Understanding Population Variation in Brain Structure at the Individual Level

No Thumbnail Available
Issue Date
Journal Title
Journal ISSN
Volume Title
Many brain disorders are characterized by heterogeneity in terms of symptoms and their underlying biology, which has often been neglected in previous studies on brain disorders. Alternatively, an approach that takes into account this heterogeneity is by means of nor- mative modeling, which is a Bayesian method used to estimate a biological variable based on clinical covariates, aiming to model the variation in a large group of healthy controls. The method demands large-scale datasets to model this variation, but this comes with two new requirements. Namely, the normative model should be able to computationally scale to this large amount of data, and handle batch e ects coming from di erent medical imaging centers. We used the log-transformed Jacobian maps of T1-weighted structural MRI images along with corresponding age and gender information of 1114 subjects to t and evaluate six hierarchical Bayesian regression models and a Gaussian process regression model using random feature approximations. The hierarchical Bayesian regression models are linear and have either xed or varying intercepts, slopes and noise terms across batches. Furthermore, there is one hierarchical Bayesian regression model that ts a second-order polynomial curve. We solved the problem of computational scalability and our models were able to handle site e ects, as the models tted well to the healthy controls. Furthermore, the performance of each model in a clinical setting was measured by using a dataset of 199 subjects, consisting of healthy controls (60 subjects) and patients diagnosed with schizophrenia (49 subjects), bipolar disorder (49 subjects) or attention de cit hyperactivity disorder (41 subjects). Ex- treme values within the tted distribution of healthy controls were predicted to be patients. The di erences in tting performance between the models were subtle. However, allow- ing for varying noise terms across batches improved the detection performance and when a model was given too much exibility, it showed signs of over tting.
Faculteit der Sociale Wetenschappen