Michael Kahana
Subject: A Breakthrough in Restoring Memory
Bio: Former Professor of Psychology at Penn and Founder of Nia Therapeutics
Transcript:
Larry Bernstein:
Welcome to What Happens Next. My name is Larry Bernstein. What Happens Next is a podcast that covers economics, politics, and history. Today’s episode is A Breakthrough in Restoring Memory.
Our speaker is Michael Kahana who is one of the world’s leading experts on memory. He is a former Professor of Psychology at Penn, and he is now the founder of Nia Therapeutics which is a company seeking to bring a medical device to patients that can radically improve memory for individuals with a brain injury, dementia, or Alzheimer’s.
I want to hear from Michael about how memory function works and how a medical device can improve human memory retrieval to improve cognitive abilities.
I met Michael 10 years ago at a previous Penn Days event, and I was incredibly impressed by his research. As a result, I became one of the founding investors in his business venture.
This podcast was taped at a conference of UPenn professors that I hosted so you will hear questions from me as well as my friends.
Micheal, please begin with 6 minutes of opening remarks.
Michael Kahana:
My talk is about human memory and my journey from professor at Penn to building a company that will hopefully change the world. After 22 years at the University of Pennsylvania’s Department of Psychology, I’ll be dedicating my life to building Nia Therapeutics. A company that will be advancing therapeutics for memory loss.
Let me first start by telling you how I got into human memory. I started out studying physics in college. I was fascinated by mathematical descriptions of the physical world and by the human mind. I started reading up about neural network theory, the work that’s at the foundation of modern AI by John Hopfield, David Rumelhart and Jay McClellande.
I did my PhD working on mathematical models of how the brain stores and retrieves information. I was fortunate to have a wonderful academic career for over three decades studying memory, its mathematical description, and its neural basis.
The major focus of my early work was disproving the theory that says that memory involves two separate systems, short-term and long-term memory. That is a dated theory that has been disproven by extensive experimental data and theoretical work. I developed a new theory of memory that was an amalgam of ideas that is the basis of AI models, neural network theory and a concept called recursive retrieved context that endows a neural network with the ability to have memories that exist at a particular time and place in the system storing those memories.
I’m going to share with you two stories in my academic journey that changed my life. The first was in 1997, a young neurosurgeon Joseph Madsen at Children’s Hospital in Boston took me to the ninth floor North Wing and I saw a young man aged 16 sitting up in bed. He had a big ribbon cable going out of his brain into a port on the wall, and he was playing an Atari video game.
I asked, “who’s analyzing the data?” Joe said, “We don’t have enough memory storage to save the data. Once we see a seizure, we know what we’re going to do, and we delete the data at the end of the day.” That was the first time that I took a major risk in my academic career. I took my money and bought computers and hard drives. I set up a collaboration to harvest brain neural recordings from neurosurgery patients with intractable epilepsy who need to have electrodes implanted in their brain, record the signals, figure out where the seizures are coming from and then remove it if they can pinpoint it in the brain. That began a journey that lasted for two decades identifying the brain signals at a micro level within the brain that predict memory function.
The next major event happened in 2013, at the Society for Neuroscience meetings in New Orleans, we studied a patient where we had stimulated the brain and his memory got better. Now, usually, when you stimulate with electrical currents in the brain, memory get worse. This one patient his memory got better. Was that a one-off? How could we systematically investigate brain stimulation on a large scale? Because every hospital did it in a very ad hoc manner. There was no real science to it.
We got a big grant from Jeff Lang at DARPA, the Defense Advanced Research Projects Agency, who was a colonel in the US Army to help wounded warriors who had memory or cognitive loss due to IED blasts in Afghanistan and Iraq. We were given a $24.5 million-dollar four-year grant, which is about 10 times as big as a typical NIH grant and it allowed us to ramp up the study of electrical stimulation memory.
We invented a device here at Penn together with Medtronic, which hooks up to electrodes in the brain. For three years we mapped out the neural signals in the brain associated with memory and collected data from hundreds of neurosurgical patients. We gently stimulated the brain to learn how stimulation modulates neural circuits.
