Interpretable machine learning helps clinicians classify EEG anomalies

A Duke University-led research team has developed a machine learning (ML) tool designed to assist clinicians in accurately reading the electroencephalography (EEG) charts of patients in intensive care.

EEGs are currently the only reliable method for determining when an unconscious patient is prone to suffering a seizure or experiencing a seizure-like event. These events can be life-threatening, but reading EEGs correctly presents a challenge.

During an EEG, sensors attached to a patient’s scalp record the brain’s electrical signal in the form of wavy lines. In the event of a seizure, these lines jump up and down in a telltale pattern that is easily recognizable to many clinicians. However, seizure-like events can present in more subtle ways, making them difficult to catch on an EEG.

“The brain activity we’re looking at exists along a continuum, where seizures are at one end, but there’s still a lot of events in the middle that can also cause harm and require medication,” explained Brandon Westover, MD, PhD, associate professor of neurology at Massachusetts General Hospital and Harvard Medical School, in the news release. “The EEG patterns caused by those events are more difficult to recognize and categorize confidently, even by highly trained neurologists, which not every medical facility has. But doing so is extremely important to the health outcomes of these patients.”

To enhance the detection of seizure-like events, the researchers turned to interpretable machine learning. Unlike black box AI tools, interpretable models must provide rationale for how they reach their conclusions, making them potentially useful for healthcare applications like EEG classification.

The research team underscored that while seizure-like events often manifest on EEGs as particular graph shapes or repetitions in the appearance of the lines, the variability in how an EEG can look, combined with the possibility of data “noise,” can make the chart confusing to read and interpret.

“There is a ground truth, but it’s difficult to read,” said Stark Guo, a PhD student at Duke who worked on the study. “The inherent ambiguity in many of these charts meant we had to train the model to place its decisions within a continuum rather than well-defined separate bins.”

To develop the model, the research team gathered EEGs from more than 2,700 patients and asked 120 experts to flag the relevant features in each graph, categorizing them as a seizure, one of four types of seizure-like events or “other.”

Using this data, the model was tasked with visualizing each EEG on a chart. The chart, which resembles a multicolored starfish, displays the continuum on which an EEG can exist. Each colored “arm” on the chart represents one type of seizure-like event, and the model places each EEG on the arm in relation to how sure it is of its classification; EEGs placed near the tip of an arm are those the model are more sure about, while those toward the center of the chart are less certain.

In addition to the visualization, the model also surfaces which patterns in the brainwaves it used to make its determination and provides three examples of professionally reviewed and annotated EEGs it thinks are similar to the one in question.

“This lets a medical professional quickly look at the important sections and either agree that the patterns are there or decide that the algorithm is off the mark,” said Alina Barnett, a postdoctoral research associate at Duke. “Even if they’re not highly trained to read EEGs, they can make a much more educated decision.”

To test the model, the research team tasked eight experts with classifying 100 EEG samples across categories, both with and without the ML’s assistance. Without using the tool, the group achieved 47% accuracy, which rose to 71% when the tool was applied. The group also outperformed participants who utilized a black box model for the same task in a previous study.

“Usually, people think that black box machine learning models are more accurate, but for many important applications, like this one, it’s just not true,” said Cynthia Rudin, the Earl D. McLean, Jr. Professor of Computer Science and Electrical and Computer Engineering at Duke. “It’s much easier to troubleshoot models when they are interpretable. And in this case, the interpretable model was actually more accurate. It also provides a bird’s eye view of the types of anomalous electrical signals that occur in the brain, which is really useful for care of critically ill patients.”