EEGs, Learning, and Deep Sleep
Who would have thought a summer fellowship would grant you sleep sessions during work hours! Not just that, but it also comes with the ability to explore the deepest phases of sleep and access to unlimited Delta waves that come in all shapes and heights! Well, that can only happen at Backyard Brains, right from the interesting sleep lab I’m running this summer in collaboration with Om of Medicine.
Om of Medicine: Where the Magic Happens! Om is letting us use part of their lounge as our sleep lab, where subjects come and perform the study.
For the past couple of weeks, I have been working diligently on designing and implementing the experimental procedures to test if inducing consolidation during sleep by cuing certain auditory stimuli can improve memory recall. This is done using the TMR or Targeted Memory Reactivation technique, where we selectively target memories, reactivate them, and compare them to ones that are not targeted or cued with any stimuli. Such methodology allows us to explore different parameters to learn more about the specificity of memory formation and bias in learning. From here, my project splits into two main parts: The memory task and EEG recording/decoding.
For the first part, I am collaborating with Dr. Ken Norman from the Princeton Computational Memory Lab and two of his students: Robert Zhang and Everett Shen to develop an iOS software for the memory tasks. The goal is to have this be a fully functional app users can download from the App Store and run their own sleep studies.
The memory task is simply comprised of 3 main parts. The first part is the learning task, where subjects would watch 48 different images being displayed on random locations on the screen, each with a distinct sound correlated with it, for example: cat with a meow sound. Subjects should try to memorize the location of where each image was displayed. Following this phase comes two consecutive rounds of testing with feedback, where subjects would see each image and should then click on where they think it’s correct location should be based on what they remember. Following this multi-stage learning phase, the subject would do the actual pre-sleep test. This is essentially the same as the previous two rounds, but without the feedback. The second part of the app, is the cueing phase that will be played during the nap when the subject is sleeping. The idea is to cue 24 targeted sounds out of the 48 the subject listened to before the nap. For the other 24 untargeted sounds, we play a baseline sounds that the subject did not listen to before the nap (so different than all the 48 presented). Part three is the post-sleep test which is again the same as the pre-sleep test.
Part 2 of the app: cueing phase, should play only in the Slow Wave Sleep cycle, where Delta waves are observed. Here comes the second cool aspect of my project: EEG recording and Decoding.
Some screenshots from the current working version of the app. It is still being developed and improved upon. Code can be found soon on GitHub. CC
Scoring Sleep Stages and spotting Delta waves in real time can be very challenging. The end goal of this project, is to be able to detect deep sleep automatically and cue the sounds accordingly. For now I am using our EEG setup and Spike Recorder to observe Delta waves in real time as the subject is sleeping, once I see them, I start cuing the sounds from the app.
My beautiful Delta Waves in different shapes and height taken from our subjects. Delta waves typically
have a frequency of 0.5-3 Hz with an amplitude of around 75 microvolts
After recording, I am performing signal analysis and plotting of frequency and power graphs in different variations to check that Delta waves are happening at the same time we did the cueing in real time. So far, the results are on point and matching!
Top Left: Subject 1, Top Right: Subject 2, Bottom Left: Subject 3, Bottom Right: Subject 4
One of the most challenging tasks in my project is to find subjects willing to volunteer, perform the task and sleep. As this step is very crucial, I designed a brochure and gave it out during Tech Trek and to various parties. There is a doodle poll where subjects can sign-up for sessions.
Throughout this time, I learned Matlab from scratch and worked more with electronics and soldering. During the sleeping session, I play white noise using a generator, and the cueing sounds from a speaker placed next to the subject’s head. The trick is to not have the cueing sounds be more than 4dB higher than the white noise in order not to wake up the subject. Setting this up took a lot of testing and playing around with different wires, sound meters, and speakers. All subjects were asked after waking up if they listened to any sounds while sleeping. All assured they did not, which is good for us because we can make sure the procedure is working. Next to the speaker sits the EEG shield connected to the Arduino. The electrode placements are as follows. Reference electrode on the mastoid, active electrode and ground on frontal lobe using our EEG headband.
Top: iPad running the Memory Task. It is connected to the speaker placed inside the room by the subject’s head. I cue the sounds from it once I observe Delta waves. Mac for recording EEG in real time and scoring/observing SWS. We have both extended outside the room so that I don’t wake up the subject by sitting in the room with them.
Bottom: Speaker, white noise generator, and sound meter.
Subjects during the session. Photos were taken with the permission of the subject and taken at the very end of the nap, right before waking them up.
Finally, here comes the best part!! Getting our data that agrees with the expected literature and published papers.
This is the basic plotting of the data we got. More statistical analysis regarding error bars and figure labeling will be applied. The graphs show the mean distance in pixels of the 48 images for each category: Cued and Uncued, before and after sleep. The distance is between where the user clicked and where the original location of the image should be. This distance is compared to a certain threshold we set and compares to it. Larger distances mean more error and thus quantifies as incorrect. Smaller distances mean less error thus quantifies as correct. We can clearly notice that both subjects performed worse on the uncued sounds after sleep compared to before. Subject 4 also clearly shows an improvement in recall for the cued images after sleep compared to before. This supports the TMR technique and shows the selectivity of memory consolidation and recall.
The upcoming final month will be filled with more exciting work and experimentation. I will be running more experiments on more subjects to double-check our results. Then I will start with the control experiments, were some subjects would not sleep, and other would but have no sounds cued at all. Stay tuned!!