There were three major discoveries that came out of that project, but before I tell you, I want to tell you about an unsolved puzzle in memory research.
Memory is way more variable than any theory predicts. The theories predict some variation over time in the same individual but the data show massive variation. So, the same person one day their memory is pretty good, another day their memory is not so good and that range of memory function within each of us, especially people with a brain injury, neurological disease, dementia, that range is enormous. How large is it? It’s so large that if you take somebody in their bottom 50th percentile versus their top 50th percentile, there is a two-fold difference in memory function on every metric. So, when you’re not doing so well with your memory, you’re doing half as well as when you’re in your upper half of your own distribution. Huge variability and theories don’t explain it. I thought maybe the neural recordings could explain the variability.
Discovery one was we could use AI techniques to forecast when your memory was going to be good or bad. It’s a forecasting model, just like in financial time-series models. In this case, we can forecast memory very reliably.
Discovery two was if you looked at when brain stimulation pulses arrived, did it matter what the forecast said? It turned out it had a huge impact because if the forecast is that the brain is doing well, then stimulation made memory worse. But if the forecast said that the brain was doing poorly, then brain stimulation sometimes didn’t do anything and sometimes it made memory better.
So, the third question was, how do we learn when the stimulation does nothing, and when it makes memory better? That’s a simple engineering problem because it’s a search of the parameter space of stimulation, locations, amplitudes and frequencies in each person and you build a model. You take this massive data set on each human brain and you learn when I stimulate a bad state, what’s the probability that I jostle the brain into a better state or stay in the bad state? And we learned that.
How do we deploy this? $14 million of DARPA money went into inventing this gadget and we tested the therapy in human subjects who had electrodes implanted in their brains. In five published papers and four data sets, we showed that we could reliably improve memory consistently in human patients with this algorithmic stimulation.
How do I take this to market? Well, you can’t wear this device as a therapy. It’s too big. You can’t have electrodes coming out of your brain. You need a fully implantable system. The government is not going to fund that. So private investors, started a company, a spin out from Penn, and we developed a device similar to deep brain stimulation for Parkinson’s disease, which is now a widely used therapy. This device is a cranial implant. It doesn’t go in the brain but goes in the skull under the skin. It has electrodes that go into the brain. It has a hundred times the capability of recording than any existing device, but notably it has a computer inside of it and it can record the signals, analyze and deliver stimulation just like this larger device connected to a computer did in our research studies, and it harvests all the data and it goes on to a wearable earpiece like a hearing aid.
The plan is to bring it to human trials in the next six months. The first study will be in patients with traumatic brain injury. There’s no treatment for traumatic brain injury memory loss. The FDA was excited for these younger, healthier patients in their 20s or 30s who have debilitating memory loss. This therapy should work for any memory loss due to any neurological disorder or neurodegenerative process.
We have a team assembled to do a study in early Alzheimer’s disease based at Rutgers Alzheimer’s Center and the University of Texas Alzheimer’s Center. Then the third big activity that my company is going to be doing is to make the device a lot smaller.
Larry Bernstein:
Tell us about how a robotic neurosurgeon will install wires inside the brain.
Michael Kahana:
The implantation of EEG electrodes is done with a ROSA robot by neurosurgeons. 20 years ago, this was a risky procedure. Now it has essentially zero serious risks associated with it. The biggest aspect of this surgically is the craniotomy, where you put in the device. They wires go inside the memory centers of the brain. One of the reasons why no existing therapy can do what this does is because this does a different therapy for every patient.
There are 64 contacts that are placed along the key axes that cut through the memory areas of the brain with these tiny wires. It learns to model the memory of every patient, which is going to be a different model for me, for Larry, and for each patient, which is why it has the capability of learning to get better and better over time.
Larry Bernstein:
All of us are in good and bad memory states all day long. Even in a good day, there are bad moments. Tell us about how an electric shock does a reset and why it takes us back to a good memory state.
