Researchers at Duke University have developed an assistive machine learning model that could significantly improve medical professionals' ability to read electroencephalogram (EEG) charts of intensive care patients.
Because measuring EEG is the only way to know if an unconscious patient is at risk of having a seizure or is experiencing seizure-like symptoms, the calculator could help save thousands of lives each year. The results are published online May 23. New England Journal of Medicine AI.
An EEG uses tiny sensors attached to the scalp to measure the brain's electrical signals, producing long lines that undulate up and down. When a patient is having a seizure, these lines jump up and down dramatically, like a seismograph during an earthquake. It's an easily recognizable signal. But other medically significant abnormalities, called seizure-like events, are much harder to spot.
“The brain activity we see exists on a continuum, and seizures are at one end, but there are many events in between that can be harmful and require medication,” said Dr. Brandon Westover, associate professor of neurology at Massachusetts General Hospital and Harvard Medical School. “The brainwave patterns caused by these events are more difficult to confidently recognize and classify, even for highly trained neurologists, who are not available at every medical facility. But doing so is crucial to the well-being of these patients.”
To build tools to help make those decisions, doctors turned to the lab of Cynthia Rudin, the Earl D. McLean, Jr. Professor of Computer Science and Electrical and Computer Engineering at Duke University. Rudin and her colleagues specialize in developing “interpretable” machine learning algorithms. While most machine learning models are “black boxes”—meaning humans can't understand how they arrived at their conclusions—an interpretable machine learning model essentially needs to show how it works.
The researchers first collected EEG samples from more than 2,700 patients and had more than 120 experts extract relevant features in the graphs to classify them as a seizure, one of four types of seizure-like events, or “other.” Each event type appears as a specific shape or repeating wavy line on the EEG chart. But these charts are often inconsistent in appearance, and telltale signals can be interrupted by bad data or blended together into a confusing chart.
“The ground truth is there, but it's hard to read,” says Stark Guo, a doctoral student in Rudin's lab. “There's ambiguity in a lot of these charts, so we had to train models to place decisions on a continuum, rather than in discrete, clearly defined bins.”
Visually, the continuum looks like a colorful starfish fleeing a predator. Each different-colored arm represents one type of seizure-like event that the EEG could indicate. The closer the algorithm moves a particular chart to the tip of the arm, the more certain it will be in its decision; the closer it moves it to the center, the less certain it will be in its decision.
Alongside this visual categorisation, the algorithm also shows the EEG pattern it used to make its decision, and provides three examples of diagnostic charts by experts that appear to be similar. “This allows medical professionals to quickly review the key sections and either agree that a pattern is present or decide that the algorithm is off the mark,” says Alina Barnett, a postdoctoral researcher in Rudin's lab. “It allows them to make a more educated decision, even if they're not highly trained in reading EEGs.”
To test their algorithm, the joint team had eight medical professionals with relevant experience classify 100 EEG samples into six categories, once with the aid of the AI and once without it. All participants performed significantly better, with overall accuracy increasing from 47% to 71%, and outperforming previous studies using similar “black box” algorithms.
“Typically, people assume that black box machine learning models are more accurate, but in many critical applications like this, that's not true,” Rudin said. “If the model is interpretable, it's much easier to troubleshoot, and in this case, the interpretable model was indeed more accurate. And it gives us a bird's-eye view of the types of abnormal electrical signals occurring in the brain, which is incredibly helpful in caring for critically ill patients.”
This research was supported by the National Science Foundation (IIS-2147061, HRD-2222336, IIS-2130250, 2014431), the National Institutes of Health (R01NS102190, R01NS102574, R01NS107291, RF1AG064312, RF1NS120947, R01AG073410, R01HL161253, K23NS124656, P20GM130447) and DHHS Nebraska Stem Cell Grant LB6066.
-Note: This news release was originally published on the Duke University Pratt School of Engineering website. Because it has been republished, it may not follow our style guide.
