G-ICA supported region-based feature selection technique for fMRI data. By: Indranath Chattterjee Department of computer science, University of Delhi, Delhi-110007, E-mail: indranath.cs.du@gmail.com Run the code in the following order: 1. Run GICA on 3 runs of SIRP task fMRI data 2. Taking mean of all the runs for each subject for each component. 3. Now we have 13 components for each of the subjects. 4. Select 116 brain regions from AAL atlas. 5. For each component, a. Mapped the connected brain regions identified by ICA with the marked regions in the atlas. b. From each component, take those regions of the AAL atlas map which is also present in that particular component. c. For each region, we measured five statistical features (say a1 to a5) on the basis of voxel values of that particular region for all the subjects individually. d. Find the FDR score for each of the five measures for each region for all the subjects. e. Sort and arrange the FDR scores in descending order. Higher the FDR score, more important the particular measure is. f. All the statistical measures for every region are arranged in descending order of FDR score. (5 measure*116 region = 580 features) g. Runs the scheme in LOOCV leaving one subject for testing. i. Iteratively add each of the features (a1 to a5) for all one by one according to descending FDR score. (Sequential forward selection) ii. Training set and test set are built with the selected feature subset. iii. Feature subsets are selected, if it results the high accuracy in classification task. This way the best set of features can be obtained. iv. Check to see which region has more contribution in selection of feature subset, i.e. the more number of statistical features in a particular region. v. Count and save those regions. vi. Repeat the same for all 68 subjects, as LOOCV runs. h. Check the average classification accuracy for each component and note the identified brain regions after all rounds of LOOCV classification. 6. Now check and note the brain regions occurring most often in each of the components and also check the frequency of its occurrence during 68 times LOOCV classification.