AI model to improve EEG reading for ICU patients

DURHAM, N.C: Duke University researchers have unveiled an assistive machine-learning model to significantly enhance the interpretation of electroencephalography (EEG) charts for patients in intensive care with diagnostic precision and to study unusual brain activity for timely intervention aimed at saving lives.

Published online in the New England Journal of Medicine AI recently, this innovation holds the potential to save countless lives annually by helping medical professionals in identifying critical neurological events.

EEG readings, derived from sensors affixed to the scalp, capture the brain's electrical activity in the form of intricate patterns resembling undulating squiggles. While seizures manifest as unmistakable spikes in this data, other clinically significant occurrences, termed seizure-like events, pose a greater challenge for detection.

Dr. Brandon Westover, Associate Professor of neurology at Massachusetts General Hospital and Harvard Medical School, highlights the complexity of interpreting EEG data: "The spectrum of brain activity we're examining spans from seizures to subtler events that still necessitate intervention. Identifying these nuanced patterns is crucial for patient outcomes, yet it poses a formidable task, even for seasoned neurologists," he says.

To address this challenge, the research team used the expertise of Cynthia Rudin's laboratory at Duke University, renowned for developing interpretable machine learning algorithms. Unlike conventional "black box" models, which obscure their decision-making process, interpretable models offer transparent insights into their rationale.

The study commenced with the collection of EEG samples from over 2,700 patients, analyzed by 120 experts to explain distinct features corresponding to seizures, seizure-like events, or other anomalies. However, due to the inherent variability and ambiguity in EEG charts, the model was trained to navigate a continuum of interpretations rather than rigid categories.

Stark Guo, a Ph.D. student in Rudin's lab, says: "Deciphering EEGs entails a degree of uncertainty. Our model embraces this complexity, akin to observing a multifaceted spectrum of signals rather than discrete classifications."

Guo says visualizing this continuum resembles a vibrant starfish, with each arm representing a distinct type of neurological event. The algorithm positions each EEG chart along these arms, indicating its confidence level in the diagnosis. It also elucidates the specific brainwave patterns informing its decision and offers comparative examples for validation.

Alina Barnett, a postdoctoral research associate in Rudin's lab, says: "By highlighting salient features and providing reference points, our model empowers medical professionals to make informed judgments, irrespective of their expertise in EEG interpretation."

In a comparative evaluation involving 100 EEG samples, eight medical professionals demonstrated a marked improvement in accuracy when aided by the AI model, improving their overall performance from 47pc to 71pc. This interpretable approach surpassed the efficacy of traditional "black box" algorithms.

Professor Rudin emphasizes the superiority of interpretable models saying: "Contrary to common belief, transparency enhances accuracy, particularly in critical domains like healthcare. Our model not only enhances diagnostic precision but also furnishes a comprehensive overview of abnormal brain activity, thereby enhancing patient care."

In essence, Duke University's pioneering machine learning model represents a paradigm shift in EEG analysis, heralding a new era of precision medicine and improved outcomes for intensive care patients.