Melanie Cole (Host): Researchers at the University of Alabama at Birmingham have developed an algorithm that may predict when patients with seizure disorders such as epilepsy might be at risk for a cluster of seizures. Welcome to UAB Med Cast. I'm Melanie Cole, and joining me today is Dr. Adeel Ilyas. He's a resident in the Department of Neurosurgery in the UAB Marx e Hearsy School of Medicine, and first author of the study published in the journal. Dr. Ies, it's a pleasure to have you with us. This is such an interesting study. As we get into this topic, please tell us a little bit about epilepsy and the seizure aspect of this condition, the prevalence of it in this population of affected patients, and up until now, could we predict when or how severe a seizure would be?

Dr Adeel Ilyas: First of all, thank you so much. It's a pleasure to be here. and I'm really interested in epilepsy and understanding, predicting, seizures and how they happen and how, often they occur. And one of the, aspects of epilepsy, which is really interesting, is that seizures occur in clusters. And this is something that we've, known about and it's, shown in the literature in different studies. but there is no kind of universal, or maybe a better definition is there's no, mathematical or probabilistic definition for what constitutes the cluster. And the reason why seizure clusters are so important is because the notion is that seizure clusters represent a sort of heightened risk, to have seizures.

In other words, the brain is in an elevated state where it, has a higher propensity to produce seizures and therefore patients have, clusters of seizures. The analogy would be like lightning for example. as individual events, lightning strikes are pretty random. But when you think about it, lightning in general isn't really that random. In the sense that it's more likely to occur when there's a thunderstorm, if there's a clear blue sky, it's unlikely that there'll be a lightning strike compared to when there's a, thunderstorm occurring.

So maybe there's something about the brain where there are certain brain states that promote seizures in which seizures are more likely to occur, and other states in which seizures are relatively unlikely, quite random. So with this background or this notion, we try to develop a way to capture, seizure clusters, to define them and then forecast them and see if we could predict when they would occur in the future.

Melanie Cole (Host): That's so cool Dr. Ilyas. So tell us a little bit about not only the complications, but how would reliable seizure cluster prediction immensely benefit individuals with epilepsy? How would knowing that this patient is at an increased risk for this cluster allow you as medical professionals to intervene to reduce that risk whether it's increased monitoring, adjustments to medication, tweaks to the simulator, whatever it is?

Dr Adeel Ilyas: So, we'll maybe talk about what constitutes the seizure cluster and the definition of the seizure cluster a little bit later. But the idea here is that, when the brain, let's say, is in this heightened seizure cluster state, then there are certain, things that we know about patients who, have seizure clusters or are more prone to have, clusters or seizures. And for example, one thing that we know about them is that they tend to be, treatment resistant, whether that's with medical therapy or with, surgery or stimulation, neuromodulation. They tend to have a resistant to, or be resistant to treatment.

They also tend to have an increased risk of going into a state of status epilepticus, where the brain basically is in a seizure mode for an extended period of time, and that is quite dangerous. Seizure clusters are also associated with, increased, hospitalization risk and with sudden death in epilepsy. this is an very unfortunate, event that can happen in patients with epilepsy, it's a feared complication where, patients, are found dead basically, and presumed after having had a seizure.

Melanie Cole (Host): Wow. So it can be really, really detrimental to the health of the patient. So then speak a little bit about what you've done, this algorithm and how researchers at the University of Alabama at Birmingham have developed this method to predict when patients with Seizure disorders such as epilepsy might be at risk for these clusters? This is fascinating.

Dr Adeel Ilyas: The first hurdle that we had to overcome is we had to first define what a seizure cluster is and in the literature there are different definitions and you can make the argument that's many of them are arbitrary. Three common definitions. This is based on a, meta-analysis, some of the authors of whom are also part of a UAB. So one definition is, having. Three or more seizures in a 24-hour period. Another definition is having more than two seizures in a 24 hour period. And yet a third is, more than two seizures in a six hour period.

Like this, there are other, definitions of what constitutes a seizure cluster, but the notion is that a seizure cluster happens when seizures occur. Earlier than expected, or they occur kind of back to back, in rapid succession. So that's sort of the general notion of, a seizure cluster. So we devised a probabilistic or and mathematical definition of seizure clusters based on, Cole Agora statistic. So the idea here is that you input the seizures as events in a time series, and you sort of look at the distribution of these events and compare that distribution to an expected distribution or an alternative or expected distribution.

As if the events were equally spaced. So in other words, the events that you're seeing, in actuality compare them to events that were equally spaced out. And the difference between the observed events versus the expected equally distributed events, can be computed or calculated as a probability. And that basically, is how we define our seizure clusters, a probabilistic definition of how the seizures are occurring more rapidly as compared to an equally distributed, or equally spaced out distribution.

