EEG data can bolster rare disease drug research and trials. An epilepsy drug trial, for instance, could ask patients to log the number of seizures they experience in a day. “If you look at a seizure diary, there might be 10 seizures a day. But if you look at EEG, there could be 150,” said Dr. Jacob Donoghue, a neurologist, neuroscientist and CEO of Beacon Biosignals.
EEG data can thus help improve the treatment of epilepsy by uncovering hidden patterns in patients’ brains. Such data also holds promise for testing experimental drugs for sleep disorders and neurodegenerative conditions such as Alzheimer’s disease, Donoghue said in a recent interview.
In the following Q&A, Donoghue explains how EEG data — paired with machine learning — can inform clinical trials and how such data can make the fields of neurology and neuroscience more objective.
The responses have been lightly edited.
Drug Discovery & Development (DDD): Could you tell us about your background and what led you to cofound Beacon Biosignals?
Donoghue: I did my MD/Ph.D. at Harvard and MIT. I had the luxury of being around cutting-edge technologies and seeing them be applied in digital pathology and radiology. Those spaces were already queued up to benefit from advanced machine learning techniques and platforms. At MIT, I worked on neurophysiology, studying drug effects on the brain and got a little taste of what it was like to apply cutting-edge machine learning techniques to brain data. That involved studying the effects of different drugs on the brain changes in brain states.
I was fascinated with the pairing between deep scientific research and taking care of patients — seeing the fruits of all that labor when caring for patients receiving oncology services. The cutting-edge techniques enable measurements for biotechnologies to help late-stage lymphoma and leukemia patients. I had the pleasure of helping take care of these patients.
After that, I transitioned to the space I knew best in neuro, psychiatry and sleep medicine. I saw that we often use subjective measurements to understand the patient’s disease and don’t always have the best treatment options. For instance, Alzheimer’s patients were advised to put sticky notes on cabinets to help them remember where things were. Meanwhile, we’re giving gene therapies and other precision medicines to help cure different types of patients.
Seeing some of the clinical work from our co-founders at Beacon Biosignals got me excited about using vast datasets to help address some of these unmet needs.
DDD: What are some ways that EEG paired with AI or machine learning could help inform the treatment of conditions like Alzheimer’s disease?
Donoghue: There are a few different ways. In general, diseases of the central nervous system tend to evolve, but also there’s a significant amount of fluctuation within a given day for these patients. Quantitative monitoring of electrophysiology and brain function allows a better assessment of the patient’s true aggregate state.
For Alzheimer’s in particular, I think one of the things that’s immediately available is that some patients with Alzheimer’s disease who have never reported having a seizure have signs of epileptiform activity below the surface when getting an EEG.
Patients with these waveforms have faster cognitive status decline. So, one of the things we think is exciting for EEG is we can identify those patients using our machine learning tools. We can identify these patients who represent a different phenotype, and we can either intentionally exclude them from trials because we know they’re declining faster. There’s something different about what’s happening in their brain. Not all types of Alzheimer’s are identical.
Based on the mechanism of their drug, a drug that targets hyperexcitability might work particularly well for a subset of patients. So, this is one great example where we get more precision into the subpopulations of brain function of these heterogeneous diseases. And then, data can inform how a novel therapy may affect these patient populations differently. For example, if you could throw out 20% of the patients in a trial that were non-responders, the remaining patients might have a better chance of actually responding.
DDD: What are the most promising areas for this marriage of AI and EEG?
Donoghue: I think three spots are epilepsy, psychiatric disease and neurodegenerative disease.
EEG is already the clinical gold standard for epilepsy diagnoses and sleep medicine studies.
Through machine learning tools, we can now interpret the full depth and breadth of this data to understand different types of epilepsy better and understand which patients are responding to which types of treatments.
When an EEG is ordered for kids with a rare type of epilepsy or adults with temporal lobe epilepsy, often they get these long EEGs to look for epileptiform spikes, which can be these sharp waves, almost like a QRS complex in EKG. And sometimes, they can be rare — like a needle in a haystack. Other times, there can be thousands of these abnormal events.
