Laura Freire Lyra: Comparison of different UK Biobank data modalities for prediction of mental health phenotypes

BCCN Berlin / Technische Universität Berlin

Abstract

The triumph of the entry of neuro-imaging into the big-data domain has been partially suppressed by recent results questioning whether these techniques can extract significant and replicable brain-phenotype associations. In this study, we investigate if neuro-imaging improves phenotype prediction compared to other cheaper and widely available biological samples at increasing sample sizes. We compare T1 and diffusion-weighted imaging-derived phenotypes (IDPs) with blood count and biochemistry, retinal features, and omics data from the UK Biobank. As representatives of mental health targets, we focus our analysis on major depressive disorder, alcohol intake, and fluid intelligence. Our results show that the best-performing feature modality for mental health prediction depends on the target and sample size. Brain imaging has little to no advantage over other modalities in most prediction tasks. Particularly in major depressive disorder prediction, there are no differences between aggregated results from proteomics, metabolomics, and biochemistry and aggregated brain imaging IDPs results.

 

 

Additional Information

Master Thesis Defense

 

Organized by

Prof. Dr. rer. nat. Kestin Ritter (Charité)   & Prof. Dr. med. Dr. phil. Henrik Walter   / Lisa Velenosi