AI Draws New Value from Old Medical Technology - Dr. Jacob Donoghue, Co-Founder & CEO of Beacon Biosignals
Michael Carrese: Hi, everybody. I'm Michael Carrese. Today, we're going to dive into the fascinating story of a very old technology, EEGs, being mined for data using a very new technology that's changing the way that patients are treated for disorders of the brain. With us to explain is Dr. Jake Donoghue, co-founder and CEO of Beacon Biosignals, a startup that applies artificial intelligence to EEGs to unlock precision medicine for neurological, psychiatric, and sleep disorders.
Dr. Donoghue earned his MD from Harvard Medical School and PhD in neuroscience from MIT, leading research into the effects of neuroactive compounds on brain network activity. His published works span epilepsy, cognition and machine learning methods for quantifying pharmacological effects on neural activity. Thanks so much for being with us today.
Dr. Jake Donoghue: Thanks for having me.
Michael: It's going to be interesting to dive into this. But first, we want to learn more about you and your background and what first got you interested in medicine.
Dr. Donoghue: Yes, so a long journey. Probably didn't fall too far from the tree with a mother that's a pediatric neurologist and a father that's a neuroscientist. I think in early days, I was interested in the brain and the opportunities to take care of patients and seeing that the power of being able to take care of groups of particular children with pediatric neurologic disorders and the impact you can make on their lives really stood through as I paved my path through my training at Brown and then at Harvard and MIT.
Michael Carrese: So, once you were in med school, was there anything else that kind of confirmed that that was the right direction for you? Did you consider other things?
Dr. Jake Donoghue: I was in medical school, so I got my MD, PhD. It was a long journey there. It was a great opportunity to really dive deep and understand, in my particular instance, the neurophysiology and how that relates to cognition and the way that drugs affect brain activity. I think that that's the beauty of the MD, PhD program in general and at Harvard and MIT in particular where they really allow you to think about the number one goal which is having big, harnessing, deep scientific insights to make an impact on patients, and this path provided an opportunity to make that impact.
Michael Carrese: Sure. I said at the beginning that EEGs are very old technology...I think about 100 years old. Could you give us kind of a grounding on what they are and what they do before we figure out what you're doing with them?
Dr. Jake Donoghue: Sure, absolutely. EEG -- or electroencephalography -- records the electrical potentials from the brain activity from the scalp. It is a 100-year-old technology. Really, the first rhythm -- the alpha rhythm -- was discovered by Hans Berger from the posterior parts of the head when you close your eyes, and you can see this ten hertz rhythm up here in the back of the head. So, from these electrodes on the scalp, we're able to understand both disease processes and healthy brain activity by looking at how the electrical activity patterns change over time.
Michael Carrese: And give people an example of like how that is typically used.
Dr. Jake Donoghue: Yeah, so EEG is the sort of standard of care as part of diagnosing epilepsy. Looking for epileptiform activities, such as seizures on EEG or epileptiform spikes, maybe subclinical events. And also it's worth remembering that EEG is a standard part of the polysomnogram suite, so folks getting a PSG for diagnosing sleep disorders are also getting their brain activity monitored.
Michael Carrese: So, here this is sitting out there for 100 years, and at some point you must have had an insight about applying newer technologies to maybe get more information out of those. Tell us about that and kind of the backstory of how Beacon came about.
Dr. Jake Donoghue: I think the beautiful thing about time series data from the brain is it represents one dimensional images. So, what I was -- and all of us have been -- witnessing as part of taking care of patients and just generally in the field, was how image classification and recognition tasks have really transformed the way we think about digital pathology and radiology, and how we can take those tools and apply them to this data modality where we record large quantities of data but don't have the tools to necessarily extract all the features of interest and understand their links to how patients feel, function, survive, and those underlying diagnoses.
So, at Beacon, I’m really spending my time thinking a lot about brain waves and how drug effects to brain activity happen. During my training, it became very clear that we could harness the same types of tools that were making massive improvements in clinical workflows and in driving forward precision biotechnologies in oncology and medicine in general, and thinking about a similarly heterogeneous space of psychiatric disease and neurologic disease where we have this physiologic activity that we've never really been able to understand the full depth and breadth of to really capture the disease heterogeneity.
So, Beacon emerged from realizing we can apply these tools to these datasets and really understand the ways that we can drive forward precision medicine for the brain.
Michael Carrese: So, are you training the AI on like the dataset of a patient, or a huge dataset of patients?
Dr. Jake Donoghue: Important question. Yeah, so machine learning, as I'm sure everyone's aware now, with large language models and ChatGPT, benefits from really large datasets for training that need to be diverse, heterogeneous, so that they can generalize to any patient.
At Beacon, we have partnerships with academic medical centers and some of our EEG acquisition vendors where we can aggregate over 50,000 EEGs and polysomnograms to be able to use that data to label it and train machine learning algorithms to identify features within those large recordings as well as a consensus of experts. So, what we do there is actually have multiple experts label hundreds of thousands of examples of epileptiform spikes. We can have eight epileptologists label hundreds of thousands of these individual sharp waves that last under one second and we can train the model to guess what percentage of epileptologists would call this individual event an abnormal feature.
