PHILADELPHIA — Real-time data can help train a machine-learning algorithm to detect seizures prior to clinical symptoms, results from a small study indicate. The data were presented at the American Epilepsy Society Annual Meeting in Philadelphia.
Previously, few technologies have been able to appropriately warn patients of an oncoming seizure due to the great variation in seizure onset patterns and seizure location.
“The main advantage of our dynamic algorithm is the use of a novelty detecting machine-learning algorithm that constantly adapts to the variation in brain activity of the patient,” Daniel Ehrens, study author and PhD candidate at Johns Hopkins University, told Neurology Advisor. “This allows the algorithm to have a very robust behavior that is not affected by fluctuations in brain activity like arousals, medication withdrawal, and daily variations. It can also extract more advanced information from the patient’s own EEG activity.”
Ehrens and colleagues used 5-hour EEG recordings that ended in seizure from 5 patients with focal epilepsy who were undergoing presurgical evaluation. Utilizing features previously known to be helpful in seizure detection, including spectral parameters and complexity and entropy measures, the researchers implemented a novel detector that trained on 20-minutes windows of features and shifted every minute.
In all of the participants, seizures were detected within a delay of 0 to 4 seconds. There was a false positive rate range from 0.11/hr to 0.5/h.
“Tuning the detection method to the patient’s characteristics is a critical part of seizure detection, and we are now pushing the method to mimic the process that experts used when reading data from a new patient: a combination of experience and ability to recognize novelty,” Ehrens said.
The adaptive algorithm allowed for the detection of “previously unseen seizure events without prior training of the ictal signature,” the authors wrote, indicating that the algorithm could allow for real-time detection of seizures.
“Currently, the detector has a very generic behavior, where it is able to detect many types of seizures from different types of patients using a large number of channels,” Ehrens said. “False positives are given by epileptogenic activity or paroxysmal events that don’t always indicate a seizure; however we believe it is still important to look at those events for characterization purposes in the epilepsy monitoring unit.. Further down the road, the classifier will learn to recognize the various types of novelty events and be able to classify their importance and react accordingly.”
“Our plan is to make this detector more adaptive to each patient, capable of recognizing known seizure patterns and new ones and capable of helping neurologists get better and faster answers.”
For more coverage of AES 2015, go here.
Ehrens D, Cervenka M, Bergey G, Jouny C. Abstract 3.092. Dynamic Training of Machine-learning Algorithm for Real-time Seizure Detection in the Epilepsy Monitoring Unit. Presented at: American Epilepsy Society Annual Meeting; Dec. 4-8, 2015; Philadelphia.