Prediction of disease classes using resting state neuroimaging data: application of new statistical learning tools.
Cor Ninaber started his PhD with the Methodology and Statistics group just a couple of weeks ago. Cor Ninaber studied the psychology master specialization of Methodology and Statistics at Leiden University. Ninaber focused his master thesis on the field of multinomial classification, and completed an internship at TNO Quality of Life.
- The Methodology and Statistics Lab
- Can we use resting state fMRI in order to diagnose a single patient?
- The classification area
- Why do we need new methods for classification?
During his internship, Ninaber discovered an interest in the combination of programming and statistics. Therefore, he considered starting a PhD on the prediction of disease classes using resting state neuroimaging data through the application of new statistical learning tools a great oppurtunity. And, the combination between cutting edge statistical techniques and the practical appliances these methods could have, made it an easy choice for him.
Ninaber will work in the Methodology and Statistics (M&T) group under the supervision of Willem Heiser, Serge Rombouts and Mark de Rooij. Willem Heiser’s main research interests are multidimensional scaling and unfolding, cluster analysis and classification. The research of Mark de Rooij focuses on classification and longitudinal data analysis. Serge Rombouts’ interests lie in several aspects of fMRI methodology.
Resting state functional magnetic resonance imaging (RS-FMRI), studying the brain during rest, has become a very popular way to study functional connectivity in the brain. It appears that the brain is very active even in the absence of explicit input or output behaviour. And by comparing levels of activity among different groups of brain cells, researchers can determine which areas are communicating with one another (i.e., functional connectivity). Recent functional connectivity analyses have shown that certain brain networks are especially active when the brain is at rest, and these have been termed 'resting state networks'.
The networks obtained in rest resemble networks that are typically observed activated during cognitive, sensory or motor tasks, and provide further insight into the intrinsic functional architecture of the brain. Furthermore, functional connectivity measures have improved our understanding of variability of behavior and associated brain activity. And RS-FMRI has provided insight in alterations in brain activity in healthy aging, sleep, brain development, dementia, depression, ADHD, autism, schizophrenia, Parkinson’s disease, and MS. These opportunities explain why, in just a few years time, resting state brain connectivity studies have grown from a limited niche area to a large and popular field of research.
Most investigations are limited to studying whether brain signals differ between patient and control groups. These studies provide important new insights about average (group mean) functional brain connectivity changes in diseases. However, to understand to what extent this innovative technique can be applied for (early) diagnostics and treatment predictions, it is of great interest to study whether we can classify a subject based on his/her RS-FMRI scans and see whether RS-FMRI scans of a single subject allow to determine whether a subject has Alzheimer’s disease, a frontotemporal lobe dementia, a depression, schizophrenia, ADHD, MS, or is possibly healthy.
Classification entails that we like to predict the class of a patient. Suppose there are brain scans of n subjects, which are known to come from different groups, for example, healthy subjects, Alzheimer patients, patients with schizophrenia, ADHD. The question is then whether we can distinguish these groups on the basis of the brain scans, and whether we can accurately predict the status of a single subject based on earlier obtained rules. And during his PhD, Ninaber will investigate the application new statistical learning tools with which he will try to predict of disease classes using resting state neuroimaging data.
The implication of Ninabers study are twofold. First, the new classification techniques are developed that can deal with the enormous amount of data. Traditional methods require that the number of subjects is larger than the number of variables. In this case the variables are the measurements on every voxel in each volume. The number of subjects can never exceed this number. This requires new statistical methodology. Second, new insights in the differences of brain functioning between patient groups will be obtained. These results might be valuable for early diagnosis of diseases and/or differentiation between diseases.