More than 80 people are diagnosed with epilepsy every day. About 60 million people worldwide live with this neurological disorder, experiencing varying degrees of seizures caused by unusual nerve cell activity in the brain.
Epilepsy can be genetically acquired and appear in childhood, or may be brought on by a car crash as an adult. Often the cause is unknown.
Some epilepsy patients experience several seizures a day while others have only a few in their lifetime. Fortunately, most patients can keep episodes at bay with medication, but some have drug-resistant forms that require more invasive treatment.
“There’s a ton that we don’t know about epilepsy, that’s why it’s such an active area of study,” says Alex Gardner, a recent graduate of Rice University’s George R. Brown School of Engineering
Gardner and two other computational and applied mathematics (CAAM) students, Wendy Knight and Evan Toler, teamed up with Dr. Nitin Tandon, a professor of neurosurgery at The University of Texas Health Science Center at Houston, and Beatrice Riviere
to develop BrainGuide, a software-based automation tool that doctors can use to more efficiently treat patients with drug-resistant forms of epilepsy.
Some epilepsy patients who find that medication doesn’t work for them undergo a procedure that involves drilling long metallic probes into the brain at pre-determined trajectories.
Once the patient has anywhere between eight and fourteen of these probes planted in their skull, electroencephalography (EEG) data is recorded over several hours. This information tells doctors what areas of the brain need to be surgically removed.
“Mathematically speaking, it’s an inversion problem in trying to determine which region is being activated when they have a seizure,” says Riviere, Rice’s Noah Harding Chair and a professor of CAAM.
According to Gardner, only a few thousand people undergo the procedure a year, making the pool of study participants fairly small.
“This is basically a last resort. If there is no drug that treats your epilepsy properly, they want to remove part of your brain,” Gardner says. “Most people don’t want you going in and touching their brain, but their quality of life is so poor that they’re willing to just let people get in there and prod and poke in their brain and get data. They really have no other option.”
Currently, optimal trajectories for the probes are calculated manually. But BrainGuide is changing things by computing the trajectories faster and with greater accuracy.
“We don’t want to hit any major blood vessels or any really big vascular regions because that could cause hemorrhage,” Gardener says. “We also wanted to make sure that the probes enter the brain as perpendicular to the surface of the skull as possible. That makes it more likely that the probes are going to follow their prescribed trajectory.”
With the help of Dr. Tandon, Riviere, and Kiefer Foreseth, a researcher in Tandon’s lab, the students used MRI and CT data from forty of Dr. Tandon’s patients to build models of each of their brains.
According to Riviere, after the data was put into BrainGuide, it predicted optimal trajectories in a mere thirty seconds. This is a significant drop from the four to eight hours it usually takes doctors to do these calculations manually.
Although BrainGuide has great potential, both Gardner and Riviere say more work must be done before it can be implemented by other doctors. In order for the machine learning to improve, data needs to be collected from a much larger pool of patients.
“The problem is that some doctors do things differently,” Gardner says. “Is it technically universal? Yes. Is it practically universal? Maybe not. The big limiting factor there was just our lack of data.”
Until then, Gardner says BrainGuide gives neurosurgeons something to aim for, a potential solution to get excited about. The key to success will be collecting the extra data. Hopefully this will happen now that the project is moving back to UTHealth for further testing.
“The next step is to get more data and test it on Dr. Tandon’s patients, but also on patients that aren’t his,” Gardner says. “That can be a really big thing in its design, we just don’t have the proper data to actually do it, yet. The next step for that is to add more data and really see what we can do.”