Find the weights that maximize the distance between recovery not recovered group

[yacineMahdid]

The for sure recovery group is composed of: WSAS02,09,19 and 20
The for sure not recovered group is composed of : everyone but `WSAS10.

[yacineMahdid]

For this we can try out a linear model with the three sum of the contrast matrix and then the weight and bias are found easily. We use the two class for the modelization. We can then visualize the points in three d space with the hyperplane made by the model.

[yacineMahdid]

With attempt #4 we have an index that illustrate that there is something there in terms of prediction of RoC. However, it is not fullfilling the 'prediction' part of our narrative. We need to use either statistical analysis of significance or machine learning to build a predictive model.

My (Yacine Mahdid) exchange with my collegue:

@sbmoraes and @cduclos could you remind me of the big picture for the dpli-dri analysis. What are we trying to say with that index?

cduclos  8:17 PM
We are trying a capture, with a single index, how much the brain network re-organizes under anesthesia. Our aim is to show that this reorganization can reliably predict potential for consciousness (i.e. eventual recovery of consciousness) in unresponsive patients.

8:19
Our other argument is that baseline connectivity/network hubs alone aren’t sufficient to predict recovery, and that the adaptive reconfiguration under anesthesia is what is most predictive of potential for recovery.

So the objectives of this index of consciousness is two-fold:

• Aim 1: show that reorganization under anesthesia can reliably predict RoC in WSASXX patient.
• Aim 2: show that baseline connectivity alone is performing worse in predicting RoC in WSASXX patient than reorganization.

If we want to attain aim 1 or 2 we will need more than the number of timepoints we currently have (only 10 points). A solution for this problem is to make a classifier of the binary recovery outcome based on all of the windows of data we currently have and make it interpretable.

I propose that we go with the following plan:

• Calculate dPLI on 10second windows for each state (~30 windows)
• Generate a dataset where we take all the permutation of the three state windows to generate BvA, RvA and BvR.
• Store in a dataframe sum(BvA), sum(RvA) and sum(BvR). This will generate a 270 000 data points * 3 feature data set.
• Train a logistic regression or a decision tree (interpretable models) on the dataset using a Leave-One-Subject-Out (LOSO) cross validation. This will allow us to quantify the predictive power of our index.
• Train the best model on the full dataset and look at the resulting weights to define our dpli-dri index (it will be a linear combination of sum(BvA) sum(RvA) and sum(BvR).

We are making the assumption that by using the full sum of the contrast matrix we will have separable state, but we can see that by eye that it is the case for the average so I am pretty confident we can get a robust classifier.

[yacineMahdid]

Charlotte did her analysis with a similar setup, it wouldn't take too long to put something together.