Article In Brief
In research settings, the use of artificial intelligence (AI) in epilepsy detection and care looks promising. But clinicians and researchers say the field must address scientific and ethical concerns before they roll out new AI technologies related to the diagnosis and treatment of epilepsy.
Remarkable gains in the diagnosis, treatment, and prognosis of epilepsy are heading from the lab to the bedside with the aid of artificial intelligence (AI), according to neurologists and computer scientists who spoke at the first international devoted to the subject.
Among the findings from studies presented at the conference in Breckenridge, CO, researchers reported that an AI program can diagnose genetic forms of epilepsy in children 3.6 years sooner than clinicians can, and that a machine learning program can predict with 85 to 90 percent accuracy who will be seizure-free following ablative surgery.
But clinicians and researchers also emphasized that scientific and ethical concerns must be addressed before the new technologies are rolled out, since few AI programs are ready for the clinic.
“We need to come together as a community and begin to discuss what's out there, set standards, and lay down some guidelines,” said Samden Lhatoo, MD, FRCP, the John P. and Kathrine G. McGovern Distinguished University Professor of Neurology and director of the Texas Comprehensive Epilepsy Program at UTHealth Houston's McGovern Medical School.
The learning curve for neurologists seeking to understand the advanced mathematical techniques used to construct AI programs for epilepsy can be steep. One of the prize-winning papers presented at the conference, for instance, described what it called a dynamic brain network model to predict seizures.
“Specifically, we use the source-sink (SS) metric to quantify each node by its connectivity properties to other nodes in the network,” stated the paper, whose first author was Amir Hossein Daraie, a PhD student in the biomedical engineering department at Johns Hopkins University School of Medicine.
While few neurologists are prepared to evaluate such abstruse methods, it's also true that few AI researchers understand the nuances of epilepsy, said Dr. Lhatoo, who led the conference's organizing and scientific committees.
“We need to address the shortcomings we currently face in judging the veracity and utility of these technologies and algorithms,” he said. “If we don't discuss these things now, it will be too late.”
Sandipan Pati, MD, associate professor of neurology and director of the epilepsy fellowship program at McGovern Medical School at UTHealth, said the challenge of applying AI to epilepsy is in making the transition from the laboratory to the patient.
“These programs require a lot of computational power, so how does a patient make use of it when they're home?” Dr. Pati said. “It's one thing to have the patient in the hospital to undergo evaluation, where we can track them for days. The problem comes when the patient goes home. Perhaps the device will link to the cloud, where the computations happen. Or perhaps the devices will have to get smaller and more powerful.”
Despite the misgivings and shortcomings, experts told Neurology Today that the field is experiencing a palpable feeling of excitement about the potential of AI.
“The conference was the first of its kind,” Dr. Pati said. “It was small compared to the large professional society meetings, but it gave us a great view of where the field is going.”
Promising Studies
Vikram Rao, MD, PhD, associate professor of clinical neurology and chief of the epilepsy division at the University of California, San Francisco, is researching the use of AI to discern hidden multi-day periodicities of up to 30 days in interictal epileptiform activity (IEA), with seizures occurring preferentially during the rising phase of the IEA rhythm.
“At a one-day horizon, we can forecast seizures better than chance in about two-thirds of people,” he said. “It's not perfect, but it's much better than existing methods. It's still not something we can offer to patients yet, but it's well beyond being just an idea or theory. We want these programs to be as accurate as possible so patients can plan based on it. That's the holy grail. It won't happen this year or next year. I think we're looking at a five- to 10-year horizon.”
“We need to come together as a community and begin to discuss what's out there, set standards, and lay down some guidelines.”—DR. SAMDEN LHATOO
While Dr. Rao's study used data generated from implanted devices, other scientists use data obtained from noninvasive wearable devices. Benjamin H. Brinkmann, PhD, associate professor of neurology and a clinical support scientist for the division of epilepsy at the Mayo Clinic in Rochester, MN, has published papers describing an algorithm that forecasts seizures minutes to hours in advance by measuring heart rate, accelerometry, electrodermal activity, and temperature. Like Dr. Rao's research, Dr. Brinkmann also found that seizure risk varies in daylong and multi-day cycles.
One of three prize-winning abstracts presented at the meeting in Colorado described a natural-language processing tool, which extracted clinical features from the free-text electronic medical records of all 32,112 patients at Children's Hospital of Philadelphia (CHOP) diagnosed with childhood epilepsy. After analyzing 4.5 million clinical notes, the AI was able to distinguish those children whose epilepsy has a genetic cause—and did so a median of 3.6 years before clinicians made the diagnosis.
