Subscribe to eTOC

Data from Electronic Medical Records Helps Identify Epilepsy Symptoms in Patients Before Genetic Testing

​By pulling data from the electronic medical records of people with epilepsy, researchers were able to identify clinical features associated with genetic epilepsies, which they hope may lead to earlier diagnosis and better treatment of those conditions, according to a study presented in December at the American Epilepsy Society annual meeting in Nashville.

The researchers used a machine learning technique—applying customized natural language processing (NLP)—to extract 4.5 million full-text clinical notes from  the EMRs of  more than 32,000 people with epilepsy available since 2010 at the Children's Hospital of Philadelphia (CHOP). The patients ranged in age from 0 to 25.25 years, and 1,925 “had a known or presumed genetic etiology [and] verified causative genetic variants," according to the study.

The researchers were able to find these clinical features much earlier in children than genetic testing did. Their data identified the symptoms in infancy, whereas the genetic testing led to a diagnosis later.

Lead author Peter Galer, MSc, a bioengineering doctoral student at the University of Pennsylvania who has studied genetic epilepsies since 2017, said that generally not enough genetic tests happen at early ages, and that is partially because of clinicians missing clear symptomology.

“Despite this increased availability of technology and knowledge about genetic testing, there still can be—as we've shown by our research—months to years of delay in receiving a genetic diagnosis," Galer told Neurology Today.

He added that the work can be “tremendously impactful" to families of young patients since early diagnoses with these genetic epilepsies can inform treatment.

“There were a lot of clinical features I expected—that we saw in SCN1A, SCN2A—but there were some surprising results," Galer said. “The median delay to median age of genetic diagnosis was 3.6 years, which is pretty striking."

The researchers also found that patients with epilepsy who presented with almost any neurologic developmental abnormality between the ages of 6 and 9 months were approximately five times more likely to receive a genetic diagnosis.

“We also saw that very broad-level abnormalities are highly predictive of receiving a genetic diagnosis," Galer said, adding that patients who experience such abnormalities, particularly in the first year of life, should receive genetic testing.

The research was made possible through the Helbig Lab and Epilepsy Neurogenetics Initiative (ENGIN) at CHOP.

Weighing In

Kristina Julich, MD, assistant professor in the department of neurology at The University of Texas at Austin's Dell Medical School and ​chief of the Pediatric Neurogenetics Center, said the study offers an “important new approach to help identify patients with genetic epilepsies."

Dr. Julich pointed out that the researchers found that children who presented with status epilepticus between ages 9 and 12 months had a very high likelihood of having a pathogenic (or disease-causing) variant in the SCN1A gene​, which is related to a mutation on the sodium channel. Using a sodium blocker could make the condition worse for a patient with that gene, Dr. Julich said; knowing early on whether patients have a particular gene can help doctors determine the best medication for their situation.

“Because genetic factors play a really important role in childhood epilepsies and because of the increased availability of targeted treatment options, it's clinically relevant to identify these kids as early as possible," she said.

In an email, Phillip L. Pearl, MD, FAAN, director of epilepsy and clinical neurophysiology and the William G. Lennox Chair at Boston Children's Hospital, and professor of neurology at Harvard Medical School, said that the study “exemplifies how machine learning and big data can lead to some very real, clinically relevant​ discoveries."

“The investigators creatively extracted an amazing amount of data based on EMR notes … encompassing 200,000-plus total patient years and extracted associations using [NLP]," he added.

In hindsight, Dr. Pearl said, some of the associations made—including between status epilepticus in children aged 9 to 12 months and the SCN1A gene, and between infantile hypotonia and the STXBP1 gene—were “at times intuitive, nonspecific, and familiar to experienced clinicians." But, he added, other associations were not as familiar or biased by the population in the study, including between myoclonias occurring at age 3 and IQSEC2, between involuntary eye movements happening at 6 to 9 months and synaptic transmission genes, and between infantile myoclonus and GABA pathway variants.

“This work highlights risk factors for genetic diagnoses and furthermore demonstrates that automated EMR analysis is a treasure trove waiting to be explored," Dr. Pearl said.

Adam P. Ostendorf, MD, a neurologist at Nationwide Children's Hospital in Columbus, OH, agreed, noting that the researchers extracted data from disorganized information.  

“On aggregate, they had 63 patients with the genetic etiology who had neurodevelopmental abnormalities in infancy," he said. “That's 63 patients out of 32,000; [it] is a relatively small number, but they did it in an automated fashion. That's pretty amazing."

He pointed out that a lot of practice-based learning health systems and the [EMRs] themselves have attempted to make data more discrete through forms and structured fields. “These researchers bypassed that problem by using [NLP]," Dr. Ostendorf said. “That has implications not just for epilepsy and genetics but [also] neurology more broadly or outside of the neurology field. It's exciting to see successful application of NLP to find clinically meaningful information from such a large data set."

Galer and Dr. Pearl had no disclosures.

Link Up for More Information

​Galer PD, Parthasarathy S, Xian J, et al. Clinical signatures of genetic epilepsy precede diagnosis in electronic medical records of 32,000 individuals: An analysis of more than 4.5 million clinical notes. Accessed 13 December 2022.