Many people with epilepsy can recognize the signs when they’re about to have a seizure. But predicting a seizure long before it happens, giving the person enough time to get to a safe place, is still very difficult. Scientists know that seizures are caused by a problem with the electrical signals in the brain, but just what causes a seizure is often not known. If scientists could monitor patients’ brain waves normally and during a seizure, maybe they could detect the small changes that could precipitate a seizure. Now neuroscientists from the University of Melbourne and researchers from IBM are teaming up to create a device to constantly monitors patient’s brainwaves and find predictive patterns for when they will occur, according to Wired and a study that the researchers will present at the ACM Computer Frontiers conference in Italy next month.
The technology at the heart of this project is IBM’s TrueNorth, a computer chip designed to mimic the structure of the brain. The researchers wanted to prove that this system would be useful for epilepsy data, which is their main goal, so they used this study as a proof of concept. A participant completed a series of tasks that involved a left hand or right hand squeeze, all while hooked up to an EEG that was detecting his brainwaves. The researchers used that data to train their deep learning system, then used it to predict which hand the participant was squeezing. The algorithm, they found, was 76 percent accurate. That’s not as high as they had hoped—past trials had generated a model that was 86 percent accurate—but the researchers were hearted that more training data would improve their models.
The study showed the researchers that TrueNorth can predict brain waves. But picking up brain waves from people with epilepsy is a bit more challenging, since they can’t yet predict when a seizure will occur. What they want to do, the researchers told Wired, is implant a sensor in patients’ brains that would gather constant data about their brain waves. If they gathered data over enough time and from enough people, that would provide plenty of data with which to train the model. Ideally that could predict seizures with enough warning so that patients can make sure they’re in a safe place before the seizure starts.
The warning system is still years away, the researchers note. But understanding the electrical signals that go haywire before and during a seizure might someday help scientists correct them, preventing seizures altogether.