Melanie Cole (Host): So how have been the outcomes? And up until now, what have you been able to forecast? I mean, identifying this window two and a half days prior to the onset of these clusters and giving you really time to establish an intervention plan. How's it been working now?

Dr Adeel Ilyas: This study we did in 15 patients who had a specific type of, device called, a responsive near stimulator, RNS device that is made by a company called NeuroPace. And what this device does is it detects. Seizures in the brain, electrical activity in the brain, and can then stimulate the brain to reduce or modulate, seizures over a period of time. But this detection of seizures then can be recorded over a long span of time. So in our study, in the 15 patients that had this RNS device, we recorded electrographic activity on average, about 30, 39 months or so of data that we had on these 15 patients.

And across that span of time, we looked at the seizure events, and kept track of when the seizures occurred. And then we applied our probabilistic, algorithm to compute seizure clustering in these 15 patients. And then we, use what's known as a generalized auto aggress model. Basically a model that can predict the fluctuations and probabilities of the seizure clustering, and then use that algorithm to forecast future events. And so the way we did that is, let's say we had, 100 events that we captured in one patient. So 100 seizures.

We divided those 100 seizures and took the first 75% of the data. So 75 seizures, trained that auto aggress model on the first 75 seizures, and then use that model to predict the remaining 25 seizure. So trained on 75% of the data and predict on the remaining 25% of the data, and then we computed how accurate our predictions were. And so on average across the 15 patients, if we predict out to one day, we had really good accuracy. So computed the accuracy based on what's called a may score or a mean absolute scaled error.

And a value less than one indicates that the algorithm performs better than chance and our values were, 0.78 on average for a one day time horizon. That means that on. across the 15 patients. we could predict the seizure clustering the probability of seizure clustering out to one day, and we did that subsequently for two and a half days and so forth. But the accuracy gets worse as you, forecast out further and further, from the last known time, as you increase the horizon.

Melanie Cole (Host): So where are you now, Dr. Ilyas, in this research and for other providers that are listening, as interested as I am with this, how can they be able to use this research? Take us from bench to bedside, and how can they get involved?

Dr Adeel Ilyas: So I would say, first of all, this study is a pilot city or feasibility study in 15 patients. So the next step, for us would be to validate these findings in a larger cohort of patients. And assuming that we're able to validate this and to show that, Algorithm works, in general for a large number of patients with epilepsy, then it can be very useful in potentially developing an app or something like this where patients can be informed of, impending seizure risk. In this way, either they can modify their own medications or therapy that they're receiving or they have a plan to do so, or their providers can, intervene appropriately.

They can adjust their daily activities accordingly. There's a lot of therapeutic potential in seizure forecasting . It may allow us to investigate further, regarding brain states that promote seizures or increase the propensity of having seizures. In general, the notion is that epilepsy or seizures are random. but research like this and there other research out there as well is sort of challenging that notion and suggesting that seizures may be partly random, but also can be predicted, to certain extent and with certain probabilities.

Melanie Cole (Host): As we wrap up, this is just an amazing study that you're conducting. What would you like other providers to know about this algorithm that may predict the onset of seizure clusters in patients with disorders such as epilepsy, and anything you'd like them to know about this study as we summarize?

Dr Adeel Ilyas: I'd like providers to know that, our algorithm is a promising algorithm and it seems to be, a practical algorithm that can, forecast seizures, at least in this small cohort, effectively. And if a larger validation study shows similar and promising results, then it really, brings about this notion that seizures may be able to be forecasted to some extent, at least in a probabilistic way. And if that's true, then it means that we can deliver therapy based upon a projected or forecasted seizure risk. And that can really change the field of epilepsy because knowing that a patient is going to have an increased seizure risk in the near future might actually allow us to deliver.

Or at least inform the patient that there's an increased risk that can, affect how the patient, he or she conducts his daily life, but then also deliver additional medication, perhaps change neuro stimulation settings. In this case, these patients all had, the RNS device, so those stimulation settings can be adjusted. But any sort of intervention can be done based on the understanding that there's an increase or impending, seizure risk. And I think this is a really, novel and, very, beneficial, practical application of this work, if it's validated in a larger study.

Melanie Cole (Host): It has very far reaching implications. It's so interesting, and I hope Dr. Ilyas, you'll come back and join us and give us an update as this research continues. Thank you so much for joining us and explaining this all to us, and a physician can refer a patient to UAB Medicine by calling the missed line at 1-800-UAB-MIST. Or by visiting our website at uab medicine.org/physician. That concludes this episode of UAB Med Cast. I'm Melanie Cole.