Weak and Unstable Prediction of Personality from the Structural Connectome

Personality neuroscience aims to discover links between personality traits and features of the brain. Previous neuroimaging studies have investigated the connection between the brain structure, microstructural properties of brain tissue or the functional connectivity (FC) and these personality traits. Analyses relating personality to diffusion weighted MRI measures were limited to investigating the voxel-wise or tract-wise association of microstructural properties with trait scores. The main goal of our study was to determine whether there is an individual predictive relationship between the structural connectome (SC) and the big five personality traits. To that end we expanded past work in two ways: First, by focusing on the entire structural connectome (SC) instead of separate voxels and tracts; and second, by predicting personality trait scores instead of performing a statistical correlation analysis to assess an out-of-sample performance. Prediction of personality from the SC is however not yet as established as prediction of behavior from the FC, and sparse studies in this field so far delivered rather heterogeneous results. We therefore further dedicated our study to investigate if and how different pipeline settings influence prediction performance. In a sample of 426 unrelated subjects with high quality MRI acquisitions from the Human Connectome Project, we analyzed 19 different brain parcellations, three SC weightings, three groups of subjects and four feature classes for the prediction of the five personality traits using a ridge regression. From the large number of evaluated pipelines, only very few lead to promising results of prediction accuracy r > 0.2, while the vast majority lead to a small prediction accuracy centered around zero. A markedly better prediction was observed for a cognition target confirming the chosen methods for SC calculation and prediction and indicating limitations of the personality trait scores and their relation to the SC. We therefore report, that for methods evaluated here, the SC cannot predict personality trait scores. Overall, we found that all considered pipeline conditions influence the predictive performance of both cognition and personality trait scores. The strongest differences were found for the trait openness and the SC weighting by number of streamlines which outperformed the other traits and weightings respectively. As there is a substantial variation in prediction accuracy across pipelines even for the same subjects and the same target, these findings highlight the crucial importance of pipeline settings for predicting individual traits from the SC.