Michael Kahana:
The memory system is a complex electrical network that has stable configurations. It can fall into different stable attractor states. Some of those attractor states are conducive to good function and some are conducive to bad function. If I’m in a bad state, I deliver a small pulse to white brain matter tracts that coax the brain from a state that’s poor into a state that’s better.
Larry Bernstein:
And how does that work exactly?
Michael Kahana:
A short quarter of a second wave of electrical activity. You have an anode and cathode, two contacts that are made out of platinum and you deliver biphasic electrical pulses between these two. You don’t feel or hear anything, but it makes a little electrical pulse. And what we do with the device is we see the state of the brain before and after the burst. We see whether the burst coaxes the brain in a good direction or not and that’s how we can learn and adapt the therapy.
Larry Bernstein:
When you go to bed, you would charge up your device and download all of your brainwaves.
Michael Kahana:
All your brain data goes to the cloud encrypted. When we download the data and then we can analyze and improve the models of your memory system. It knows when your brain is in a better or worse state and it recalibrates the model on the fly, meaning if the brain is shifting its context, it can recalibrate the AI model.
When you do memory tests, you can tune the therapies. The therapy will learn that when your memory test scores are better or worse, it will refine the AI model. And you can rebuild the model every night and get a better model based on the new data that you’ve downloaded.
Larry Bernstein:
The next question comes from Rory MacFarquhar.
Rory MacFarquhar:
Why do you encrypt the data?
Michael Kahana:
For confidentiality of your brain data.
Rory MacFarquhar:
If someone got your brain data, what could they extract from that information?
Michael Kahana:
We published a paper where we showed that when you had people look at letters, we could tell what letter they were thinking of. If you thought of a word, we could tell whether that word was one of 25 different categories. You could probably build models to predict what people were thinking about.
Rory MacFarquhar:
Have you tried to rebuild something from the data that you’ve collected on all your patients? Is that literally possible? Can you put it up on a screen and see what they’re thinking?
Michael Kahana:
We published a number of papers showing this. The idea that you can use brain data to read the mind. Speech recognition systems are 90 odd percent accurate. These systems are way less accurate, but they’re way better than chance. If the question is what exact word am I thinking of, they’re going to be very unreliable. If the question is, am I thinking of a living or a non-living thing, they’ll be very reliable. The whole basis of the neural prosthetics work in motor systems is that you can read out the motor activity of the brain and build a model of speech or motor activity and that’s been advancing very rapidly in the last three years.
Larry Bernstein:
What kind of improvement in memory are you expecting?
Michael Kahana:
I would expect an improvement of about one standard deviation, so it’s a pretty big improvement. A standard deviation improvement in terms of an SAT score would be 200 points out of 1600, or for IQ, it would be 15 points with a 100 mean.
Larry Bernstein:
What does a one standard deviation improvement in memory entail?
Michael Kahana:
To evaluate memory function in real life with implications for whether somebody goes back to work, goes back to school after a head injury, or whether they’re going to develop dementia, it’s context dependent recall tasks. If you give people a list of words and then you ask them after a brief delay, what were the words in the list? That’s the kind of test. And there are many varieties of those tests and they all load on the same underlying factor, which is the same factor that we’re decoding.
Larry Bernstein:
I gave you 10 words and the average is six. Would a standard deviation be something like two?
Michael Kahana:
Yeah, that’s about right.
Michael Kahana:
We’re talking about people who give them a list of 10 words and either they get zero or one or they get three. And so that’s when we’re talking about patients with memory loss. We’re saying three is not great, but you can function. It’s not embarrassing. Zero or one, you don’t even want to leave the house. You don’t want to go to work.
That’s the thing about memory loss. People have a bad memory. It’s not such a big deal. A colleague of mine in my department once commented at a faculty meeting that “my memory isn’t as good as it used to be. I had this legendary memory. I would walk into a class with a hundred students and it would take me a few weeks to learn everybody’s name. I can’t do that anymore.” That’s not a problem.
The problem is when you’re sitting in the meeting like this and you don’t realize why you’re here and what’s going on and what am I supposed to be talking about? And the problem is that people with impaired memory, they have days like that and then they have other days where they’re just a little bit forgetful. It’s not such a big problem.