“Our machine learning models could predict a genetic diagnosis in general at 1.5 years of age with an accuracy of 80 percent,” said the paper's first author, Peter D. Galer, MSc, a PhD student with the Epilepsy Neurogenetics Initiative at CHOP and the Center for Neuroengineering and Therapeutics at the University of Pennsylvania. “All of our work uses only preexisting free-text clinical notes prior to any clinical diagnosis.”
“AI is a small but rapidly growing area in epilepsy research,” Galer said. “I do not think it will be long before it begins to be an essential part of clinical care. This meeting had a lot of brilliant presenters and researchers. There was a strong push for representation by the many subfields of epilepsy, including electrophysiology, imaging, genetic, and informatics.”
Rather than predicting seizures, another presentation at the meeting described the use of machine learning to predict who will be seizure-free following ablative surgery.
“Using classical computational tools, the kind commonly used in statistics, we've been able to come up with models that are about 60 percent accurate,” said Lara Jehi, MD, director of the outcomes research program for epilepsy and chief research information officer for Cleveland Clinic Health System. “With machine learning, our algorithm looks at the MRI itself, and that has bumped up the accuracy to 85 to 90 percent. It's a lot better.”
Dr. Jehi said her next step will be to combine MRI data with EEGs, clinical history, and other tests. The AI could predict not only patients' likelihood of being seizure-free after surgery but also where precisely in the brain that surgery should occur. “We need to know exactly where the best spot is to help the patient,” she said.
Potential Pitfalls
Some of the studies that look most promising may actually be of little practical value, said GQ Zhang, PhD, professor and distinguished chair in digital innovation, co-director of the Texas Institute for Restorative Neurotechnologies, and chief data scientist at UTHealth Houston.
“Many of these studies have such excellent results, they seem to leave no room for improvement,” Dr. Zhang said. “But the metrics they use might only show how their data fits into their AI model using self-defined metrics. As a result, they may have little real-world relevance.”
Although hundreds of studies have been published on the use of AI in epilepsy, he said, “it hasn't yet fundamentally changed the way doctors monitor or care for patients. The data from EEGs, MRI, and other modalities are invaluable raw material. However, we still have a long way to go in the translatability of AI tools that harness such data to generate real-world, life-changing impact.”
Dr. Zhang compared the challenge of predicting seizures to the forecasting of earthquakes or tornados. “There's unpredictability inherent in these things,” he said. “We cannot say for certain when they will happen.”
Dr. Rao agreed that the accuracy of predicting a seizure, as in weather forecasting, will never be perfect.
“Even if you say there's a 95 percent chance of rain, 5 percent of the time it won't rain,” he said. “There may be ethical or medical-legal issues about making forecasts. What if you tell the person they're at a very low risk of seizure and they drive to work and have one? Or if you say they're at high risk and they stay home and don't have one? Will these forecasts actually help people, or will it make them more anxious? The way we use this information will have to evolve. It won't be as simple as saying, ‘OK, you're in the clear, so no need to worry.’”
In truth, Dr. Rao said, “the current state of our knowledge is very limited. I can't tell a person much about when a seizure will or won't happen. Relative to that, these approaches can represent a huge step forward.”
Kathryn Davis, MD, MS, FAES, associate professor of neurology at the University of Pennsylvania and director of the Penn Epilepsy Center, said she sees both promise and peril in the use of AI for treating epilepsy.
“I do think we will be able to use AI across many aspects of epilepsy, including imaging interpretation and the prediction of seizures, drug resistance, and responsiveness to device therapy,” Dr. Davis told Neurology Today. “But there are pitfalls. One is that researchers need to be sure they're using the proper data set when they develop their algorithms. If the patients you're treating in the clinic aren't well represented in the dataset used to train the algorithm, it won't be very useful.”
Dr. Davis also pointed out the danger of an over-reliance on an AI tool. “If you aren't using clinical intuition and observation, you could make the wrong recommendations for patients,” she said. “We will always need a human in the loop.”
Disclosures
Dr. Rao is a consultant for Novela Neurotechnologies Inc. and EnlitenAI, Inc. (for which he receives stock option). He has also received consulting fees from NeuroPace Inc. and served on the advisory board for LivaNova PLC.
