Novel methods for estimating homogeneous clusters and underlying processes: combining independent component analysis and cluster analysis with applications to neuroimaging data

Abstract

This dissertation develops statistical methods for identifying homogeneous subgroups of individuals based on neuroimaging data. Traditional clustering approaches in psychiatry and neurology often rely on clinical and behavioral measures alone, which neglect the rich multivariate structure of brain data and fail to capture patient heterogeneity. Recent advances in functional and structural MRI provide opportunities to uncover brain-based subtypes that may improve diagnosis, prognosis, and treatment.

The core methodology builds on Independent Component Analysis (ICA) to extract functional connectivity networks and multi-modal brain components. Novel clustering approaches are introduced that either sequentially or simultaneously combine ICA with clustering, allowing data-driven discovery of neurofunctional subtypes. Specifically, the Clusterwise ICA (C-ICA) method is proposed for resting-state fMRI data, while Clusterwise Joint-ICA extends this framework to multi-modal datasets integrating structural and functional measures. Both approaches are validated through simulations and applied to Alzheimer’s disease patients and healthy controls.

In addition, methods for model selection (MultiCHull) and open-source R software packages are developed to support applied researchers. Together, these contributions provide robust tools for uncovering brain-based subtypes in heterogeneous populations, offering new insights into neurological and psychiatric disorders and advancing the path toward precision medicine.