The goal of this therapy is to keep people out of those situations where they’re going to not want to go to work, not want to show up in front of their colleagues. We have prominent lawyers who get into a car accident. They have a head injury and they say, “I’m embarrassed to show up in a meeting. I don’t remember what’s going on. I have notes and I’m reading them over and over again just to figure out, to get my bearing. And it’s impossible.” That’s the kind of memory that we’re trying to blunt that horrible situation that people face.
Larry Bernstein:
Our next speaker is Brande Stellings.
Brande Stellings:
For degenerative memory loss that’s going to worsen over time, does starting this therapy arrest the deterioration of memory? Does it have any impact on it getting worse?
Michael Kahana:
The one standard deviation increase is about five to six years of degeneration. The question that you’re asking is critical. Does the rate of degeneration change? We will only know that when we do the trial.
Larry Bernstein:
Our next speaker is David Wecker.
David Wecker:
I was just curious how big you think this market is in terms of revenue. Does the device ever have to be replaced and is there a recurring revenue stream from the cloud service?
Michael Kahana:
It has a sales price of about $35,000, Manufacturing price at scale would be less than $10,000, so you’re talking about a $25,000 margin. The total possible number of patients who could benefit is 27 million. My expectation is that over a period of years, it’ll ramp up from small numbers to 10%. You could imagine different curves based on different levels of effectiveness.
In terms of recurring revenue, the idea of putting this earpiece or eyeglass as an element that can be easily updated every year you could get a new version. People can pay for that privately. Some insurance will pay, some insurance won’t, same as hearing aids. Some insurance will not let you upgrade your hearing aids. Medicaid probably won’t, but private insurers sometimes will. You can have recurring revenue from that and from cloud services and from the updates to the memory model that would be done periodically.
Larry Bernstein:
Michael didn’t say you need to replace the thing that’s in the brain.
Michael Kahana:
No, that’s intended to be permanent. We got breakthrough device designation, which gives us a special priority reimbursement where they give you an upgraded reimbursement above the base Medicaid- Medicare reimbursement level, the whole issue is market penetration. The market is enormous and it’s just about how quickly you capture the market.
Larry Bernstein:
The market is enormous now, but even you, Wecker, years from now, you’re a great candidate.
David Wecker:
Hopefully many years. Do you plan on trying to commercialize this yourself with a sales force or do you expect to sell or team up with a larger organization that has those resources?
Michael Kahana:
I want it to be successful and to scale and if the best way for that to happen is through an IPO, then we’ll do that. If the best way is through strategic acquisition by a larger company, then whatever makes the thing grow.
Larry Bernstein:
Our next speaker is Rory MacFarquhar.
Rory MacFarquhar:
There were memory impaired human subjects that you were testing. I didn’t think that you could do that with human subjects. How does that work?
Michael Kahana:
There’s only one neurosurgical treatment that requires chronic implantation of electrodes for a period of one to four weeks where you can record all those data. That was when I gave that talk at Harvard and they invited me to the ninth floor of the children’s hospital where patients with epilepsy who failed all their medications were having debilitating seizures, those patients also often have cognitive impairment and they get electrodes implanted in their brains for three or four weeks to get seizure data for a treatment involving the resection of the seizure focus. I saw the opportunity, the patient is sitting there, they’re playing video games in the hospital, let me do my memory experiments, have them play memory video games and then build models of memory. They were doing stimulation clinically so that when they would suction out a part of the brain, they didn’t cause harm.
Larry Bernstein:
Can you give an example where when you had a portion of the skull open and you put a probe in, what response, if you shocked it, how they would have a memory release?
Michael Kahana:
The patients would be playing memory games or remember lists of words and the computer unbeknownst to the subject would apply therapy on some trials and not apply therapy on other trials. And sometimes there was one patient at the Mayo Clinic, they got the stimulation and they’re remembering almost the entire list of 12 items and on the other trials they were remembering three items. That was a pretty dramatic case. Every time the stimulation therapy was applied, this patient had phenomenal memory and the patient reported they didn’t feel anything. They were just like, “My brain was clear on those trials.”
