Cross-Decoding Analysis of Action Representations using CoSMoMVPA

A MATLAB implementation of cross-decoding analysis to investigate the neural representations of person-directed and object-directed actions in the human brain using CoSMoMVPA.

Overview

This script performs cross-decoding analysis to examine how different brain regions represent action features across different contexts. Specifically, it tests:

1. Person-directedness across objects: Can we decode person-directed vs. non-person-directed actions when generalizing across object-directedness?
2. Object-directedness across persons: Can we decode object-directed vs. non-object-directed actions when generalizing across person-directedness?

Academic Context

This script was prepared for the Advanced fMRI Analysis course at CIMeC (Center for Mind/Brain Sciences) under the supervision of Prof. Wurm.(2024)

Requirements

• MATLAB (tested on R2023a)
• CoSMoMVPA toolbox

Data Structure

The script expects the following data organization:

fMRIset2/
├── glm/
│   ├── SUB01_video_twoPerRunwise_sm3mm.mat
│   ├── SUB02_video_twoPerRunwise_sm3mm.mat
│   └── ... (19 subjects total)
└── msk/
    ├── univarConjunction_spherical_12mm_MTG.mat
    ├── univarConjunction_spherical_12mm_PMC.mat
    ├── univarConjunction_spherical_12mm_SPL.mat
    └── univarConjunction_spherical_12mm_IFG.mat

Analysis Pipeline

1. Data Loading

• Loads GLM results for each subject
• Applies ROI masks (12mm spherical ROIs)
• Four ROIs analyzed: MTG, PMC, SPL, IFG

2. Cross-Decoding Design

Test 1: Person-Directedness

• Training: Object-directed actions (targets 1-4)
• Testing: Non-object-directed actions (targets 5-8)
• Classification: Person-directed (3,4,7,8) vs. Non-person-directed (1,2,5,6)

Test 2: Object-Directedness

• Training: Person-directed actions (targets 3,4,7,8)
• Testing: Non-person-directed actions (targets 1,2,5,6)
• Classification: Object-directed (1-4) vs. Non-object-directed (5-8)

3. Classification

• Classifier: Linear Discriminant Analysis (LDA)
• Cross-validation: Leave-one-run-out
• Metric: Classification accuracy

4. Statistical Analysis

• Computes mean accuracy across subjects
• Calculates standard error of the mean (SEM)
• Performs one-tailed one-sample t-tests against chance (50%)

Usage

% Run the entire analysis
cross_decoding_analysis

% The script will:
% 1. Loop through all 19 subjects
% 2. Perform both cross-decoding tests for each ROI
% 3. Generate bar plots with error bars
% 4. Output statistical results

Output

Variables

• allRes: 3D matrix containing accuracies [subjects × ROIs × tests]
• meanAcc: Mean accuracy across subjects
• semAcc: Standard error of the mean
• H, P, CI, STAT: T-test results

Visualizations

The script generates two bar plots:

1. Person vs Non-Person Directed across objects
2. Object vs Non-Object Directed across persons

Each plot shows:

• Mean accuracy for each ROI
• Error bars (SEM)
• Chance level line (50%)

Key Functions Used

CoSMoMVPA Functions

• cosmo_fmri_dataset: Load fMRI data with mask
• cosmo_match: Select specific conditions
• cosmo_slice: Extract subset of dataset
• cosmo_nfold_partitioner: Create cross-validation partitions
• cosmo_crossvalidation_measure: Perform cross-validation
• cosmo_classify_lda: LDA classifier

Experimental Design

The analysis uses a 2×2 factorial design:

• Factor 1: Person-directedness (person-directed vs. not)
• Factor 2: Object-directedness (object-directed vs. not)

This creates 8 unique conditions that allow for cross-decoding analyses to test the independence of neural representations.

Notes

• The script uses 3mm smoothed data
• ROI masks are 12mm spherical regions
• Statistical tests are one-tailed (testing for above-chance performance)
• Chance level for binary classification is 50%

References

• Oosterhof, N. N., Connolly, A. C., & Haxby, J. V. (2016). CoSMoMVPA: Multi-Modal Multivariate Pattern Analysis of Neuroimaging Data in Matlab/GNU Octave. Frontiers in Neuroinformatics, 10, 27.

This script is for educational purposes as part of the Advanced fMRI course at CIMeC.