Noninvasive Approach Detects Epileptiform Activity

Pauline Anderson

May 06, 2022

A computer algorithm can noninvasively detect hippocampal epileptiform activity (HEA) using only information from a standard scalp EEG, new research suggests.

This machine learning approach may improve diagnosis of epilepsy and help guide decisions surrounding surgical interventions, researchers note.

Dr Alice Lam

"The hippocampus is a super important part of the brain when we think about epilepsy, but we don't have great ways of monitoring activity from the hippocampus that are noninvasive," study investigator Alice D. Lam, MD, PhD, assistant professor, Department of Neurology, Massachusetts General Hospital, and Harvard Medical School, Boston, told Medscape Medical News.

"This algorithm allows us to monitor that activity in a noninvasive way, which is something human experts can't do," Lam added.

The findings were published online May 2 in JAMA Neurology.

A Novel Approach

Scalp EEG is used to assess the brain's electrical activity, but is relatively insensitive to such activity arising from deep brain regions, including the hippocampus.

In the diseased state, the hippocampus is highly prone to generating abnormal spiking activity, which can cause memory impairments and psychiatric disturbances, and signal impending seizures.

Currently, detection of HEA requires an invasive procedure where intracranial electrodes are surgically inserted into or adjacent to the hippocampus, where they can directly record electrical activity.

The investigators developed and validated a deep-learning algorithm called HEAnet in order to noninvasively detect HEA using only information extracted from a standard scalp EEG recording.

To do this, they used datasets derived from a convenience sample of 141 participants. Of these, 97 had temporal lobe epilepsy (TLE), whereas 44 did not have epilepsy and acted as the healthy controls group (HC).

One dataset included the simultaneous scalp EEG and intracranial electrode recordings of 51 patients with TLE (mean age, 40.7 years; 59% men).

Better Than Humans?

Results showed the computer algorithm could classify positive from negative HEA examples, achieving a mean area under the curve (AUC) receiver operating characteristic (ROC) of 0.89 and mean AUC precision recall (PR) of 0.39.

"Thus, HEAnet accurately detected HEA from scalp EEG at the single-event level, a task that human experts cannot perform," the investigators note.

As well as single-event detection, the algorithm could also detect deep brain activity over a period of hours, Lam added.

Researchers trained the detector to assess scalp EEG recorded during sleep. This was done because epileptiform discharges are more frequent during sleep in patients with TLE and there is much less artifact "or noise" on recordings when people are asleep, said Lam.

The algorithm also performed reasonably well on awake EEG data, maintaining excellent specificity — although with reduced sensitivity.

Investigators validated the performance of HEAnet on two independent scalp EEG datasets. One included 24 patients with TLE and 20 HCs (mean age, 42.3 years; 61% women), and the other included 22 patients with TLE and 24 HCs (mean age, 43 years; 57% women).

The algorithm accurately distinguished the TLE group from the healthy controls group, with an AUC ROC of 0.88 for the first dataset and 0.95 for the second. It also predicted epilepsy lateralization — with an accuracy of 100% and 92%, respectively.

Independent Information Source

The study showed that combining human expert review with the HEAnet increased the sensitivity of diagnosing TLE, the researchers note.

The findings also showed that HEA-net can detect changes in HEA in response to medication adjustments, they add. Patients in the epilepsy monitoring unit are gradually taken off seizure medications in order to record seizures on EEG and then gradually put back on medications at the end of the monitoring session.

"It seems the algorithm has enough of a dynamic range that it can actually track how people's hippocampus spike rates respond to medications," said Lam.

Although the algorithm may not totally replace intracranial recordings, "it provides us with an independent source of information about whether or not a patient is having these discharges and what side someone's seizures might be coming from — the right or left hippocampus," she added.

The researchers foresee the algorithm eventually being used for earlier diagnosis of TLE and better monitoring of brain activity in relation to medication status. It also might be used to identify the source of memory issues that patients with epilepsy often complain of, and determine whether reducing HEA with medications improves cognitive function.

HEAnet could also serve as a noninvasive biomarker, independent of scalp EEG epileptiform discharges, to guide surgical decision-making in TLE patients, the investigators note.

However, some "fine-tuning" needs to be done before the algorithm is ready to be applied clinically, said Lam. For example, researchers need to determine how specific the algorithm is for epileptiform discharges that come from the hippocampus as opposed to other nearby areas in the brain.

The HEAnet's performance should also be validated on larger datasets, she added.

"Huge Victory"

Commenting for Medscape Medical News, Daniel M. Goldenholz, MD, PhD, Harvard Beth Israel Deaconess Medical Center, Division of Epilepsy, Boston, Massachusetts, said the study "represents a huge victory" for artificial intelligence (AI).

"I would place a study like this in the category of 'outstanding application of artificial intelligence to the study of epilepsy' because of the careful design, the high accuracy, the multiple layers of external validation, and the fact that fundamentally the AI is accomplishing a superhuman feat," said Goldenholz, who was not involved with the research.

He also praised the exploration of how the algorithm can be used. "This includes diagnosing epilepsy coming from the hippocampus, monitoring for evidence of irritability as medications are withdrawn in the epilepsy monitoring unit, and uniquely identifying the time of spikes, which could have consequences for cognitive tasks," he said.

Goldenholz noted the researchers investigated combining this tool with a human expert to show that AI plus human input achieves higher diagnostic accuracy.

"If this type of tool could be used clinically, it would open up many practical possibilities, including improved detection of epilepsy in challenging cases, additional localization accuracy, and better predictions about driving safety," he said.

He also speculated the tool may allow noninvasive access to other "hidden information" that previously was obtainable only through invasive intracranial monitoring. "Maybe these kinds of approaches may liberate some of our patients from dangerous surgical procedures," said Goldenholz.

The study was funded in part by grants from the National Institutes of Health (NIH) and by the American Academy of Neurology Institute (AANI). Lam reported having received grants from the NIH, the AANI, and Sage Therapeutics; and personal fees from Sage Therapeutics, Neurona Therapeutics, and Cognito Therapeutics outside the submitted work. Goldenholz has disclosed no relevant financial relationships.

JAMA Neurol. Published online May 2, 2022. Abstract

For more Medscape Neurology news, join us on Facebook and Twitter


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.