Larry Bernstein:
Our next speaker is Colin Teichholtz.
Colin Teichholtz:
There are other companies doing neural implants such as Neuralink. If each one of these types of devices is totally different and unrelated, are there additional applications to what you’re doing in addition to what you’re doing with memory?
Michael Kahana:
There are a number of competitors to Neuralink that are focusing on what I would call the input-output problem. They’re trying to solve the fact that some people have a damaged spinal cord or have a damaged input-output system. They’re recording the brain to read out motor signals. And they’re also trying to plan to do similar things for patients who have perceptual disabilities. A cochlear implant would be an example that’s very commonly used but now they’re developing therapies for people who are cortically blind to stimulate the cortical system to create sight. So those are input-output systems.
The other category of systems are central. They’re focused on the apex of the neural system neural processing tree. That’s like my device. There’s nothing else like it when I tried to do this. People said, “There’s no possible way you’re going to build a de novo medical device. You have to use an existing device.” I’m not Elon Musk, so I couldn’t just capitalize it on my own. But you could imagine a class of devices of which this is a version that treats central disorders that would include depression, PTSD, higher level language dysfunction. Memory is the big one. There’s a company that has an idea for treating severe PTSD, and they’ve recently approached me to see if we would license our technology to them because they didn’t imagine that they would want to try and build their own.
Larry Bernstein:
Our next speaker is Paul Orlin.
Paul Orlin:
In your paper you talk about random stimulation does not help memory but specific predictive pulses from the AI does. Can you explain why you think that is?
Michael Kahana:
Random stimulation is going to interrupt the brain during good states and bad states, whereas the AI predictive stimulation is going to be tailored to only perturb the brain when it’s in a bad state with parameters that are designed to coax it into a better state. So the random stimulation is not going to work. That makes sense because if random stimulation would work, this problem would have been solved a while ago with DBS type devices.
Paul Orlin:
What do you think’s happening to create the memory?
Michael Kahana:
You could imagine a version of this therapy that would create memories. We’re not doing that. I’ve thought about that. It’s science fiction. You could imagine the study that I proposed once, which I haven’t done, is you could barcode memories.
For example, if I wanted to, if I click my phone and I could say, “I want to remember that guy and his question and put a barcode in there and then later recapitulate that memory by triggering that barcode.” The therapy that we’ve developed is designed to take the brain and help it get to states that it knows how to facilitate the formation and the retrieval of memories. Naturally move the brain into a state that’s more conducive to creating and retrieving memories.
Larry Bernstein:
Our next question comes from Paul Rozin who is a Professor of Psychology at UPenn.
Paul Rozin:
Someone is in a bad memory state and you stimulate. When you get them in a better state, when you turn the stimulus off, did they go right back to a bad state or is there a residue?
Michael Kahana:
There’s a residue. We’ve shown that. There’s a residue of staying in the good state.
It will stop after a half or a quarter of a second and then it will wait for the brain to drift back into a bad state before it reapplies the therapy seconds or minutes later.
There was a conversation about kosher laws and rabbis before, so I just wanted to add a little fun note. I asked a famous old rabbi, “What do you think about this?” Is it kosher? Can you use it on the Sabbath? And he looked at me and he said, “Is this going to help me remember the whole Talmud?” And I said, “I think so. “ And he said, “You should do that.”
Larry Bernstein:
Our next speaker is Ron Miller.
Ron Miller:
How long is it going to take to finish these trials and how much money is it going to take?
Michael Kahana:
Five years. After you develop the device, you do the animal testing, which we’re finishing now. Then you do in humans for five years altogether. Because we got this breakthrough device designation and there’s a special program called TAP, which is run by the director’s office at the FDA. They try to fast track devices. The whole project as like a 10-to-12-year project from start to finish and it costs altogether about a hundred million dollars, the whole business.
