Kathy Garcia
SfN Global Connectome 2021 · Poster

Predicting Dimensional Symptoms of Psychopathology from Task-Based fMRI using Support Vector Regression

Kathy Garcia1 Zach Anderson1 Iris Ka-Yi Chat2 Katherine S.F. Damme1 Katherine Young3 Susan Y. Bookheimer4 Richard Zinbarg1 Michelle Craske4 Robin Nusslock1
1Northwestern University · 2Temple University · 3King's College London · 4UCLA
Project workflow from screening through SVR fitting and evaluation
From questionnaires and MID-task fMRI to a cross-validated support vector regression.

00Overview

Can reward-related brain activity predict dimensional symptoms of anxiety and depression? Using SVR on whole-brain task fMRI from 209 people, the answer was no: linear models fit worse than a flat line, pointing toward targeted networks and non-linear methods.

209
participants at risk for anxiety & depression
264
ROI features from a whole-brain atlas
10
-fold cross-validation

Machine learning has been applied to disease diagnosis and treatment prognosis with structural and functional neuroimaging, but few studies have related task-based fMRI activation to symptom dimensions of anxiety and depression. Support vector regression (SVR), whose complexity doesn't depend on input dimensionality, offers a data-driven, multivariate approach with sensitivity to small effects. We trained SVR on fMRI from the Monetary Incentive Delay (MID) task to predict dimensional symptoms.

01Study design

Tri-level model of general distress, fear, and anhedonia
Tri-level model. General distress, fear, and anhedonia z-scores, derived from scales across multiple questionnaires, are the SVR targets.
Monetary Incentive Delay task timeline
Monetary Incentive Delay task. An anticipation cue signals possible reward or loss; participants respond as fast as possible, then see the outcome. fMRI is recorded at each stage.

02Results

Across 12 contrast × symptom tests, R² and RMSE were averaged over 10 folds. R² ranged from −4.24 to −0.65 for whole-brain tests and improved slightly, to −1.69 to −0.46, using only the single best-correlated ROI feature. Negative R² means a horizontal line at the mean predicts better than the fitted model.

Model fit plots
Poor fit and underfitting. High dimensionality, noise, and unrelated activation lead the linear model to underfit.
R-squared and RMSE per test
Goodness of fit. R² and RMSE for each of the 12 tests; high RMSE also indicates poor prediction.
Residual distributions
Classifier comparisons. Normally distributed residuals point to the input variables as the cause of poor fit.

03Where to go from here

Model-based features. Restrict to theory-derived networks — default mode, cingulo-opercular, frontoparietal, salience — to reduce dimensionality and noise.
Non-linear methods. Explore non-linear models and clustering to capture the relationship between reward-related activation and symptoms.
Candidate brain networks for future analyses
← PreviousLarge-Scale DNN Benchmarking