WASHINGTON -- Researchers presented new methods for predicting seizures, involving both continuous EEG and machine-learning tools in several posters at the American Epilepsy Society meeting. Others attempted to characterize which patients would take longer to diagnose with seizures on cEEG based on certain characteristics.
Patient-Reported Data for Predicting Seizures
Algorithms developed using data from online seizure diaries were , according to Tobias Loddenkemper, MD, of Boston Children's Hospital, and colleagues.
Action Points
- Note that these studies were published as abstracts and presented at a conference. These data and conclusions should be considered to be preliminary until published in a peer-reviewed journal.
While seizure prediction efforts have typically focused on EEG data, not all patients can provide that level of information. But providing clinical data in an online seizure diary at SeizureTracker.com was more achievable, the researchers said. This program allows patients to enter information about seizure occurrence, including demographic features and the time and duration of seizures.
Loddenkemper's group used machine-learning techniques to train a predictive model and subsequently validated it in an independent testing subset. They compared predictive performance to null predictive models, and included the first 100 seizures in 3,716 patients -- 2,418 in the training subset and 1,298 in the testing subset.
They found better prediction for their various machine-learning algorithms, including robust regression (0.8 days), multivariate adaptive regression splines (MARS) with no interaction (2.6 days), MARS allowing two-way interactions (2.6 days), and linear regression (2.8 days) -- all much shorter than the null model (12.6 days).
"Machine learning algorithms can predict the time of seizure occurrence better than a simple null model," they wrote. "The implication of our results is that time of occurrence of future seizures can be predicted better than chance using clinical data."
A New Machine-Learning Strategy for Predicting Seizures
Very early results from a "wearable" seizure prediction system , IBM researchers reported here in partnership with academics from the University of Melbourne.
The usefulness of a seizure prediction device depends on being able to tailor it to the preferences of physicians and patients, the researchers said. So IBM developed a system that uses intracranial EEG, paired with deep-learning techniques, that can be adjusted to warn about seizures.
Isabell Kiral-Kornek, PhD, of IBM Research Australia, and colleagues studied 10 patients. A deep-learning classifier was trained to distinguish preictal and interictal signals. A second classifier was tested on iEEG data from all patients, and then patients were allowed to tune the prediction system so that its sensitivity or its time in warning suited their preferences.
Overall, the system achieved a mean sensitivity of 69%, with a mean time in warning of 27%, Kiral-Kornek's group reported. After tuning, however, that sensitivity rose to 92%, they said, adding that seizure prediction performance was significantly above chance level for all patients.
They said the study provides a proof-of-concept that seizure prediction alerts can be managed on wearable devices, and that doctors and patients can adjust their warnings to improve accuracy.
More Time Needed to Detect Seizures in Some Patients
It may on continuous EEG (cEEG) in patients with stuprous mental status and seizures secondary to hemorrhage, trauma, and tumors, researchers reported.
Studies have been done to understand how long patients should have cEEG monitoring to exclude seizures, but it's unclear how long it should be continued in different patient populations based on their etiology of seizures and their mental status.
Ifrah Zawar, MD, of the Cleveland Clinic, and colleagues looked at all 2,425 patients who had cEEG monitoring at Cleveland Clinic in 2016. A total of 309 patients had seizures and were included.
The median time to seizure onset was 3 hours, and 39% had their first seizure within the first hour of monitoring. Seizures were detected in the first 24 hours of cEEG monitoring in 80% of patients and within the first 36 hours of monitoring in 90% of patients.
Some etiologies took longer to detect seizures. For instance, for those with hemorrhage, the time to detect seizures in 50% of patients was delayed to 15 hours (P=0.0004) and these patients were more likely to have their first seizure recorded after 24 hours of monitoring (P=0.02).
Patients whose seizures were secondary to trauma and tumors were also more likely to have delayed onset of seizures, as were those with stuprous mental status, Zawar reported.
In these groups, physicians may need to monitor patients longer in order to assess their seizure risk, he concluded.
Disclosures
Kiral-Kornek is an employee of IBM.
Loddenkemper and Zawar disclosed no relevant relationships with industry.
Primary Source
American Epilepsy Society
Loddenkemper T, et al "Machine learning methods for seizure prediction using patient-reported clinical data from a digital diary" AES 2017; Abstract 2.036.
Secondary Source
American Epilepsy Society
Harrer S, et al "A mobile and tunable seizure prediction system using deep learning" AES 2017; Abstract 2.149.
Additional Source
American Epilepsy Society
Zawar I, et al "Time to detection of seizures with continuous EEG: The correlation with seizure etiologies, neurologic status, and EEG findings in a large patient sample" AES 2017; Abstract 3.3063.