Seizure Forecasting Using Machine Learning & AI - Part 2 - Ben Brinkmann, Mayo Clinic, USA
Hear about developments in seizure forecasting (using AI & machine learning), the benefits to people with epilepsy, trust, and managing expectations. With Clinical Support Scientist at the Mayo Clinic: Ben Brinkmann.
Find Epilepsy Sparks Insights episodes on YouTube, Spotify, Apple Podcasts, Google Podcasts, Amazon Music, Audible, and RadioNorge.
Reported by Torie Robinson | Edited and produced by Pete Allen.
Epilepsy Sparks Insights episodes are meant for informational purposes only, and not as clinical or medical advice.
-
00:00 Torie Robinson
Fellow Homo Sapiens, welcome back to Epilepsy Sparks Insights. Today is part two of two with the fabulous Ben Brinkman: a clinical support scientist (amongst many other things) at the Mayo Clinic, who shares with us his latest in seizure forecasting and remote monitoring technologies, their accuracies and plans, and hopes for the future.
A quick one, please don't forget to like, comment, and subscribe because your comment and like will help spread awareness and understanding of the epilepsies around the world. Now onto our start of the week, or the past two weeks!00:52 Ben Brinkmann
I’m a Clinical Support Scientist with the epilepsy group at the Mayo Clinic, I do a bit of work with mostly our patients who are going towards surgery - these are the people who medications haven’t worked for.
01:08 Torie Robinson
Seizure forecasting is something that is, you know, another passion of yours that you’re very involved with, whether I understand charities, as well as for-profit organisations, and getting everybody together to contribute. Can you just give us a quick insight into what you're doing there, please?
01:22 Ben Brinkmann
You know, one of the things about epilepsy that really makes it a challenge is, of course, people have seizures rarely. It's really difficult to have someone in the office have their seizure so you can observe it. And typically we bring people into the hospital and it's not much fun. They're stuck there for a week or however long. So, I've been really interested in these remote monitoring technologies; whether it’s EEG, whether it's wearable devices. And I think there's an immense amount of promise there. We've done quite a bit of research with non-invasive wearables. There are new devices: the subcutaneous EEG devices. It's just a small electrode that goes under the skin, over the skull, very, very safe to implant. And it's like having a very limited coverage scalp EEG system in place for, you know, a year - you can monitor someone for a year! And you can get someone who has five seizures a year; you can capture all their seizures. And then that information is so incredibly valuable. The other thing, of course, as you mentioned, that's all build up to your question(!), seizure forecasting!
02:34 Ben Brinkmann
We found over the years - through a couple of research systems and a company that nearly made it in the United States - but it was a startup company and they just couldn't get through the regulatory process. But these sorts of ambulatory recorders with invasive electrodes: we've been able to record people's seizures over, as I described, a year, or maybe more. And we found that with that kind of long-term data, you can start running machine learning, AI models on the data, pull out different characteristics of the data. And what we find is that there are very, very subtle changes to the EEG before seizures. And, different groups have studied it differently, but, you know, we think 60 to 90 minutes is probably a good timeframe before a seizure to look. Everybody's a little bit different, and it depends on the person. But, you know, looking at those changes before a seizure: you're able to make a forecast of a seizure. You're able to forecast people's seizures. It's a bit like forecasting the weather! Sometimes you get it wrong. Sometimes you rain. You know, the weatherman says it's going to be sunny and it rains! But you're right more often than you're wrong. So that tells us there's a signal there. It isn't perfect. If we keep working, we'll continue to make it better and better.
So we've extended that work to…we're working with subscalp EEG now. We've got some evidence that we can do seizure forecasting with subscalp EEG. And we were able to do a bit of seizure forecasting with the wearables. We think we need to do much more work there to make it more accurate so people, you know, it's clear there's a signal, but that can still mean that if we tried to launch a system, it would be frustrating for people ‘cause we’d be missing seizures and they'd be getting false alarms. We're still better than random, so we do think this is worth pursuing and we are pursuing it.
04:38 Torie Robinson
And what sort of timescale are we looking at? Do you reckon, I mean, it's probably a really awful question, but would you foresee in the next five years, for instance, where the algorithms are going to be much cooler and we're just going to really see, we're going to be able to predict seizures more accurately, do you think?
04:56 Ben Brinkmann
Absolutely. It's a great question. Our data suggests that maybe on average about 30 minutes of warning we'll be able to give people. And of course that's really important because that's enough time to take perhaps a medication. If you've got a VNS or a you know a DBS system you might be able to swipe a magnet increase your stimulation. Maybe you could prevent the seizure. Certainly enough time to call someone you love and say hey can you just stay with me make sure I'm okay…
05:23 Torie Robinson
Or if you're at work for instance, you say, oh, oh god, this is, this is going to happen. All right, cool: I'll go and, you know, the medical room or whatever, or I'll tell my colleagues where I'm going to go, what's going to happen, why I'm not going to be there, da, da, dah. And yeah, gosh, that could be horrifically empowering.
And what we were talking about though, accuracy and seizure prediction: this is, as you previously mentioned, is a bit of an issue at the moment! And it's actually, from my personal experience speaking to people, it's one of the key things that really annoys people. It's like:, they might get an alarm and it's not gonna happen. And that can stop people from actually using seizure forecasting devices (if it's obviously, if it's not an implant). So where do you think we are? How accurate do you think in say five years time, these devices are gonna be in predicting seizures?
