Merge pull request #15 from nhejazi/slides

The outline for our presentation

slides/progress.md

@@ -2,35 +2,83 @@
 % Nima Hejazi, Feng Lin, Luyun Zhao, Xinyue Zhou
 % November 12, 2015
 
-# Background
 
-## The Paper
+# Background (~30sec)
 
-- from OpenFMRI.org
-- ds001
+## The Paper (Nima)
 
-## The Data
+- "Working memory in healthy and schizophrenic individuals"
+- Accession number: ds000115 (from OpenFMRI.org)
+- The data is mainly functional connectivity imaging (fcMRI)
 
-- 12 subjects
-- 2 conditions per subject
+## The Data (Nima)
 
-## The Method
+- 102 subjects, consisting of healthy individuals and those with Schizophrenia, as well as their siblings
+- 2 conditions per subject???
 
-- linear regression
+## The Method (Nima)
 
-# Initial work
+- The paper used linear regression (ANOVA) after transformation to Talairach space
+- The goal was to examine connectivity within four brain network regions (e.g., the default mode network)...
 
-## EDA
 
-- downloaded data
+# Initial work (~30 sec)
+
+## Exploratory Data Analysis (Slide 1) (Lynn)
+
+- We downloaded and worked with a subset of the subjects (8 chosen from among the first 15 subjects in the dataset)
 - simple plots, summary statistics
 
-# Next steps
+## Exploratory Data Analysis (Slide 2) (Lynn)
+
+- fit a basic linear model using OLS on a per-voxel basis within subjects
+- used the residuals from the linear model to examine normality of the errors
+- tried to find activation regions using the time-coruse (without convolution)
+- multiple testing...
+
+
+# Next steps (~3 min)
+
+## Pre-processing and Validation (Xinyue)
+
+- dropped the first 5 volumes
+- for some methods we normalized the data
+- on/off times, sometimes we dropped the off times
+- Since convolved neural prediction has a unit of 0.1 second while TR has a unit of 2.5 seconds, our strategy picks the median of the span of neural prediction values that correspond to a TR.
+- Read on-off neural prediction from condition file in FILENAME, convolve the neural prediction with a predefined HRF function.
+
+## Statistical Analysis (Xinyue)
+
+- fit a basic linear model using OLS on a per-voxel basis within subjects
+- used the residuals from the linear model to examine normality of the errors
+- tried to find activation regions using the time-coruse (without convolution)
+- used convolved neural prediction values as a baseline function, we attempted to visualize the activation regions based on the correlation coefficients...
+
+## Unsupervised Learning with K-Means (Liam)
+
+- used K-means by clustering voxels (in groups of 5) with different features
+- features used: (1) mean BOLD over timecouse per voxel, (2) BOLD over timecourse per voxel, (3) normalized BOLD over timecourse per voxel
+- pictures...
+
+## Future (Liam)
+
+- stuff we want to do
+
+
+# Process (~1 min)
+
+## What has been the hardest part? (...)
+
+- problem 1: fMRI data and having an open assignment (figuring out how to use resources to develop a coherent analytic plan) Xinyue
+- problem 2: how to figure appropriate analyses and understand validity of results Liam
+- problem 3: dealing with functional connectivity in the data
 
-## Preprocessing / Validation
+## Success in overcoming obstacles (...)
 
-- PCA
+- Found out that the condition files were not ideal for our analyses (mismatch between the theoretical techniques and the messiness of provided data -- spoke to Matthew about TR issues with time alignment, and had to decide between mean/median/first unit in the timecourses)
+- What parts of the class have been the most useful? (1) assignment too open/vagueness and learning to deal with ambiguity in research, (2) collaboration with git
 
-## Statistical Analysis
+## Issues with reproducibility? (...)
 
-- linear model
+- Reproducibility not as much of a problem as initially thought...
+- Learning to work with the git model (using code review with merging, etc.) has made reproducibility not too much of an obstacle