For the roughly 50 million people worldwide with epilepsy, the exchange of electrical signals between cells in their brain can sometimes go haywire and cause a seizure—often with little to no warning. Two researchers at the University of Louisiana at Lafayette have developed a new AI-powered model that can predict the occurrence of seizures up to one hour before onset with 99.6 percent accuracy.
“Due to unexpected seizure times, epilepsy has a strong psychological and social effect on patients,” explains Hisham Daoud, a researcher who co-developed the new model.
Detecting seizures ahead of time could greatly improve the quality of life for patients with epilepsy and provide them with enough time to take action, he says. Notably, seizures are controllable with medication in up to 70 percent of these patients.
Daoud and his colleague Magdy Bayoumi are by no means the first people to explore ways to predict seizures. Other research groups have worked on ways to analyze brain activity using electroencephalogram (EEG) tests and have used the data to develop predictive models. However, each person exhibits unique brain patterns, which makes it hard to accurately predict seizures. Previous models were designed to do this in a two-stage process, where the brain patterns must be extracted manually and then a classification system is applied, which Daoud says adds complexity to the model.
In the new approach, described in a study on 24 July in IEEE Transactions on Biomedical Circuits and Systems, the features extraction and classification processes are combined into a single automated system, which enables earlier and more accurate seizure prediction.
Furthermore, the researchers incorporated another classification approach whereby a deep learning algorithm extracts and analyzes the spatial-temporal features of the patient’s brain activity from different electrode locations, boosting the accuracy of their model. And finally, EEG readings can involve multiple “channels” of electrical activity, so Daoud and Bayoumi applied an additional algorithm to identify the most appropriate predictive channels of electrical activity; this also speeds up the prediction process.
The researchers developed and tested their approach using long-term EEG data from 22 patients at the Boston Children’s Hospital. Although this is a small sample size, the results proved exciting for the team. Not only is their model very accurate, at 99.6 percent, but it also has a low tendency for false positives, at 0.004 false alarms per hour.
The system does require some setup before it can produce such results. “In order to achieve this high accuracy with early prediction time, we need to train the model on each patient,” says Daoud, noting that training could require a few hours of non-invasive EEG monitoring around the time of a seizure, including during the seizure itself. “This recording could be [done] off-clinic, through commercially available EEG wearable electrodes.”
With the software component complete, Daoud says the next step is to develop a customized computer chip to process the algorithms. “We are currently working on the design of an efficient hardware [device] that deploys this algorithm, considering many issues like system size, power consumption, and latency to be suitable for practical application in a comfortable way to the patient,” he says.