This week I was working on the REU presentation video, getting feedback from my mentor and the PhD student I work with in order to do a concise/coherent product. I started the week by plotting curves with the correlation values between Amplitude Coupling and Phase Coupling states based on the optimal alignment of temporal profiles per each subject. We expected these curves to look like a bell-curve. I tried normalizing the curves to plot them with the temporal profile averages. The correlation curves of z-scored data weren’t interpretable since the values are insignificant. On the other hand, the curves of non z-scored data don’t showed a noticeable variance, so I went on to perform a two-way ANOVA test to test the interaction between the Amplitude Coupling and Phase Coupling Temporal Profiles. There was indeed a significant interaction between AmpC and PhC. I then plotted the non z-scored spatial correlations (y-axis data) using the pairs of states extracted from the temporally-matched states of z-scored data.

I started this week by comparing the State Time courses (Gammas) of the two HMM models using a function in the toolbox. I then used the outputs to find matching states between the two models. Then I compared Lifetimes, Interval Times, and Fractional Occupancies of matching states and plotted it. At first, I only generated figures for the first subject. It turned out that calculating the Lifetimes, Interval Times, and Fractional Occupancies would optimally be done using the Viterbi Path extracted from the model rather than the Gammas, so I had to switch around the Viterbi Path to match the new state pairings. Then I adapted my code to process all this for all subjects of the same frequency band and plotted this. (The plots you saw today). Additionally, for visualization purposes, the Amplitude Coupling data was organized in descending order (Phase Coupling data was arranged accordingly).

This week, I worked on creating null distributions for the data by phase-permuting the data and computing the correlations. I researched and met with my PhD student to better understand the Fourier Transform and the idea behind the phase-permuting. Then I used the null distributions to determine the significance of the correlations in the actual data with a 95% confidence. This entailed creating graphs to better display the data.

This week I extracted HMM from non z-scored data (I had been working with z-scored data before) to extract state specific functional conductivities. From this data, I computed correlations between z-scored data (which filters out baseline activity to only keep task-related activity) and non z-scored data. We discussed the findings which aligned with the previous paper’s fidnings of the PhD student I’m working with. Following this, I extended the analysis to more subjects (previously I had only been working on one). The results were pretty robust. Also, I learned more about the purpose and idea behind the research.

This week, since I have already tested my code and it seems to be yielding viable results, I’m implementing it on the actual data for the study.

This week, I continued my work on the implemented toolbox for analyzing the brain data. I learned a few Matlab features that allow me to find the code for functions of the toolbox, which is very helpful. So far, I discussed by email on how to find the optimum number of clusters for the model. We considered a method suggested in the toolbox, Bayesian Information Criterion, and the elbow method. I have mostly been involved with understand the tools in the toolbox to analyze the results in order to extract features out of the trained model. The end is to decipher and quantify the behavior of each state. The results so far seem reasonable, which could imply that the pipeline is working properly. Now I will be working on better understand the research behind this current project and tentatively go on to compare the behavior of the states across different conditions.

In the past two weeks, I familiarized myself with the Hidden Markov Model (HMM) algorithm, Fourier Transform, and a Matlab toolbox that uses HMM to analyze brain data. Right now, I am using that toolbox to analyze data collected from epileptic patients undergoing brain surgery, so the device, unlike EEG, is directly on top of their brains. So far, implementing the toolbox has not been particularly difficult, given that I previously read the entirety of the documentation and took notes. However, I have found that the documentation is not as thorough as it should be and many functions are not outlined, which is a bit of a hassle. I will continue looking around the wiki and hopefully find something, but I have come to terms with the possibility of there not being anything. We will see. To end on a lighter note, I am in the Eagle County in Colorado and have gone to two hikes on the mountains. The first one was not particularly dangerous, but the last one went by a deep chasm, a huge waterfall, and up a rocky mountain. I drank water from above the waterfall, which comes from a lake on the top of the mountain, and it was pristinely pure.