An EEG in epilepsy gives us a lot more precision in describing the disease, characterizing the features of brain activity that represent the pathology, and then seeing how that might change with either a new therapy or a change in therapy for patients.
I think we’ll be doing much more brain monitoring in the future on patients with epilepsy to help with more targeted therapies. The industry doesn’t need any convincing that’s valuable.
We’re bullish long-term on sleep in using EEG for psychiatric medicine.
The core diagnostic criteria for diseases like major depressive disorder (MDD) and bipolar include a significant sleep component. For MDD, you can meet the criteria by either having hypersomnia (sleeping too much) or insomnia (not sleeping enough). But, of course, something very different is happening in the brains of these patients. And it wouldn’t surprise us at Beacon that a therapy that might work on depressed patients with insomnia might not work on patients with hypersomnia.
Understanding one of the core symptoms that represents the neurophysiological underpinnings of the disease will help us build more targeted therapies and then give us more endpoints to study if they’re working.
Lastly, I think there’s a lot of hope for neurodegenerative disease. The key is going to be intervening earlier.
Clinical trials tend to focus on minimal cognitive impairment or early-stage Parkinson’s. So how do we reach patients earlier when they may have the best shot at responding to treatment?
We’re focusing on how sleep interplays with neurodegenerative diseases. So, for example, you may have a REM sleep disorder, where you have limb movements. That can be an early onset sign that Parkinson’s may later develop. And in Alzheimer’s disease, you see significant changes in sleep architecture in terms of the amount of slow waves and how fragmented your sleep is. So I think that EEG offers a lot of opportunity for early diagnosis and then brings a way to help quantify the changes that might come with a new therapy.
Right now, the symptoms of these patients are often reported in snapshots. So the opportunity to have a longitudinal measurement of these patients over time, especially in patients whose disease may make it hard to accurately reflect the subjectiveness of their disorder, makes it an exciting modality to target.
DDD: EEG data can inform which drug or combination of medications might work better for a patient with a specific sleep habit?
Donoghue: Exactly. And it can inform when to treat such a patient.
Beacon started with over 50,000 patients’ EEGs. So we have the world’s largest annotated EEG data set built for interrogation. For example, a drug company might want to know how the sleep patterns changed for patients with major depressive disorder in a specific age range who were also taking another drug. So we can continuously query these types of questions to help our pharma partners think about how their new mechanism might work or look at the real-world evidence of their drug to think about how to expand the label of an approved drug.
You can easily imagine the companion diagnostic of the future that deals with sleep. If a drug is sleep-promoting for major depressive disorder in some patients, it may be the right drug to take at 10 p.m. for some patients. For others, it might be right to take it at 6 p.m. There’s a lot of interest in precision medicine so that the individual patient can get the optimal therapy.
DDD: What are the main customer types that Beacon Biosignals has?
Donoghue: We’re focused on biopharma drug development. We’re working in the epilepsy space heavily. One of the most exciting things we can offer is the state-of-the-art EEG quantification of seizure and spike burden. And so that’s been interesting for phenotyping patients as endpoints for clinical trials. Does this new drug decrease the amount of EEG seizures? Does this new drug reduce the amount of EEG spikes?
We’re also doing work in neurodegenerative diseases such as Alzheimer’s. We also have multiple groups in sleep medicine and are working in major depression and schizophrenia. We are working directly with biopharma partners and are also starting to talk to CROs. We believe that every neuro, psychiatric or sleep-medicine drug-development program would benefit from more quantitative tools to measure the effect on the brain. The actual organ of interest, the brain, is the most complicated organ in the known universe. And I think studying how brain activity changes will be critical, as long as we have the tools to look at and obtain that data.
DDD: How is the company exploring how long COVID affects the brain?
Donoghue: COVID definitely affects neurons. The virus can destroy olfactory sensory neurons, which can affect the sense of smell.
For long COVID, the question is why some patients have more prolonged symptoms. The ones that are interesting to us are, of course, the neurocognitive. Those include things like brain fog and sleep disorders.