David Wecker:
Remembering can be an intentional act but it can also be a passive act because memories sometimes just parade through one’s consciousness. There’s also the issue of sleep where we seem to be incorporating memories. I’m curious about how this might affect our consciousness where memories are popping up.
Michael Kahana:
I believe and my research supports the claim that the same basic machinery of memory is used when you try to remember a memory, you pull back something from the past, you update memory when some cue from the environment or from inside of your head involuntarily triggers the memory. Now those two are not exactly the same, but they involve the same mathematical rules. They involve the same equations and related to that is the idea that when you do have a spontaneous thought, it is a memory and it operates as a memory that alters your future beliefs, which is a fascinating idea because it means that circumstances that make one set of memories more likely to pop up than another set of memories will shape your beliefs separate from the ground truth of what happened to you.
David Wecker:
When you do the human trials, are you going to be able to collect any data around those types of issues?
Michael Kahana:
Absolutely. From a clinical and a pragmatic point of view, the therapy helps people in their daily lives and the best we can do right now is we can do tests that evaluate memory in a way that’s predictive of memory in your daily life. What you’re saying is that maybe the person with memory loss is having fewer of these spontaneous memories. And so, if you restore some of those spontaneous memories, maybe that will give them the capability to alter their beliefs in both good and not good ways based on their experiences. There may be situations where not learning and remembering is good. I don’t think that’s generally the case, but yes, if you’re learning and remembering things that are painful that then recur, we know that among people who’ve experienced a traumatic event, about 15% go on to develop PTSD and that’s because of the machinery of the memory system going awry.
The Nia Therapeutic trial with this device, one patient undergoing the therapy for a year will generate more data by a factor of 10 than all the data we’ve collected in these 20 years of NIH and DARPA funded studies. That’s one patient in one year. You can imagine what you would learn from the first 5,000 patients.
Larry Bernstein:
Next speaker is Brande Stellings.
Brande Stellings:
For many companies their real business is data.
Michael Kahana:
The data will be enormously valuable for a given patient to improve their therapy. The company by pooling data in an encrypted, safe, anonymized way will be able to deliver far better therapy to the next patient in the study. Data are extremely valuable. The only thing I’ve been focused on is the therapeutic side.
Larry Bernstein:
Many of us might be reticent to install a wire system in their brain, but when we see Mitch finally being able to function, and we see incredible improvement in his cognitive skills and memory capabilities.
Michael Kahana:
In neurosurgery, things that were unimaginable two decades ago are becoming accepted today and once you know people who benefit from whether it’s a cochlear implant or DBS for Parkinson’s disease, I think this will be more effective than either of those therapies, then there’s no more stigma. Nowadays, people don’t worry about getting a cardiac pacemaker, but 25 years ago, very few patients who needed pacemakers got pacemakers. What you see is you see a course of adoption where a therapy that is clearly beneficial to people is adopted at 1% in the beginning and then later it’s adopted at 20%. It’s never adopted at 100%.
Larry Bernstein:
Do you want to end in a note of optimism?
Michael Kahana:
When I was teaching, students get to do one lecture where I just answer their questions. One student Kobe asked me, “How did the study of memory help you think about how you could be a better person, live a better life?” The answer I gave him is, “Every day try to make good memories for yourself because the memory system is going to think about those memories when you’re asleep, when you’re walking around and make good memories and that will benefit you in many ways.” That’s my note of optimism. And this is a good memory, so thank you.
Larry Bernstein:
Thanks to Michael for joining us.
If you missed our previous podcast, it was Predicting the Midterms.
Our speaker was Patrick Ruffini who is a pollster and a Substack writer for The Intersection that does analysis on topics like gerrymandering, realignment, and polling errors. He is also the author of a book Party of the People.
Patrick discussed who will win in the midterms in the House and Senate. Our discussion got into the weeds about campaign issues, the current state of play, and who will win the close races.
Previous episodes and transcripts are available on our website as well as at Apple Podcasts and Spotify.
I am Larry Bernstein with the podcast What Happens Next.
Check out our previous episode, Predicting the Midterms, here.
