Centre for Youth Mental Health - Theses
Permanent URI for this collection
Search Results
1 - 1 of 1
-
ItemNormative modelling in large-scale multi-site neuroimaging data sets to investigate brain abnormalities in depression( 2023-09)The search for diagnostic biomarkers using cortical thickness alterations in depression has been impeded by a combination factors, including small sample sizes underpowered to detect small effect sizes in cortical thickness differences, clinical and pathogenic between-subject heterogeneity, group average comparisons and brain ageing effects. Sample sizes in clinical neuroimaging can be increased by creating data sets by pooling. However, pooling studies across acquisition sites can create site effects, which stem from differences in during image acquisition and pre-processing. Site effects can induce biases into the estimates of cortical thickness measures and can interact with the effects of additional variables on cortical thickness, which can make their removal difficult. The aim of this thesis is to 1) Provide an educational review of retrospective site effect correction methods, their benefits, drawbacks and use cases, including normative modelling 2) Develop and test a normative model based on the covariates age, sex and site that allows to correct for site effects on cortical thickness measures in a pooled, public normative modelling data set 3) apply the best performing normative model to cortical thickness data of a large pooled depression neuroimaging data set, in order to compare z-score deviations of depressed individuals to those of healthy controls and link those deviations to clinical characteristics. Regarding aim 1) I summarise several retrospective site effect correction methods that have been published. The evaluation of the statistical foundation of each method reveals that each method has different used cases, advantages and disadvantages that the user should be aware of when choosing a method. To address aim 2), linear and non-linear versions of a normative model based on Hierarchical Bayesian Regression were developed and tested against alternative common site-effect correction methods. All models were evaluated based on their interference with making predictions from cortical thickness measures in a test set containing 35 cortical measures (34 bilateral regions and one whole brain average). ComBat 1–3, regressing out site and predictions from raw data led to a shrinkage of variance when predicting cortical thickness measures from the test set, which suggests the removal of both site variance and shared co-variation with other variables, such as age and sex. Normative modelling, in contrast, was able to retain a larger spectrum of variation. 3) Finally, the non-linear version of the normative model was applied to a large, pooled neuroimaging data set in that contained the same 35 cortical thickness measures of 5300 healthy individuals (training set: n = 3181, test set: n = 2119) and 3645 individuals with depression. The results show large between subject variability in cortical thickness alterations within depression and a large overlap of alterations with healthy controls. This thesis highlights the large between -subject variability in brain measures in clinical cohorts, that may be partially due to site effects in pooled neuroimaging studies. The findings of this thesis stress the need for methods and models in clinical neuroimaging that allow for individualised predictions and for site effect correction, stepping beyond the average patient. Last, the large overlap in the distribution of cortical thickness measures between individuals with depression and healthy controls suggests that cortical thickness might not be a suitable marker for the diagnosis of depression.