Boost in Test-Retest Reliability in Resting State fMRI with Predictive Modeling.

Area of Science:

• Neuroimaging
• Cognitive Neuroscience
• Machine Learning in Neuroscience

Background:

• Recent studies highlight low test-retest reliability in functional magnetic resonance imaging (fMRI) features.
• This raises concerns for the reproducibility of neuroimaging research findings.

Purpose of the Study:

• To systematically evaluate the test-retest reliability of multivariate predictive models in neuroimaging.
• To compare the reliability of predicted outcomes from models versus individual fMRI features.

Main Methods:

• Applied 10 predictive modeling methods to resting-state fMRI connectivity maps from the Human Connectome Project dataset.
• Predicted 61 outcome variables using these models.
• Assessed reliability across various scanning and processing choices.

Main Results:

• The mean reliability of predicted outcomes from predictive models was substantially higher than the mean reliability of individual resting-state connections.
• Reliability improvements were consistent across all modeling methods, scanning, and processing choices.
• The most reliable methods achieved "good" reliability (above 0.60) for predicted outcomes.

Conclusions:

• Multivariate predictive models offer a promising avenue for enhancing test-retest reliability in neuroimaging research.
• Researchers can potentially achieve higher reliability by integrating predictive modeling approaches.
• Identified mechanisms explaining the superior reliability of predictive model outcomes compared to individual features.