Taking this consensus-based modeling, we're able to remove a lot of that inter-rater variability and produce robust, validated pipelines to be able to understand and run inference over very large recordings really quickly.
Michael Carrese: So, I'm a patient and you get my scan and the AI is able to tell, based on this huge dataset, maybe what I've got or what I haven’t got?
Dr. Jake Donoghue: Yeah, I think two big examples that we like to use and stand out is some of these patients with developmental and epileptic encephalopathies may have hundreds or thousands or hundreds of thousands of epileptiform spikes depending on the duration of the EEG recording. So, you can imagine in this indication where you start a new treatment in one of these children and you'd want to know, is there a change in the background abnormal epileptiform activity?
Well, it's infeasible for a human to count the number of spikes that may be present, or number of electrographic seizures, and so machine learning tools allow us to measure that burden quantitatively and see that maybe there was a 10-20% reduction in some of these events. That may have huge implications for that kid's long-term behavioral and cognitive motor development. They really just get missed through the gestalt and sort of subjective terminology we use in our medical records to describe those features.
On the converse, a patient may arrive with an individual spell and they will get an EEG and there may only be one event of potential suspicion in a twenty-minute recording where one second has an event that sort of caught the eye of the epileptologist. Well, if that person's a pilot or someone that drives their car, making that diagnosis based on that one sharp wave event -- and, of course, in context of the clinical symptoms -- can have a huge impact for that person. So, one way the machine learning tools here can be added is basically providing this expert consensus modeling so that you don't have an under-caller or an over-caller of abnormal events reading that EEG and making a diagnosis that could transform that person's life positively or negatively.
Michael Carrese: So, as the provider, is it just sort of a confidence issue that this is what I'm seeing but I've got this Beacon information that's telling me that, yeah, I'm on the right track?
Dr. Jake Donoghue: So, Beacon right now is really focused just on our biopharma partners.
We utilize our tools to bring quantitative endpoints into clinical trials, so we're able to do things like measure every single epileptic form spike, subclinical seizure, and sleep. We can do automated sleep staging and count sleep microarchitecture events like spindles and arousals. These are the types of tools that we're validating in our clinical trial partners to help understand target engagement, whether a new treatment is affecting brain activity and whether there might be certain responders to certain medications.
Some patients with depression may have hypersomnia, sleeping twenty-two hours a day. Some of the patients I remember seeing -- as opposed to insomnia -- sleep just a few hours every night. So, we're really thinking about what would it mean in these novel clinical trial settings to bring more quantitative insights that might define these two different classes of patients with depression in a data-driven way, and there probably are even more to really understand which types of patients, what is their underlying sort of endophenotype and why certain groups may respond to certain treatments better? Again, this is the mentality that everyone sort of really thinks about in oncology that we think we can translate to the neurology and psychiatry space.
In the longer term, we do see this opportunity to translate to the clinic where we can bring these tools to really improve individual patient care.
Michael Carrese: Ah, gotcha. You've mentioned epilepsy and sleep and psychiatric disorders. Are there other clinical trial areas where this is particularly useful or effective?
Dr. Jake Donoghue: Yeah. One big one is in Alzheimer's disease. We put research out last year building off of a foundation put forth by a few groups over the last five years or so that concerns why some patients with Alzheimer's disease decline faster, have cognitive decline faster, than other subgroups of patients with Alzheimer's disease. One area of research we were interested in is seeing if patients with Alzheimer's have abnormal electrical activity patterns. So, we actually looked at our large data set -- over 50,000 patients -- to find the patients with Alzheimer's disease that had no comorbid epilepsy -- so they had no history of seizures -- but we looked at their EEGs, and we actually found subtle hints of subclinical epileptic form activity.
That had been seen a few times before, but never in a really large real-world data setting. The idea is there are potentially subgroups of patients that have their hippocampus involved in a particular way, where there's a hyper-excitability feedback loop that might define them as having faster cognitive decline, and therefore may be better or worse responders to some of these novel Alzheimer's next-generation therapies.
Michael Carrese: Ah, so in other words, if you have this tendency toward epileptic seizures, your Alzheimer's could progress more quickly?
Dr. Jake Donoghue: That's what previous research has shown, and we've replicated elements of that.
Michael Carrese: That's terrific. Obviously pretty profound to be able to advance understanding of Alzheimer's. And what else are you working on that you think holds a lot of promise?
Dr. Jake Donoghue: Seeing The Year of The Zebra logo behind you, I think it’s worth noting that we're really focused a lot in rare disease in developmental and epileptic encephalopathies. I think that there's tremendous unmet need there, where these diseases that can range from a couple thousand patients to tens of thousands really are at the bleeding edge of needing next-generation precision technologies, and there's a huge wave of antisense oligonucleotides, gene therapies emerging that really could have a huge impact for these patients.
