The Structural Connectome to Select Functional Connectivity Features for Prediction of Example Phenotypes

This preregistration documents a methodological evaluation of feature-selection strategies for functional connectivity-based prediction of alcohol consumption (AUDIT-C score) and, as additional benchmarks, age and fluid intelligence using UK Biobank resting-state fMRI data. We test whether biologically informed feature selection can reduce the dimensionality of functional connectivity (FC) features while preserving predictive performance, which would improve interpretability of the input, e.g. more reliable feature relevance maps, while, at the same time, reducing the computational load. We compare a structural-connectivity-guided selector (SC), derived from an external diffusion MRI connectome, against a connectivity-based selector derived from the functional input itself, and random selectors as a baseline. To this end, we identified 17,152 UK Biobank participants with data on alcohol consumption, fluid intelligence, and RS-fMRI, and divided this set into an exploration set (N = 13,686) and a validation hold-out set (N = 3,466). Using the exploration set, we optimised the processing pipeline and evaluated prediction performance across a range of feature selection percentages (FSPs). Performance metrics were Pearson’s correlation (r) between predicted and observed target values and the area under the r vs FSP curve (AUC). Here, we are reporting the results from the exploration set analyses and preregister confirmatory testing in the untouched validation set, in particular, that SC-based feature selection can be used to reduce the number of input features substantially without a significant drop in prediction performance. Acknowledgement: This project is co-funded by the European Union and co-financed from tax revenues on the basis of the budget adopted by the Saxon State Parliament. This research was supported by the Deutsche Forschungsgemeinschaft (DFG) grants 402170461 (TRR 265), 454245598 (GRK 2773/1), and 521379614 (TRR 393). Data was provided in part by the Human Connectome Project, WU-Minn Consortium (Principal Investigators: David Van Essen and Kamil Ugurbil; 1U54MH091657) funded by the 16 NIH Institutes and Centers that support the NIH Blueprint for Neuroscience Research; and by the McDonnell Center for Systems Neuroscience at Washington University. This research has been conducted using the UK Biobank Resource under Application Number 43880. GJ, MB, MM, FB, SE, MNS, received funding from the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation), grant number GRK 2773/1 – 454245598."