06:09 Ben Brinkmann
I think we'll continue to get better. I think we can get to, say maybe 90% accuracy - so we'll miss a few seizures and we'll give you a few false alarms. But of course, I think one of the great ways forward is to pair it with a therapy or something that you can do. And in those cases, you'll be able to intervene. And if, I guess if you, for example, a false alarm: if you have a warning and you take a medicine, did the medicine work or was it a false alarm? I mean, maybe you don't care (!) because you didn't have a seizure and you're happy about that!
You know, the other thing about the timeframe which is really important: there's new evidence that's coming out (and this is is work that's been done by our friends in Australia and Switzerland and San Francisco); but looking over very, very long time periods, seizures are cyclical. You might be able to look at your seizure pattern and your EEG pattern and your pattern with wearables - and that was something we looked at - and found that for many people it was true. And that's really empowering because that might be able to tell you, okay, your seizure cycle is around 11 days. So, you had a seizure; 11 days from now, you're probably at high seizure risk. So, we wouldn't say it's an alarm and you need to, maybe you take a medication (obviously talking to your physician, of course, or have some intervention), but you might just say, gosh, I'm going to work from home that day.
07:47 Torie Robinson
Yeah, or I'm not going to go out on the lash - go to a party; I'm actually going to stay at home, chill out, watch some really, really bad Netflix or Amazon Prime or whatever…
07:55 Ben Brinkmann
Exactly!
07:56 Torie Robinson
And have a nice glass of mineral water and chill, you know, rather than putting yourself at risk, maybe.
08:03 Ben Brinkmann
Absolutely, absolutely. Yeah, and those kinds of things people can do and it's no risk to modifying your activities and just being more careful on certain days. So that's really, really a promising technology.
08:15 Torie Robinson
And this is something you have to keep us all up to date then because it's really exciting stuff. And I love your honesty as well: you're not saying “Okay, darlings, you know, tomorrow we're going to have something that predicts 100% of your potential seizures, you know, life is perfect”! That's not going to happen. But, we very, very much appreciate your, yes, you being frank with us, nothing's 100%. But this is how it can, where we're likely to be, and how these things can potentially improve quality of life - by giving people more control over their lives, I guess. Especially if an alarm is 30 minutes prior to a seizure. That's amazing.
08:49 Ben Brinkmann
Absolutely. Absolutely, and as a scientist, we're wrong as often as we're right! So it breeds humility and I think, you know, just I'm glad, I really want that people to hear that message too. You know: don't think that because your seizure forecaster/whatever device you have tells you that you're at low risk, that you should get in your car and drive or do something that's unsafe because we're not that good! The technology is not that good! Will it ever be that good? Gosh, I hope so. I just don't know, though. I really don't know.
09:18 Torie Robinson
Thank you to Ben for his honesty and humility as a scientist. This is sometimes not valued enough. Do keep your eye on Ben for his developments in AI, machine learning, and the benefits to people affected by the epilepsies. Do also check out his profile on torierobinson.com, where you can also find the transcription of today's episode.
Again, if you haven't already, don't forget to like, comment, and subscribe, and see you next time. -
Ben Brinkmann is associate professor of neurology and a clinical support scientist for the Division of Epilepsy at Mayo Clinic. This unique role as an engineer embedded in a clinical practice area allows him to help patients and informs the focus of his research. His research efforts are directed toward developing seizure detection and prediction capabilities using noninvasive and minimally invasive biosensors, and improving the accuracy of pre-surgical seizure evaluation using novel methods for image and neurophysiology analysis.
Ben has developed and implemented enterprisewide analytics supporting pre-surgical epilepsy evaluation and systems for statistical processing of ictal single-photon emission computerized tomography (SPECT) images. He has also used and applied methods to verify stereotactic electroencephalogram (EEG) electrode placement for patients with epilepsy.
Ben collaborates with the Bioelectronics Neurophysiology and Engineering Lab at Mayo Clinic.
Ben’s focus areas are:
Noninvasive wearable devices for seizure detection and forecasting
Machine-based learning and signal-processing methods for seizure forecasting using ambulatory intracranial EEG with applications to neuroimaging studies to identify abnormal and potentially epileptogenic brain regions
Detection, analysis and mapping of high-frequency oscillations from chronic and intraoperative intracranial EEG recordings
Ictal SPECT image processing and analysis in seizure localization
Voxel-based morphometric analysis of magnetic resonance imaging (MRI) for localization of structural abnormalities in epilepsy
Volumetric functional and structural image analysis
Significance to patient care
The goal of Ben’s research is to go beyond predicting the probable occurrence of a seizure to the actual delivery of therapy to prevent the seizure. These next-generation epilepsy management and therapy platforms will have an inestimable impact on the quality of life of patients with epilepsy. Prevention of seizures will allow patients to pursue normal activities of everyday life, such as driving. More importantly, it will counteract the comorbidities of epilepsy such as depression, impaired cognition, and sleep disturbances.
Professional highlights
Voting member, Working Group (WG)-32: Neurophysiology Data, Digital Imaging and Communications in Medicine, January 2020-current
President and CEO, 3D Medical Imaging, Inc., Byron, Minnesota, 2006-2015
-
Twitter: BenHBrinkmann
LinkedIn: benbrinkmann
Mayo Clinic: brinkmann-benjamin-h
ResearchGate: Benjamin-Brinkmann
Google Scholar: https://scholar.google.com/citations?user=ax95gtwAAAAJ&hl=en
ILAE: benjamin-brinkmann