'Spike' Detection Program Promising to Diagnose Epilepsy on EEG

October 28, 2019

A new automated computer algorithm for the diagnosis of epilepsy from electroencephalogram (EEG) readings ― developed by analyzing multiple expert opinions on thousands of such readings ― has shown better diagnostic accuracy than individual human experts.

The algorithm, known as SpikeNet, should help with diagnostic testing for epilepsy and warn of clinical decline in critically ill patients, particularly in settings in which EEG expertise is not available, say the researchers who developed it.

The process involved in creating the algorithm is described in two articles published online in JAMA Neurology on October 21. The researchers were supported by grants from the US National Institutes of Health, the American Academy of Neurology, and the Glenn Foundation for Medical Research.

"Interpreting an EEG and identifying interictal epileptiform discharges (IEDs), also known as 'spikes,' that are used to diagnose epilepsy is far from straightforward, and even experts in reading EEGs can disagree on what is and isn't an IED," the senior author of both articles, Brandon Westover, MD, Massachusetts General Hospital, Boston, told Medscape Medical News.

"To add to the problem, most people do not have access to an epilepsy specialist, and most EEGs are not read by such specialists. In the US, three quarters of EEGs will not be interpreted by an expert, and in the rest of the world the numbers are even higher," he noted.

This means the diagnosis of epilepsy is not being made correctly a lot of the time. "Patients can be wrongly diagnosed with epilepsy and can take medication inappropriately for years with the accompanying adverse effects and the real underlying problem never diagnosed, or their epilepsy can be missed so they don't receive treatment and have seizures that could have been prevented," Westover said.

"We wanted to make an automated test which harnesses the knowledge of a range of experts in a computer program which can be used for the diagnosis of epilepsy from EEG readings without an expert actually having to view the EEG themselves. This could be thought of as a general purpose spike detector," he explained.

In the first article, the researchers report a study in which they determined how well experts were able to detect EEG abnormalities indicative of epilepsy and how often they agreed or disagreed as to the presence of such abnormalities.

They devised a system in which eight experts ― all with at least 1 year of subspecialty training in clinical neurophysiology ― independently analyzed 1051 EEGs from patients with different medical problems. The EEGs contained a total of 13,262 abnormalities that may or may not have been IEDs.

"These experts are the gold standard ― they have specialty training on reading EEGs ― and we had not just one of these experts but eight different experts analyze each EEG reading," Westover noted. All the information and the experts' scores were recorded on a computer program that determined which patterns led the experts to agree and which ones led to disagreement.

Results showed that the mean percent agreement on individual EEG abnormalities was 72%. "That means that 72% of the time, the experts agreed as to whether an individual EEG pattern was a spike or not," Westover said.

For a second task ― which was regarded as more clinically relevant ― each expert was asked to look at the full EEG (lasting about 30 minutes) and decide whether there were any IED spikes in the whole reading. For that task, results indicated an agreement of 80.9%.

"While this is pretty good, it still shows how difficult it can be to diagnose epilepsy in some cases. Given that eight experts did not agree whether or not the EEG was showing an IED in 20% of cases, this shows that diagnosing epilepsy is far from a slam dunk, and this is what we also find in clinical practice," Westover commented.

The second article describes the development of a computer program ― SpikeNet ― using the data generated in the first article.

The accuracy, sensitivity, and specificity of the SpikeNet algorithm was compared with fellowship-trained neurophysiology experts for identifying IEDs and classifying EEGs as positive or negative for IEDs.

Results showed that for individual EEG abnormalities, the mean calibration error for experts was 0.183, compared with 0.066 for the industry-standard commercial IED detection algorithm (Persyst 13) and 0.041 for SpikeNet. For classification of whole EEGs, the algorithm had a mean calibration error of 0.126, vs 0.197 for the experts.

"This computer program can perform the task of spike detection and determine which EEGs have IEDs present and which don't," Westover explained. "The program has been generated by learning from the opinions of eight experts on thousands of EEG abnormalities, so it can detect an IED better than any individual expert. The program will also produce a probability of what percentage of experts would have said each pattern was an IED or not."

Westover acknowledged that developing such a computer program may seem like an obvious thing to do, and some people may question why it hasn't been done before. "It comes down to the fact that no one has gone to the trouble before of creating enough labeled examples," he said. "That is the key. It has taken us 5 years to collect enough labeled data to train a computer to make the diagnosis reliably."

The researchers are now trying to make a commercial algorithm. "It is not available yet, but we are hopeful it will become available next year," Westover said.

He believes that once the algorithm becomes available, in the short term, the workload of EEG experts may increase, because the EEGs would still need to be reviewed by an expert. This could be done through remote EEG reading services. But in the longer term, the computer program could replace the expert in reading EEGs.

"It is my expectation that patients will get more accurate and quicker diagnosis, regardless of where they live and whether they can access an expert center or not. It will then free up the experts to manage the more complicated cases," Westover suggested.

"Of the 50 million people worldwide who have epilepsy, 30 million do not have access to a neurologist, never mind an epilepsy specialist," he reported. "This algorithm has the potential to vastly improve the management of these patients."

The researchers were supported by grants from the US National Institutes of Health, the American Academy of Neurology, and the Glenn Foundation for Medical Research. One of the coauthors reports being issued US patent 10,349,888.

JAMA Neurol. Published online October 21. Article 1, Full text; Article 2, Full text

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.
Post as: