Article summary
In the research setting, the use of artificial intelligence (AI) in epilepsy detection and treatment looks promising. But clinicians and researchers say the field needs to address scientific and ethical concerns before deploying new AI techniques related to epilepsy diagnosis and treatment.
Epilepsy diagnosis, treatment and prognosis are making remarkable progress from the lab to the bedside with the help of artificial intelligence (AI), according to neurologists and computer scientists speaking at the first international conference dedicated to the issue. is being achieved.
Among the findings presented at a conference in Breckenridge, Colorado, the researchers found that an AI program could diagnose the genetic form of epilepsy in children 3.6 years earlier than clinicians, and a machine learning program could diagnose epilepsy 85 years earlier. reported that it can predict childhood epilepsy with an accuracy of ~90%. Seizures disappear after surgical excision.
But he also stressed that clinicians and researchers need to address scientific and ethical concerns before deploying new technologies, as few AI programs are available for clinical use.
“We need to come together as a community, discuss what’s out there, set standards, set some guidelines,” said FRCP, John P. McGovern and Catherine G. McGovern University of Neurology. Distinguished Professor Samden Rato, M.D., said: UTHealthDirector of the Texas Comprehensive Epilepsy Program at McGovern School of Medicine in Houston.
The learning curve for neurologists trying to understand the advanced mathematical techniques used to build AI programs for epilepsy can be steep. For example, one of the award-winning papers presented at the conference described what was called a dynamic brain network model for predicting seizures.
“Specifically, we use source-sink (SS) metrics to quantify each node by its connectivity characteristics to other nodes in the network,” the paper states. The lead author is Amir Hossein Dalaye, a doctoral student at the Faculty of Biomedical Engineering, a national university. Johns Hopkins University School of Medicine.
Few neurologists are prepared to evaluate such esoteric techniques, but it is also true that few AI researchers understand the nuances of epilepsy, according to the conference’s organizing committee and Dr. Latu, who headed the scientific committee, said:
“We need to address the shortcomings we are currently facing when judging the veracity and usefulness of these techniques and algorithms,” he said. “If we don’t discuss these things now, it will be too late.”
Sandy Pampati, M.D., associate professor of neurology at UTHealth’s McGovern School of Medicine and director of the Epilepsy Fellowship Program, said the challenge in applying AI to epilepsy lies in the transition from the lab to the patient. .
“These programs require a lot of computing power. So how do patients use their computing power when they are at home?” Dr. Patty said. “It’s another thing to be able to put a patient in the hospital, have them evaluated, and track them for a few days. The problem is when the patient gets home. Devices may need to be smaller and more powerful.”
Despite concerns and shortcomings, experts said: neurology today There is palpable excitement about the possibilities of AI in this area.
“It was the first meeting of its kind,” Dr. Patty said. “Although it was small compared to the larger professional association conferences, it gave us a great perspective on where the field is headed.”
promising research
Vikram Rao, M.D., Associate Professor of Clinical Neurology and Head of the Epilepsy Division at the University of California, San Francisco, has used AI to identify hidden multiday periodicity of up to 30 days in interictal epileptiform activity. (IEA), seizures occur preferentially during the ascending phase of the IEA rhythm.
“If we consider the duration of a day, we can predict seizures in about two-thirds of people better than by chance,” he said. “It’s not perfect, but it’s a lot better than existing methods. This isn’t something we can offer patients yet, but it’s much more than just ideas and theories. We want these programs to be as accurate as possible so that we can set up the Holy Grail.It’s not going to happen this year or next year.We’re looking out five to ten years. And I think you are.”
“We need to come together as a community, discuss what’s out there, set standards, set some guidelines.” —DR. Samden Rato
While Dr. Rao’s research used data generated from implantable devices, other scientists have used data obtained from non-invasive wearable devices. Benjamin H. Brinkman, Ph.D., an associate professor of neurology and clinical support scientist in the Epilepsy Division at the Mayo Clinic in Rochester, Minnesota, found that by measuring heart rate and accelerometers, it was possible to detect the symptoms minutes to hours earlier. Published a paper describing an algorithm to predict seizures. , electrodermal activity, and temperature. Similar to Dr. Rao’s study, Dr. Brinkman also found that seizure risk varied between daily and multi-day cycles.
