65 million people worldwide have epilepsy, with 250,000 of those living in Australia. And March 26 is Purple Day – A day in the spirit of increasing awareness about epilepsy globally. This year’s Purple Day is just a little different as we have recently made steps forward in research using artificial intelligence to predict warning signs of seizures.
Imagine a future where technology can understand the individual brain patterns of people with epilepsy and alert them to signs which may indicate an oncoming seizure.
Well, scientists in our IBM Research lab in Melbourne have recently shown how this is possible. It’s one of the many projects underway by IBM globally as we look to positively transform the world through technology.
At the end of last year, Isabell Kiral-Kornek and Subhrajit Roy were the lead researchers publishing in Lancet’s EBioMedicine, detailing progress in personalised seizure prediction using deep learning and brain-inspired computing. Using AI deployed on an ultra-low power mobile device to predict seizures has never been done before, so I’m excited to see how the project evolves from here.
I sat down with Isabell and Subhrajit recently to talk about the project, what inspires them, and how they hope to see it develop in the future.
Q: What excited you most about this research project?
Isabell: For me, it’s undoubtedly the opportunity to change the lives of people living with epilepsy.
Seizures strike at unpredictable times, resulting in an unimaginable burden for people living with epilepsy. Technology helps look at this issue differently and is the perfect opportunity to apply new approaches to analysing brain signals.
Subhrajit: Alongside the potential real-life impact, we were fortunate to be able to work with a dataset completely unique to anything else in the world.
Often when addressing important health issues, we have limited data to work with. This project was different. We were lucky enough to access longitudinal EEG recordings spanning up to three years for individual patients, thanks to an earlier study by our research partners the University of Melbourne and St Vincent’s Hospital. This means our algorithms can be demonstrated in a real-life scenario by seeing how a person’s brain signals change over time.
Q: What was the goal of your epilepsy research?
Isabell: It was important to us that we created a solution which could allow users to maintain control over the sensitivity of a seizure prediction device.
This vision was central to our research, which simulates live monitoring of brain waves every 30 seconds to make a prediction as to whether a seizure is likely. Using this knowledge, we developed a strategy for patients to effectively set an alarm threshold. For example, this means while sleeping you would be able to select low sensitivity so you wouldn’t be alerted to every possible seizure, however when walking to the shops, you would want to be alerted to every scenario for safety reasons.
Q: How do you train a machine to understand brain waves? What steps does it entail?
Subhrajit: While traditional computer systems are programmed by instructions, machine learning systems are trained by examples. In our case, we extracted the periods before a seizure and normal non-seizure signals from the brain recordings and presented them to a neural network.
When presented with different types of signals, a neural network can be trained to discriminate between them, which was in our case between pre-seizure and normal brain states.
An example illustration of research system monitoring and measurement
Q: Was there a point in this research where you thought ‘we’ve cracked it’ – kind of the ‘aha’ moment.
Isabell: We had many moments where we thought we’d ‘cracked it’. Some of them turned out to be flukes – we had missed something that made our results at the time look better than they were. And some turned out to be actual breakthroughs which led to the exciting results we published.
One surprising finding was uncovering that a person’s brain signals change over time, which meant that an adaptable/learning system would be critical in any seizure prediction capability.
Q: Where does this go next?
Subhrajit: The project is going down two different paths.
One is to continue improving our seizure prediction system by bringing in other factors like weather, biometrics and identifying the types of seizures occurring.
Secondly, through our relationships with clinicians, we want to identify other problems in the epilepsy space and explore how AI could be applied to address these problems.
What do you enjoy most about working at IBM?
Isabell – Partnership and teamwork are one of the best things at IBM. It creates a special atmosphere. I could go to anyone in our research lab and walk away with new knowledge and inspiration. People give the time to teach, listen, and learn.
Subhrajit – I agree it’s collaboration. We get to work with more than 3,000 researchers and scientists spread out across 12 global labs on six continents – including Australia. That’s unusual and so valuable to a researcher.
Author: Bhavna Antony, IBM Research Scientist Australia Many eye diseases that cause irreversible blindness are ones that develop slowly, showing little to no sign of vision threat until it is too late. Diabetic retinopathy and glaucoma are the leading and second leading cause of blindness worldwide, respectively. They currently affect 350 million individuals across the […]
Subhrajit Roy works as part of the IBM Research Australia team in Melbourne. On 2nd April 2019, he was recognised as one of Forbes’ annual ’30 Under 30′ honorees for Asia and the Asia Pacific for innovation in healthcare and science. Subhrajit has worked as a full-time researcher in IBM Research since October 2016 and pioneered the use of Big Data and AI for […]
For over 80 years, IBM has been working to solve some of the biggest issues facing Australia and New Zealand. Today IBM is helping doctors diagnose disease, predicting the latest fashion trends and creating better services for citizens.
These are our stories; this is IBM.