Using K Means clustering to identify coactivation of neurons

Hi, I want to identify clusters of brain cells (cells firing in synchrony) in a data set of 15989x36 (15989 frames of video for 36 cells). The matrix consists only of 1s and 0s, being 1 where a cell has activated and 0 everywhere else. My problem lies in how to display these clusters, since their spatial positions are stored in a completely different matrix than their activation spikes.

[Martin Mandelbaum]

I'm aware that K Means is not an ideal method of doing this. I'm just looking to provide a comparison against a better and more specific algorithm.

Answer 1

[Image Analyst]

"My problem lies in how to display these clusters" <=== use plot3() or scatter3().

[Martin Mandelbaum]

Hi, thanks for your answer. My question might have lacked some clarification: my cells aren't located in a 3D space, they are all located in a single plane, with only X and Y coordinates. Besides their 36 by 2 matrix listing these coordinates, the matrix I wish to divide into clusters is the large ~16000 by 36 one, where each column is a single cell and each row is a single frame indicating if each cell is "on" or "off".

When I try to follow the examples provided for the kmeans function, I get a plot with the centroid locations, but without any of the cell locations, do you know how I might combine both datasets to coherently showcase the clusters?

Sorry if my original question was vague!

[Image Analyst]

It would be less vague if you attached the variables in a .mat file with the paperclip icon, and posted a screenshot of what you have, and if possible, what you'd like to obtain.

[Martin Mandelbaum]

here's my very simple code, following the examples provided.

[Martin Mandelbaum]

and this is an example of how I would like my data to be organized :

the visualization is not strictly necessary, rather, I wish to know which cells correspond to which ensemble/cluster

[Image Analyst]

I don't understand. I don't see clusters at all with those two variables (x and y). The solid red spots seem completely mixed in with the hollow circles. Why do you think there are clusters?

[Martin Mandelbaum]

The red spots are specific cells that activate simultaneously and repeatedly, thus forming a neuronal ensemble. This coactivation isn't spatial as you can see, but rather temporal, because these clusters represent cells all set to "1" at a specific row in the 'binary' matrix. I thought that running this temporal matrix through a K Means algorithm might give me an idea of which cells are active at the same time for a specific frame(row)

[Image Analyst]

What do the different columns of binary represent?

But then if your data is x, y, time, shouldn't you only pass in the third column into the kmeans()? Why are you having it cluster off of combination of spatial coordinates and time?

[Martin Mandelbaum]

Each column is a specific cell, so binary(1,1) represents the state of cell 1 at frame 1, either active and spiking (1) or inactive (0). Binary(1,2) is the second frame of activity for the same cell, and so on.

The matrix 'positions' is only there for visualization purposes, and could be removed from this problem. The location of each cell is somewhat irrelevant. What I do wish to know, is which cells are active at the same time, e.g: if at row 1, cells 1,2 and 3 are all 1, and that behaviour is repeated for many different rows, then we could conclude that cells 1,2 and 3 are a cluster.

Rather than the visualization example I provided, a list of cells present for each cluster would be enough. So I would like to get as an output something like a list of vectors, where each vector contains a list of numbers, 1through 36, indicating which cells form each cluster.

I realize that this is different from the typical implementation of K Means, but I believe that it could be applied in this specific scenario.