MVPA Searchlight Analysis for Music vs Non-Music Decoding

A Python implementation of multivariate pattern analysis (MVPA) using searchlight approach to decode music vs non-music stimuli from fMRI data. This analysis uses the Nilearn library and includes robust statistical comparisons against both dummy classifiers and theoretical chance levels.

Overview

This script performs a searchlight analysis to identify brain regions that can reliably distinguish between musical and non-musical emotional stimuli using fMRI data. The analysis includes:

• Support Vector Machine (SVM) classification with searchlight
• Comparison against dummy classifier baseline
• Statistical significance testing with multiple correction methods
• Comprehensive visualization of results

Dataset

This analysis uses beta images derived from the openly available dataset:

Lepping, R. J., et al. (2016). Neural Processing of Emotional Musical and Nonmusical Stimuli in Depression. PLoS ONE, 11(6), e0156859. https://doi.org/10.1371/journal.pone.0156859

The dataset is hosted on OpenNeuro and can be accessed here: https://openneuro.org/datasets/ds000171

Requirements

numpy
pandas
nilearn
scikit-learn
matplotlib
scipy
joblib

Usage

Basic Usage

python searchlight_analysis.py

Configuration

Update the paths in the main() function:

csv_path = "path/to/your/searchlight_data.csv"
mask_path = "path/to/your/mask.nii"
output_dir = "path/to/output/directory"

Input Data Format

The script expects a CSV file with the following columns:

• beta_file: Path to beta map NIfTI files
• group: Subject group (script filters for 'control')
• target: Stimulus type (1,3 = music; 2,4 = non-music)
• subject: Subject ID for cross-validation

Analysis Pipeline

1. Data Preparation

• Loads beta maps and filters for control group
• Recodes targets: 0 = Music, 1 = Non-music
• Applies brain mask to constrain analysis

2. Searchlight Analysis

• Radius: 4mm spherical searchlight
• Classifier: Linear SVM (LinearSVC)
• Baseline: Stratified dummy classifier
• Cross-validation: Leave-one-subject-out
• Metric: Area Under ROC Curve (AUC)

3. Statistical Testing

• Comparison against dummy classifier (accounts for class imbalance)
• Comparison against theoretical chance (AUC = 0.5)
• One-tailed tests (SVM > baseline)
• Multiple comparison correction:
  • FDR (False Discovery Rate) correction
  • Cluster-based thresholding (min cluster size = 20 voxels)

4. Output Files

The analysis generates multiple output files:

NIfTI Maps

• control_music_svm_auc.nii.gz: SVM AUC scores for each voxel
• control_music_dummy_auc.nii.gz: Dummy classifier AUC scores
• control_music_diff_from_dummy.nii.gz: Difference between SVM and dummy
• control_music_diff_from_chance.nii.gz: Difference from chance (0.5)
• control_music_z_vs_dummy.nii.gz: Z-scores for SVM vs dummy
• control_music_z_vs_chance.nii.gz: Z-scores for SVM vs chance
• control_music_cluster_vs_dummy.nii.gz: Significant clusters (vs dummy)
• control_music_cluster_vs_chance.nii.gz: Significant clusters (vs chance)
• control_music_threshold_[0.55-0.7].nii.gz: Thresholded maps at different AUC levels

Visualizations

• control_auc_comparison.png: Multi-slice comparison of SVM vs dummy AUC
• control_significance_maps.png: Statistical significance maps
• control_summary_statistics.png: Histograms and summary statistics

Interpretation

• AUC > 0.6: Regions showing above-chance decoding performance
• AUC > 0.7: Regions with strong decoding ability
• Significant clusters: Regions surviving multiple comparison correction
• SVM vs Dummy: Accounts for class imbalance in the data: For this data set we have 3 runs for music and 2 runs for non-msuic 3:2

Example Results

The analysis will output:

• Brain regions capable of decoding music vs non-music
• Statistical maps showing significance levels
• Summary statistics including mean AUC and number of significant voxels

Troubleshooting

Memory Issues

• Reduce n_jobs parameter in SearchLight if running out of memory