Exploring predictive and reproducible modeling with the single-subject FIAC dataset

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Functional magnetic resonance imaging (fMRI) traditionally focuses on spatial mapping of brain activity.
• Predictive modeling offers a complementary approach to extract more information and understand brain systems by predicting brain states.

Purpose of the Study:

• To investigate the performance of various predictive models in fMRI analysis.
• To evaluate the impact of preprocessing and feature selection on predictive model accuracy.
• To identify the potential and limitations of predictive modeling in fMRI.

Main Methods:

• Utilized block datasets from the Functional Imaging Analysis Contest (FIAC) Subject 3.
• Compared five predictive models: linear discriminant analysis, logistic regression, linear support vector machine, and two Gaussian naive Bayes variants.
• Assessed the influence of preprocessing steps (e.g., temporal detrending) and feature selection methods.

Main Results:

• Temporal detrending and feature selection consistently improved predictive model accuracy across all models.
• Linear support vector machine and logistic regression generally outperformed Gaussian naive Bayes models.
• Predictive accuracy was typically lower in principal component feature spaces compared to voxel spaces.

Conclusions:

• High prediction accuracy in fMRI does not automatically imply the learning of visually interpretable brain activity patterns due to potential artifacts.
• Cross-validation provides reliable estimates of prediction accuracy.
• The trade-off between prediction accuracy and spatial pattern reproducibility is crucial; prediction alone should not be the sole optimization in fMRI analysis unless the goal is direct brain-state classification.