Learning Representations of Individual Di erences Using Low-Rank Approximations to Normative Models
Keywords
No Thumbnail Available
Authors
Issue Date
2020-09-14
Language
en
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Early detection of psychiatric disorders can help the prevention of symptoms and the detection
of relevant biomarkers. Therefore, it is important to study neurodevelopment and its relation
to symptoms of psychiatric disorders. One major problem in the study of psychiatric disorders
is the heterogeneity of the clinical groups, meaning that the symptoms and underlying
mechanism of the disorder can di er per individual. Normative modeling has been introduced
as a method to deal with heterogeneity. It is a Bayesian method developed to model individual
di erences with respect to a normal range. From the model, normative probability maps
(NPM), which contain the individual deviance scores per voxel from the normative modeling,
can be calculated. The NPMs can be used to study abnormalities in the brain, however,
interpretation of these maps is a challenging task. To study the developing brain, normative
modeling is applied to the n-back task fMRI from The Philadelphia Neurodevelopmental Cohort,
while taking gender and age into account as covariates. To assist the interpretation of the
NPMs, independent component analysis (ICA) and dictionary learning (DL), two decomposition
techniques with di erent objective functions, are applied to the NPMs. These techniques
provide a way to summarize interesting deviations in brain activation into a small number of
components. Lastly, to study the accuracy of these techniques, the ICA and DL components
are correlated to the behavioral data of the participants. Relevant components are assumed
to have stronger correlations to the behavioral data. The results of the normative modeling
show clear neurodevelopmental e ects in the brain areas associated with working memory. No
clear di erences were found between the performance of ICA and DL, indicating no in
uence
of the di erent objective functions for this problem. Lastly, no strong correlation between the
components and the clinical data were found.
Description
Citation
Faculty
Faculteit der Sociale Wetenschappen