The high-dimensional nature of resting state functional MRI (fMRI) data implies the need of suitable feature selection techniques. Traditional univariate techniques are fast and straightforward to interpret, but are unable to unveil relationships among multiple features. The aim of this work is to evaluate the applicability of clustering based techniques to the problem of feature extraction in resting state fMRI data analysis. More specifically, we devise a methodology based on consensus clustering, a particular approach to the clustering problem that consists in combining different partitions of the same data set in a final solution. Our approach was validated on a real-word data set, deriving from multiple clinical studies on Parkinson’s disease and amyotrophic lateral sclerosis. Our results show that the adoption of consensus-based techniques can indeed lead to an improvement of the results, not only in terms of feature discriminability, but also from the point of view of interpretability.
- Consensus clustering
- Default-mode network
- Feature extraction
- Independent component analysis
- Resting state fMRI