One of the three award-winning abstracts presented at the Colorado conference described clinical characteristics from open-ended electronic medical records of all 32,112 patients diagnosed with childhood epilepsy at the Children’s Hospital of Philadelphia (CHOP). A natural language processing tool to extract was explained. After analyzing 4.5 million clinical notes, the AI was able to distinguish between children with genetic causes of epilepsy, taking clinicians a median of 3.6 years to make the diagnosis.
“Our machine learning model was able to predict common genetic diagnoses with 80 percent accuracy at age 1.5 years,” said lead author of the paper, Ph.D. student in the CHOP Epilepsy Neurogenetics Initiative. Master D. Gaylor said: Neuroengineering and Therapeutic Center at the University of Pennsylvania. “All our work uses only existing free-text clinical notes prior to clinical diagnosis.”
“AI is a small but rapidly growing area of epilepsy research,” says Gaylor. “I don’t think it will be long before it starts to become an important part of clinical care. The conference was attended by many excellent presenters and researchers: electrophysiology, imaging, genetics, informatics. There was a strong push for representation by many subdisciplines of epilepsy, such as
Another presentation at the conference discussed the use of machine learning to predict who will be seizure-free after ablation surgery, rather than predicting seizures.
“Using classical computational tools commonly used in statistics, we were able to devise a model that was approximately 60% accurate,” said Chief Research and Information Officer, Director of the Epilepsy Outcomes Research Program. says Lara Jehi, M.D.for Cleveland Clinic medical system. “With machine learning, our algorithm looked at his MRI itself and improved accuracy to 85-90%. Much better.”
Dr. Jehi said the next step would be combining MRI data with EEG, medical history and other tests. AI can not only predict the likelihood that a patient will not have a seizure after surgery, but it can also predict exactly where in the brain surgery should be performed. “We need to know exactly where the best places are to help patients,” she said.
potential pitfalls
Some of the most promising research may actually have little practical value, says co-director of the Texas Restorative Neurotechnology Institute, professor of digital innovation and distinguished professor at UT. said Dr. GQ Zhang, principal data scientist at Health Houston.
“Many of these studies have given very good results, and there seems to be no room for improvement,” Dr. Zhang said. “However, the metrics they use may just show how the data fits the AI model using their own defined metrics. There may be very little potential.”
Hundreds of studies have been published on the use of AI in epilepsy, but “the way physicians monitor and care for their patients has not changed fundamentally. It’s a valuable raw material, but there’s still a long way to go in terms of the translatability of AI tools that leverage such data to create life-changing impacts in the real world.”
Dr. Chan likened the task of predicting seizures to predicting earthquakes and tornadoes. “These things are inherently unpredictable,” he says. “I can’t say for sure when that will happen.”
Dr. Rao agreed that, like weather forecasting, the accuracy of seizure prediction is never perfect.
“If you say there’s a 95 percent chance of rain, there’s a 5 percent chance that it won’t rain,” he says. “There can be ethical or medical legal issues around making predictions. Or if they say they are at high risk and stay home and don’t have it Will these predictions actually help people or will they further increase people’s anxiety? The way we use information needs to evolve, and it’s not as simple as saying, ‘Okay, you’re safe, don’t worry.
As a matter of fact, Dr. Rao said: “The current state of our knowledge is very limited. ”
Kathryn Davis, M.D., M.D., FAES, associate professor of neurology and director of the Penn Epilepsy Center at the University of Pennsylvania, said she sees both promise and danger in the use of AI in epilepsy treatment.
“I think we will be able to use AI in many aspects of epilepsy, such as interpreting images and predicting seizures, drug resistance, and response to device therapy,” said Dr. Davis. neurology today. “But there are also pitfalls. One is that researchers need to make sure they are using the right datasets when developing algorithms. It’s not very useful if it’s not well-represented in the dataset used to train the .”
Dr. Davis also pointed out the danger of over-reliance on AI tools. “If you don’t use clinical intuition and observation, you may end up making the wrong recommendations to your patients,” she says. “We need people who can stay informed.”
Disclosure
Dr. Rao is a consultant (stock option holder) for Novela Neurotechnologies Inc. and EnlitenAI, Inc. He also receives consulting fees from NeuroPace Inc. and serves on the Advisory Board of LivaNova PLC..
