Predicting Neuroticism from fMRI

Week 4 Presentation

June 5th 2020

Liz Izakson

Annabelle Harvey

Five Factor Model of Personality

openness to experience inventive/curious vs. consistent/cautious

conscientiousness efficient/organized vs. extravagant/careless

extraversion outgoing/energetic vs. solitary/reserved

agreeableness friendly/compassionate vs. challenging/callous

neuroticism sensitive/nervous vs. resilient/confident

#Usually scores derived from questionnaires

Neuroticism & Neuropsychiatric Disorders

• Found to be significantly positively correlated with psychotic experiences
• Shi et al. 2018 The Relationship Between Big Five Personality Traits and Psychotic Experience in a Large Non-clinical Youth Sample: The Mediating Role of Emotion Regulation
• Neuroticism is a marker of non-specified general risk for common mental disorders
• Ormel et al. 2015 Neuroticism and Common Mental Disorders: Meaning and Utility of a Complex Relationship
• Neuropsychiatric disorders realign connectivity along common axes
• Moreau et al. 2020 The general impact of haploinsufficiency on brain connectivity underlies the pleiotropic effect of neuropsychiatric CNVs

Questions:

Are neuropsychiatric disorders extreme cases of connectivity patterns that are found in the overall population?

Can we use fMRI connectivity to predict neuroticism?

If so, do the features that distinguish neuroticism in connectivity align with those common to neuropsychiatric disorders?

Pipeline Menu for fMRI Prediction

Raw fMRI image

Dadi et al. 2019 Benchmarking functional connectome-based predictive models for resting-state fMRI

Data: Human Connectome Project

• The HCP provides MRI data and extensive behavioral assessment from almost 1,200 subjects
• Preprocessed ICA-Fix data (second session out of the 2 possible):
• 91 subjects = 400 GB, 4D data
• Worked with this data initially then realized we could switch to much smaller data that was processed further
• Parcellation_Timeseries_Netmats (PTN) data:
• 812 subjects = 10 GB, 2D data
• Parcellation according to group ICA (200 ROIs)
• Worked with 810 subjects (excluded 2 because of missing scores in Neuroticism)

​

​

Feature Selection Strategy

1. Connectivity matrix and a behavioural measure per subject
2. Correlate each edges with behavioural measure
3. Select significantly correlated edges according to a given p-value threshold
1. Split again into positively correlated and negatively correlated
4. Fit models on each group of edges (positive, negative, combined)

1

2

3

4

Image and strategy: Shen et al. 2017 Using Connectome-Based Predictive Modeling to Predict Individual Behavior From Brain Connectivity

Models

• Linear Regression
• Connectome-based Predictive Modeling (CPM)
• Use the selected edges to compute a summary score by summing the values of those edges in each matrix
• Fit a linear regression model using the summary scores as features
• Support Vector Regression (SVR)
• Training & prediction done using leave one out strategy

CPM

LR

SVR

MSE

Results: Neuroticism

SVR

Linear Regression

CPM

MSE: 36.129

MSE: 66.480

MSE: 56.267

Results: Extraversion

SVR

Linear Regression

CPM

MSE: 54.563

MSE: 42.646

MSE: 35.261

Results: Openness

SVR

Linear Regression

CPM

MSE: 35.662

​

MSE: 46.466

​

MSE: 39.090

Results: Conscientiousness

SVR

Linear Regression

CPM

MSE: 34.334

MSE: 47.083

MSE: 35.724

Results: Agreeableness

SVR

Linear Regression

CPM

MSE: 31.737

MSE: 49.498

MSE: 33.062

Connections Positively Correlated with Neuroticism

Strength of connection indicates number of times it was chosen as significantly correlated in leave one out for 810 subjects

Connections Negatively Correlated with Neuroticism

Strength of connection indicates number of times it was chosen as significantly correlated in leave one out for 810 subjects

Trying to improve the model

Different polynomial degrees (SVR)

Multi-layer Perceptron regressor

​

Prediction of personality traits

• Low prediction scores in the literature:
• Openness to experience: r =.24, lower predictions for all other (Dubois et al., 2018)
• Neuroticism: r = .32, Extraversion: r =.21 (Jiang et al., 2018)
• Inter-correlated amongst themselves; probably very subtle representation in the brain (unlike age for example)
• Subjective measures of personality
• How can we tweak the model?
• Different parcellation method
• More complex neural network
• PCA on the five-factor scores: extract 2 orthogonal superordinate components

​

Annabelle

Goals

• Work with compute canada
• Learn how to write and submit jobs for slurm and how to set up environment
• Use S3 and globus for data
• Use nilearn to load and process fMRI data
• Use sklearn and machine learning with fMRI data
• Explore and understand different parcellation methods and connectivity parametrizations
• Data visualization
• Similarity Network Fusion

Deliverables

• Jupyter notebook
• ptn_pipeline.ipynb
• Analysis script
• ptn_script.py
• Bash files to run script
• Local: run_script_locally.sh
• HPC: netmats_predictions.sh
• Documentation
• Data visualization
• data_viz.ipynb

​

Moving forward

• Get better at git!
• Use containers and open science tools
• Learn more about:
• Parcellations
• Brain dynamics
• Genetics
• GCNs
• Clustering

​

Liz

Goals

• Being part of a community
• Learn about available open data sources - to use in my own research
• Work with rfMRI data:
• Pre-processing (nilearn)
• Parcellation (different methods)
• Connectivity matrices
• Increase my knowledge in ML and DL especially in the context of neuroimaging
• Difference between different models
• Difference between prediction methods and explanatory methods

​

Things that I have learned

• Exposed to a lot of new tools!
• Really interesting discussions about science and open science in general
• Think about the questions
• Worked with git and github
• Collaborated with a team using github
• Analysis of neuroimaging pipeline
• Tried different parcellation methods
• Tried different ML models on the data
• Jupyter notebook:
• loadAndExplore_rfMRI_HCP
• Initial_MLmodel
• Contributed to PTN_pipeline
• Worked with Bash terminal

​

​

​

Thank you!

Special thanks to Desiree Lussier and Pierre Bellec!

And to all the BrainHack School organizers.

​