Hi everyone, I'm Alex Vaz, a fifth year MD-PhD candidate at the Duke University School of Medicine. Today I want to tell you a little bit about my thesis work, particularly looking at the replay of spiking sequences during human memory retrieval. The two questions I had at the start of my PhD were these two. How does the hippocampus communicate memory-related information across the rest of the brain? What is a possible neural code of this human memory? The place I started was here with these things called ripples. Ripples are fast frequency oscillations measured in the hippocampus of many different mammalian species. Here you can see there are humans, monkeys, cats, rats, bats, etc. This is powerful because it's an evolutionarily conserved biomarker of episodic memory. The classic example of this comes from the rodent place cell literature. Here you can see a rat navigating a linear track. You can see over time these place cells tile the space. They form a sequence of firing in the hippocampus itself. Later, when the rat is now resting or asleep, you see these ripples in the hippocampus. During the same time as these ripples, you get this fast replay of the sequence that you saw before. This was the starting point of where I was interested in looking for this in the human data. The place I started was in human electrocorticography. These are recordings taken from patients undergoing invasive monitoring for seizure activity. These are patients who have typically failed two or three anti-epileptic drugs. The surgical option is then to resect the seizure focus itself. This is showing you here the intraoperative and postoperative pictures. If you do registration with the postop CT, you can also form a reconstructed map of where these electrodes are in space. The original place I started was looking at the signals that we can get from the patients while they're awake and undergoing their invasive monitoring. This is a representative right-sided implantation. If you look at signals coming from the medial temporal lobe, which I've abbreviated MTL here, you can see two representative ripples happening in the MTL of this patient. This by itself wasn't too interesting because this had been described before in the literature, but what was more interesting is if I looked at other places in the brain, so here the middle temporal gyrus or MTG, you could also see a ripple happening at the same time as the second ripple that you got in the MTL here. What the hypothesis was is that these ripples that were coupled across the brain versus the ones that were uncoupled would be the ones that would preferentially subserve memory retrieval or memory consolidation. These would be the ones that actually subserve this memory transfer across the brain. The way we tested that was now going with this simple memory task. This is a paired associates memory task where now I'm giving you finger and needle as a word pair, and then later I give you finger as a probe and you now respond with needle. It's a simple way to elicit a desired memory response. Now in a representative patient, this is a left-sided implantation. You can see right before they respond with the correct response, you can see ripple activity in the MTL. What was really nice was that this ripple activity is now coupled to the middle temporal gyrus. You can see two coupled ripples here. They were selectively coupled to the middle temporal gyrus, but not to the primary motor cortex itself, which we wouldn't expect. Now if you look at where these coupled ripples are happening across the brain, you can see here on the left is the baseline and right is the retrieval. If you just plot the ripple-ripple synchronization between the medial temporal lobe and the cortex across the whole brain, what you can see is during retrieval you get an increased level of synchronization with the anterior temporal lobe. This is nice because the anterior temporal lobe is where we would expect the semantic memories of these paired word associations to live. These coupled ripple oscillations between the MTL and the cortex could reflect memory transfer across the brain. Following that, a possible neural code of human memory could be contained within these spikes that are happening at the same time as these ripples. We implanted something called a Utah microelectrode array, or MEA. These sit directly underneath the cortical subdural grid. Here you can see four macroelectrodes are now sitting over the MEA itself. Here are the positions of the MEAs across the six patients in our study. Again, you can quickly measure these coupled ripples across the brain. Again, here there's a ripple in the MTL that's now temporally proceeding, ripples in the middle temporal gyrus. If you look at what's happening in the micro signals recorded from the Utah array, you can now see at the same time that you're having ripples in the macro EEG, you're also having ripples in the micro LFP, or local field potentials. Here I've also colored the ripples based on the instantaneous spike rate for each channel. You can see that the spikes are local to when the ripples are happening on that channel. You can more simply visualize this by seeing the whole raster at the same time. Now this is showing all of the units across all channels on this MEA. You can see there's a large burst of spiking activity that happens at the same time as these ripples. Now you can trace the neural activity across spatial scales, from the level of signal in the units to what's happening in the micro local field potentials, all the way up to what's happening in the macro ripples as well. If there are these bursts of spiking activity that you can see here, now I've just isolated one burst. Are these actually organized into sequences? Here I'm showing you on the left is one burst event, and on the right is the instantaneous normalized spike rate for each channel. Now if you reorganize the spikes on the left based on where their maximum spike rates happen on the right, what you can see is that these spikes actually happen in a really nice sequence over time. This is a 150 millisecond sequence. You can see now that the blue neurons fire before the yellow neurons, and you get this nice temporal progression of spiking activity over time for a single sequence. You can play the same game with multiple burst events over the course of the trial. Here the algorithm picks up three different burst events. What you can do is you can now reorder these units based on the relative temporal firing against each other. If you do so, the raster now becomes something like this. You can now see that the blue units are firing before the yellow units, and this average temporal relationship is preserved across multiple burst events over the course of the trial. Just simply plotting a regression line across the three different burst events makes this very obvious what's happening over time. Now we can look at the sequences during the task. Here the patient is memorizing the word pair crow-jeep, and you can see that there are 10 or 12 sequences happening over the course of their memorization period. Later the patient is presented with the word jeep, and it takes them about two seconds to recall the correct word crow. The hypothesis would be that the sequences during the memory retrieval period would be preferentially replaying what happened during the memory formation period. You can explicitly quantify this using something called the matching index, which is taken directly from the rodent place cell literature. If you plot that over time, you can now see that the retrieval sequences are increasing over time until they reach the highest point, and then the patient now recalls the correct memory. It's this idea that the sequence replay is part of the memory retrieval process, or replaying the memory from the encoding period. If you don't believe all of this, you can now just look at the sequences explicitly. Here's one sequence from encoding, and here's one sequence from retrieval, and you can appreciate right away that these are very similar to each other. This together supports the idea that the sequence replay is a neural substrate for human memory retrieval. What that means is that a possible neural code of human memory are these sequences of spiking activity. These two questions together now allow you to make a conceptual model of human memory. Here I've just highlighted in blue the MTL, and in red the cortex, here the anterior temporal lobe. With the first part of the story, I showed you that these coupled ripples happen between the MTL and the cortex during memory retrieval. During the second part of the story, I showed you sequence replay happening in the cortex itself. What we hypothesized would be that a similar sequence would be present within the medial temporal lobe itself also, but we didn't measure that in the current study. These tandem sequences across the brain would then be your neural substrate for human memory. With that, I'd like to acknowledge the Duke MSTP, the Duke Theory Center, my fellow graduate students, committee members, everyone at NIH, and my funding sources as well. Thank you.

Alex Vaz

MD-PhD candidate

Duke University School of Medicine

replay of spiking sequences

human memory retrieval