One of the issues that we see is that for these families and for these children, the way that we measure efficacy often relies on seizure diaries where we ask the families to annotate in a paper or digital diary when they have a witnessed clinical seizure. You can imagine the burden that puts on families, of course, and what's more, there are obvious events that are just unable to be witnessed because the patient is sleeping, or it’s a subclinical event -- there may be nothing to see even if you're looking at your child. Or if you, as the parent or caregiver, are sleeping and there's no one watching.
Michael Carrese: Yeah, you can't be there 24-7.
Dr. Jake Donoghue: Exactly, and so in addition, we think that capturing the underlying brain activity almost of course has to represent the brain pathology of these developmental and epileptic encephalopathies, and by definition where the epileptic activity is probably driving some of the pathology. And so we think there's a really big opportunity in clinical trials where seizure diaries are going to be fraught with inconsistencies and may miss the true underlying disease burden, and scores looking at patients’ developmental and motor outcomes may take years to see which is infeasible for a clinical trial setting because they need to be done in months or a year.
So, for groups like SYNGAP1, KCNT1, Dravet Syndrome and others, we’re really excited where we can apply our tools to measure quantitative endpoints -- such as electrographic seizure burden and electrographic spike burden -- and see maybe in really short order that these drugs are having a tremendous effect on brain activity and can help get these new therapies approved and into patients.
Michael Carrese: Oh, that'd be terrific. So, as you may know, Osmosis is a teaching company, and one thing we love to ask our guests is to give us some direction about where we should focus some attention as we make educational videos and develop courses and so forth. So, is there a topic that you're particularly interested in where you just think people don't get it? There are myths or gaps or something that should be understood that isn't right now?
Dr. Jake Donoghue: Yeah, I think one area that really we should all spend more time thinking about is in sleep medicine and how it actually applies to not just disorders like insomnia and narcolepsy, but how sleep is a fundamental part of the core pathophysiology we see in many diseases. Obviously in psychiatric disease, or it can be a part of, as we already talked about, depression or PTSD, generalized anxiety disorder. These are symptoms that may be actually linked to the core underlying etiology.
Similarly, we see sleep as a core factor in neurodegenerative disease most prominently in disorders like Parkinson's disease, where patients have REM without atonia, but also even more subtle features we think we can pull out in diseases like Alzheimer's. So, we think there's a lot of opportunity in this fundamental state that all humans go through to actually bring quantitative insights to really map some of these really robust features of brain activity to help understand disease and health.
Michael Carrese: I'll pass that along to our crew, because it's a very interesting area to explore. So, a lot of the folks listening to this are either med students, nursing students, learners of other types, or early career professionals and we always like our guests to provide their kind of “go-to” advice about getting through med school and approaching their career, particularly at this time when everybody's been through this tremendous upheaval with the pandemic and there's all these disruptive technologies like what you're working on coming through. It’s a very unsettled time, I think, for folks. So, what's your advice about navigating all that?
Dr. Jake Donoghue: I'm only a few years out, but I think no matter what path you take to have an impact for patients, it's an arduous one. It takes a lot out of you -- a lot of time and emotional energy -- so the most important thing is that you have to find what brings you the most joy and reward to know that you're having that impact. Whether that's being in the clinic, seeing patients, or thinking about how you want your impact to potentially be at a different scale or take a different form. I think at the end of the day, just figuring out what you love the most is really the most important thing.
Michael Carrese: And did you always think about becoming an entrepreneur and starting something, or was that unexpected?
Dr. Jake Donoghue: Yeah, I think it was a bit unexpected, although, you know, I think I always had an interest in finding my own path and trying to have an impact on patients was really the driving narrative for me. But I think for Beacon in particular, you know, seeing that there was this opportunity with the current state of tools and access to data and the unmet patient need... as these ideas formulated with my co-founder, Jarrett Revels -- who's a software engineer extraordinaire -- it really became clear there was no other path for what I could do and what we could do. This was just a huge opportunity in front of us that we had the ability to potentially impact millions of patients. So for us, it sort of evolved that this was the track we were on and we had to do it.
Michael Carrese: Yeah, we hear that a lot from entrepreneurs. Once you identify a problem and you think you've got a solution, it's sort of a little irresistible, despite the apprehensions about getting into business.
So, as we wrap up, is there anything else you want to share with us about yourself or Beacon Biosignals before we let you go?
Dr. Jake Donoghue: It's been a pleasure getting some time to speak to you. Obviously, I'd love your listeners to keep digging into neurologic and psychiatric diseases and really understand that there's this huge opportunity we see similar to what oncology looked like ten or fifteen year ago where we were often defining heterogeneous disorders by the histology of the slides. Now we don't think twice that we're getting molecular and genetic profiling to understand what precision therapies to give and cure these patients. I think everyone should be looking forward and thinking about how we can drive that similar narrative and that transformation in neurology and psychiatry, where I think we're really on the cusp.
Michael Carrese: That's a great note to end on. I want to thank you very much, Dr. Donoghue, for spending time with us today. We wish you the best of luck with Beacon Biosignals.
Dr. Jake Donoghue: Thanks so much.
Michael Carrese: I'm Michael Carrese. Thanks for checking out today's show, and remember to do your part to raise the line and strengthen the healthcare system. We're all in this together.