Computer Program Identified Seizures with Expert-Accuracy, Researchers Demonstrate

Automated Brain Monitoring Could Improve Care for Patients at Risk for Seizure

BOSTON – Seizures and other seizure-like patterns of brain activity can damage the brain and may signal the risk of an impending stroke. This harmful brain activity is seen not just in patients who have epilepsy; it is also common in patients who are hospitalized because of brain injuries, trauma, or critical illness such as kidney failure or heart attack. Vulnerable patients are typically monitored via electroencephalography (EEG), the data from which must be interpreted by highly trained experts in the field of neurophysiology. However, there’s a shortage of these specialists in the United States, and brain monitoring services are not available at all in much of the world.

Now, physician-scientists have trained a deep neural network – a form of artificial intelligence – to classify seizures and other seizure-like events automatically. The team used more than 50,000 EEG-generated data points to train the computer program and then assessed its performance against 20 fellowship-trained neurophysiologists. The scientists demonstrated the algorithm matched the experts’ performance. The findings, published in Neurology, suggest the computer program could help address the critical unmet need for automatic detection of these brain events without compromising accuracy.

“Fellowship-trained clinical neurophysiologists are the gold standard for identifying seizures and seizure-like types of brain activity; however, these highly trained subspecialists are scarce,” said corresponding author M. Brandon Westover, MD, PhD, a neurologist at BIDMC and the Emily Fisher Landau Professor of Neurology at Harvard Medical School. “Our results demonstrate that the computer program identified seizures and other seizure-like patterns of brain activity and distinguishes these from non-seizure like events at least as well as neurologists with subspecialty training in clinical neurophysiology.”

Westover and colleagues’ goal was to train a deep neural network, called SPaRCNet, which stands for “Seizures, Periodic and Rhythmic Continuum Neural network” to identify five specific seizure and seizure-like patterns on EEG readouts – collectively known as IIIC patterns – and to be able to distinguish them from non-IIIC patterns. The team fed SPaRCNet data from 6,095 scalp EEGs from 2,711 patients, with and without IIIC events. The dataset included a total of 50,697 EEG segments, or relevant sections of EEG readouts, which had been independently analyzed and labeled by 20 experienced neurophysiologists.

When the researchers assessed SPaRCNet’s ability to identify IIIC events, they found the computer program matched or exceeded most of the neurophysiology experts’ ability to classify seizures and other IIIC events and distinguish them from non-IIIC events. The results indicate that SPaRCNet can classify seizures and other IIIC events and distinguish them from non-IIIC events at least as well as human experts, and with calibration better than most individual experts with performance comparable to the consensus of a committee of experts.

“By training on a large set of examples annotated by multiple raters, SPaRCNet effectively learns to simulate a committee of experts,” said first author Jin Jing, PhD, an instructor in Neurology at BIDMC and Harvard Medical School. “By automating a challenging diagnostic task previously limited to specialists, SPaRCNet opens a path for expanding brain monitoring to a broader range of patients with epilepsy and critical illness.”

Westover and Jing completed this research while affiliated with Massachusetts General Hospital (MGH). Co-authors included Mouhsin M. Shafi, MD, PhD, of BIDMC; Wendong Ge, PhD, Marta Bento Fernandes, PhD, Mohammad Tabaeizadeh, MD, Kristy Nordstrom, AS, Fábio A. Nascimento, MD, Ziwei Fan, MS, Samaneh Nasiri, PhD, Sydney S. Cash, MD, PhD, Daniel B. Hoch, MD, PhD, Andrew J. Cole, MD, Eric S. Rosenthal, MD, and Sahar F. Zafar, MD, of MGH; Aaron F. Struck, MD, and Safoora Fatima, MD, of the University of Wisconsin-Madison; Shenda Hong, PhD, of Peking University; Sungtae An of Georgia Institute of Technology; Aline Herlopian, MD, Jennifer A. Kim, MD, PhD, and Emily J. Gilmore, MD, of Yale University-Yale New Haven; Ioannis Karakis, MD, PhD, MSc, and Andres A. Rodriguez Ruiz, MD, of Emory University School of Medicine; Jonathan J. Halford, MD, and Sarah Schmitt, MD, of Medical University of South Carolina; Marcus C. Ng, MD, of the University of Manitoba; Emily L. Johnson, MD, Peter W. Kaplan, MBBS, FRCP, and Mackenzie C. Cervenka, MD, of Johns Hopkins School of Medicine; Brian L. Appavu, MD, of the University of Arizona College of Medicine; Rani A. Sarkis, MD, MSc, and Jong Woo Lee, MD, PhD, of Brigham and Women's Hospital; Gamaleldin Osman, MD, MS, of Mayo Clinic-Rochester; Monica B. Dhakar, MD, MS, of Brown University; Lakshman Arcot Jayagopal, MD, and Olga Taraschenko, MD, PhD, of the University of Nebraska Medical Center; Zubeda Sheikh, MD, MS, of West Virginia University Hospitals; Hiba A. Haider, MD, of the University of Chicago; Christa B. Swisher, MD, of Atrium Health; Nicolas Gaspard, MD, PhD, of Université Libre de Bruxelles; Ji Yeoun Yoo, MD, of Icahn School of Medicine; Manisha G. Holmes, MD, of New York University; Susan T. Herman, MD, of Barrow Neurological Institute; Jennifer A. Williams, MB, BAO, Bch, FRCPI of Mater Misericordiae University Hospital; Jay Pathmanathan, MD, PhD, of the University of Pennsylvania; and Jimeng Sun, PhD, of the University of Illinois.

This work was supported by the Glenn Foundation for Medical Research and American Federation for Aging Research (Breakthroughs in Gerontology Grant); American Academy of Sleep Medicine (AASM Foundation Strategic Research Award); Football Players Health Study (FPHS) at Harvard University; Department of Defense; National Institutes of Health (R01NS102190, R01NS102574, R01NS107291, RF1AG064312, R01AG062989, R01NS111022, R01 NS062092, K24 NS088568, R25NS06574, P20GM130447); and NSF (award SCH-2014431, NS11672601); American Epilepsy Society; Marinus Pharmaceuticals and Parexel Inc.; AHA; Bee Foundation; Nutricia, Vitaflo, Glut1 Deficiency Foundation, BrightFocus Foundation. The funding sources had no role in study design, data collection, analysis, interpretation, or writing of the report. The authors report no disclosures relevant to the manuscript.