Can Personality be Predicted from the Structural Connectome?

2180 

Abstract Submission 

Amelie Rauland^1,2, Kyesam Jung^2,3, Simon Eickhoff^3,2, Theodore Satterthwaite⁴, Kathrin Reetz⁵, Oleksandr Popovych^2,3

¹Department of Psychology, University Hospital RWTH Aachen, Aachen, Germany, ²Institute of Neuroscience and Medicine - Brain and Behavior (INM-7), Forschungszentrum Jülich, Jülich, Germany, ³Institute for Systems Neuroscience, Medical Faculty, Heinrich-Heine University Düsseldorf, Düsseldorf, Germany, ⁴Department of Psychiatry, Perelman School of Medicine, University of Pennsylvania, Philadelphia, USA, ⁵Department of Neurology, University Hospital RWTH Aachen, Aachen, Germany

Amelie Rauland  
Department of Psychology, University Hospital RWTH Aachen|Institute of Neuroscience and Medicine - Brain and Behavior (INM-7), Forschungszentrum Jülich
Aachen, Germany|Jülich, Germany

Kyesam Jung  
Institute of Neuroscience and Medicine - Brain and Behavior (INM-7), Forschungszentrum Jülich|Institute for Systems Neuroscience, Medical Faculty, Heinrich-Heine University Düsseldorf
Jülich, Germany|Düsseldorf, Germany

Simon Eickhoff  
Institute for Systems Neuroscience, Medical Faculty, Heinrich-Heine University Düsseldorf|Institute of Neuroscience and Medicine - Brain and Behavior (INM-7), Forschungszentrum Jülich
Düsseldorf, Germany|Jülich, Germany

Theodore Satterthwaite  
Department of Psychiatry, Perelman School of Medicine, University of Pennsylvania
Philadelphia, USA

Kathrin Reetz  
Department of Neurology, University Hospital RWTH Aachen
Aachen, Germany

Oleksandr Popovych  
Institute of Neuroscience and Medicine - Brain and Behavior (INM-7), Forschungszentrum Jülich|Institute for Systems Neuroscience, Medical Faculty, Heinrich-Heine University Düsseldorf
Jülich, Germany|Düsseldorf, Germany

Personality neuroscience aims to explore the neurobiological basis of personality as it strongly affects interindividual differences of human behavior. We investigate if the structural connectome (SC) derived from diffusion weighted images can be utilized to predict the big five personality traits [1]. As previous work in the field showed mixed results (e.g. [2], [3]) and prediction from the SC itself is a comparably new field, we systematically evaluated different design choices of the SC and feature selection processes, illustrate a large spectrum of possible results and suggest a few conditions for improved predictions.

This study used pre-processed structural and diffusion data from unrelated subjects of the Human Connectome Young Adult dataset [4]. We analyzed the effect of the following different settings on prediction performance:
• 19 different cortical parcellations [5]
• Three SC weightings: Streamline count, mean diffusivity (MD), fractional anisotropy (FA)
• Four feature selections methods: Upper triangle of the SC (whole-brain), k first principal components (PC) of the SC, k SC edges most correlated with the target across subjects, regional connectivity profiles (RCP, separate rows of the SC)
• Three subject groups: Mixed males and females (n=426), only males (n=272), only females (n=272)
• Five personality traits
This leads to a total of 3,420 different pipelines.
The SC for each subject was calculated using probabilistic tractography for all parcellations and weightings. When feature selection was used, we investigated further options, i.e., different k's for PCs and correlated edges and different rows of the SC for RCPs. Ridge regression was used to make all predictions using a nested 5-fold cross-validation for hyperparameter tuning. Selection of most correlated edges and PCA fitting was performed on the training set to prevent data leakage. Predictions for all settings were repeated 100 times for random data splits. We additionally predicted cognition (CogTotalComp_AgeAdj) of individuals using the mixed sex dataset and the whole-brain features to give context to the prediction performance of the personality traits.

With only a few exceptions, the prediction results over all different settings yielded low correlations. The mean of the distribution of average prediction correlations was around zero at r≈-0.003 (Fig. 1A). Predicting cognition led to slightly higher test set correlations compared to personality traits (Fig. 1B). Despite the overall zero-centered prediction accuracies, some differences between the distinct settings could be observed: For the different feature selection methods, the highest correlations could be reached by applying the RCPs (Fig. 2A). Within the same parcellation strategy, there was a tendency towards better performance with higher granularity (Fig. 2B). Overall, there was no clear best SC weighting. When deriving maps of the average test correlations across 19 parcellations using the RCP feature selection, one can see higher similarity (correlation) in prediction results between the microstructural weightings MD and FA and distinct prediction patterns for streamline count (Fig. 2C).

By systematically evaluating many different pipelines for predicting personality traits from the SC, we find only a few cases of promising prediction performance (r>=0.2) which are similar to values reported in the literature for personality prediction from the FC ([6], [7]). Most results, however, are centered around a correlation of zero indicating no generalizable linear relationship between personality traits and SC. However, we did find some methodological differences between distinct settings and suggest considering RCPs for feature selection. The improved but still limited performance when predicting cognition indicates that the results for personality prediction might be influenced by both known limitations of the standard SC and the target itself, which requires additional investigations.

Emotion and Motivation Other

Classification and Predictive Modeling ²

Connectivity (eg. functional, effective, structural)

Diffusion MRI Modeling and Analysis

White Matter Anatomy, Fiber Pathways and Connectivity ¹

Machine Learning

Tractography

WHITE MATTER IMAGING - DTI, HARDI, DSI, ETC

Other - Personality