New research uses machine learning to predict epilepsy surgery outcomes based on normal scalp EEGs
Worldwide more than 50 million people are afflicted with temporal lobe epilepsy (TLE), which is typically treated with medication, diet, or electrical stimulation devices. However, when those options fail to control seizures, patients may consider a surgical procedure known as anterior temporal lobectomy (ATL) that removes part of the brain where the seizure originates.
According to bioengineering research assistant professor Yogatheesan Varatharajah, while 60-70 percent of patients who undergo ATL surgery find permanent relief from their seizures, 30-40 percent do not.
“In those cases, where there’s an abnormality visible on a patient’s MRI [scan], surgeons will remove that part of the brain and hope the seizures will stop,” said Varatharajah. “But is there a way to predict who will achieve seizure freedom after surgery and who will not?”
An expert in electrophysiology and machine learning, Varatharajah conducted a retrospective study with colleagues from the Mayo and Cleveland Clinics and University of Campinas in Brazil to find out.
The researchers discovered that a machine learning analysis of normal scalp EEG scans can help predict ATL surgery outcomes. They published their findings in the April 13, 2022, issue of Epilepsia.
In the study, Varatharajah created a machine learning-based algorithm that quantitatively assessed the normal electroencephalography (EEG) results of 41 TLE patients prior to their surgery at Mayo Clinic. EEG is an inexpensive clinical tool that doctors use to find abnormalities in the electrical activity of the brain. But when there are no visible abnormalities in an EEG (i.e., a normal EEG), that EEG study currently has no clinical value.
“We found that some spectral power and coherence properties of a normal EEG were different in the TLE patients who experienced freedom one year after ATL surgery and those who continued to have post-surgical seizures,” he said. “Specifically, the spectral power and coherence in the 10-25 Hz frequency range were reduced for people who continued to have seizures after surgery. We believe that this difference is related to the networks responsible for temporal lobe seizure generation within vs outside the margins of ATL.”
With that knowledge, Varatharajah and his colleagues analyzed differences between seizure-free and non-seizure-free patients and applied a machine learning classifier to predict surgery outcomes. He validated the technique using the normal EEGs of 23 Cleveland Clinic patients prior to their surgery in an out-of-sample fashion. Results were consistent between the two institutions, demonstrating a prediction accuracy of approximately 75%.
In the future, Varatharajah and his colleagues will apply his predictive software to the analysis of frontal lobe epilepsy, a seizure disorder that originates in the frontal lobe of the brain and can also be treated with surgery.
Varatharajah’s co-authors include Dr. Gregory Worrell, Boney Joseph, and Benjamin Brinkman from the Mayo Clinic Departments of Neurology, Physiology, and Biomedical Engineering; Dr. Lara Jehi, Marcia Morita-Sherman, Zachary Fitzgerald, Deborah Vegh, Dileep Nair, and Richard Burgess from the Cleveland Clinic Epilepsy Center; and Fernando Cendez from the University of Campinas.
This work was funded through the Mayo Clinic and Illinois Alliance, Mayo Clinic Neurology Artificial Intelligence Program, National Science Foundation (grant IIS-2105233), and National Institutes of Health (grants R01-NS097719 and R01-NS92882).
The full title of the paper is “Quantitative analysis of visually reviewed normal scalp EEG predicts seizure freedom following anterior temporal lobectomy.”