso, Telma Alves |
| Subject: | Engenharia e Tecnologia::Engenharia Médica |
| Year: | 2014 |
| Country: | Portugal |
| Document type: | master thesis |
| Access type: | open access |
| Associated institution: | Universidade do Minho |
| Language: | English |
| Origin: | RepositóriUM - Universidade do Minho |
| Summary: | Methods for the study of the functional connectivity in the brain have seen several developments over the last years, however not yet in a fully realized manner. Machine learning and complex network analysis are two promising techniques that together can help the process of better exploring functional connectivity for future clinical applications. Machine learning and pattern recognition algorithms are helpful for mining vast amounts of neural data with increasing precision of measures and also for detecting signals from an overwhelming noise component (Lemm, Blankertz, Dickhaus, & Müller, 2011). Complex network analysis, a subset of graph theory, is an approach that allows the quantitative assessment of network properties such as functional segregation, integration, resilience, and centrality (Rubinov & Sporns, 2010). These properties can be fed into classification algorithms as features. This is a new and complex approach that has no standard procedures defined, so the aim of this work is to explore the use of fMRI-derived complex network measures combined with machine learning algorithms in a clinical dataset. In order to do so, a set of classifiers is implemented on a feature dataset built with brain regional volumes and topological network measures that, in turn, were constructed based on functional connectivity data extracted from a resting-state functional MRI study. The set of classifiers includes the nearest neighbor, support vector machine, linear discriminant analysis and decision tree methods. A set of feature selection methods was also implemented before the classification tasks. Every possible combination of feature selection methods and classifiers was implemented and the performance was evaluated by a cross-validation procedure. Although the results achieved weren’t exceptionally good, the present work generated knowledge on how to implement this recent approach and allowed the conclusion that, for most cases, feature selection improves the performance of the classifier. The results also showed that the decision tree algorithm produces relatively good results without being associated with a feature selection method and that the SVM classifier, together with RFE feature selection method, produced results on the same level as other work done with a similar approach. |
|---|