There’s a tremendous opportunity to think about how antivirals like Paxlovid and vaccines might ameliorate some of these responses. It’s an open research question. To study how novel therapies change brain function to help patients with symptoms, we will need to measure brain function. An EEG is probably the most scalable way to do that.
DDD: Can you share more about how machine learning can be used to interpret EEG data?
Donoghue: AI is still a hot buzzword. What we think is unique about Beacon and other companies like PathAI is how robust and diverse of a data set you are starting with. Any algorithm is only as good as the data it is trained on. One of these massive advantages is that we have this bedrock of these unbelievably detailed datasets. The datasets have been annotated and represent heterogeneous and diverse patient populations to ensure that what discoveries we have generalize to patients worldwide.
We have patients from four hours old to 100 years old in our database. We have a whole suite of algorithms and machine learning techniques, such as deep learning. One example is our epileptiform spike detector for helping quantify the burden of disease and epilepsy. Sometimes you want to count the epileptiform spikes — these sort of sharp waves in the EEG. It is not feasible to do this in some rare diseases where thousands of these events could occur. In addition, epileptologists have a high level of inter- and intra-rater variability. How likely are they to call a single second event “abnormal” compared to their colleagues or even to themselves later in the future?
We have eight epileptologists for our spike detection who labeled hundreds of thousands of examples of spikes. Then we trained a deep learning model to learn and represent the consensus of these eight epileptologists. Now, we can deploy this algorithm on every single second of data and say it’s equivalent to having a whole roomful of epileptologists vote on every second of the whole 24-hour EEG. It’s exciting that we can now have a more quantitative look at the full EEG recordings and get new endpoints such as 24-hour spike burden, spike counts or spike amplitude. All these features can be derived by using these deep learning models on the EEG. Then, we use our traditional digital signal processing methods when appropriate. You need the right tool to give the right inputs and labels for the right question.
I think we’re the first company to engineer, discover and deploy EEG biomarkers. People have been publishing machine learning for EEG over the past decade. A few big things have made this moment in time different. The first is the datasets. As a company, we don’t believe anything if there are under 100 patients used in training and tests. That is not a big enough sample to represent the diversity of real patients. We like when our models are trained on 10,000 patients. The first differentiator is that we’ve now hit a moment in time where groups like ours have aggregated high-quality datasets and taken the arduous work of aggressively labeling them. We already know that when that work is done in places like pathology and radiology, the field sees massive benefits through improved clinical workflows and augmented clinical development. The rise of machine learning infrastructure and cloud computing helps us do training and inference at scale. The confluence of these two has enabled us to bring EEG alongside partners in digital pathology, radiology and GI medicine, where Phase 3 endpoints are already using these types of quantitative tools. EEG is just a short ride behind.
DDD: What is the main way you think that machine-learning enabled EEG data can inform drug development?
Donoghue: When you develop these new therapies, you often rely on the patient to give a subjective report of how they’re feeling. These amazing new precision biotechnologies are emerging. We’re benefiting from AAVs and gene therapies burgeoning in the oncology space. Now, we’re taking the platforms and applying them to the brain.
Now that we have these precision medicines, we need precision measurements. We can’t just rely on these like subjective snapshots.
Quantitative physiology studying brain activity allows us to see the pharmacokinetic modeling and simulation of the drug and how it’s influencing brain activity, which supports what we care about the most in feeling and function.
We’re bullish that this will continue to accelerate as more and more of our pharma partners incorporate EEG into their trials. In some areas of medicine, traditional endpoints for the approval of new therapies were traditionally subjective based on the patient-reported diary. That has largely been overcome with more quantitative measures.
We’re very excited for a future where we no longer have to burden the patients to report these outcomes, but we can get real data on how the brain responds. I think we’ll see neuro look a lot like other therapeutic areas where the endpoints in a Phase 3 trial can be approved based on quantitative biomarkers such as epileptiform spike burden or EEG seizure burden.
Filed Under: clinical trials, Drug Discovery, Drug Discovery and Development, Neurological Disease, Psychiatric/psychotropic drugs