Related Experiment Video
Updated: Jun 24, 2025
07:37
Assessment and Communication for People with Disorders of Consciousness
Published on: August 1, 2017
9.0K
Improving Clinician Performance in Classifying EEG Patterns on the Ictal-Interictal Injury Continuum Using
Alina Jade Barnett1, Zhicheng Guo2, Jin Jing3
1Computer Science, Duke University, Durham, NC.
NEJM AI
|June 14, 2024
Summary
An interpretable deep-learning system significantly improved electroencephalography (EEG) interpretation accuracy for critical care clinicians. This AI tool enhances diagnostic capabilities and understanding of brain activity patterns in intensive care units (ICUs).
Area of Science:
- Artificial Intelligence in Medicine
- Computational Neuroscience
- Clinical Neurophysiology
Background:
- Intensive care units (ICUs) use electroencephalography (EEG) to monitor critically ill patients for brain injury.
- Current EEG interpretation faces challenges due to limited clinician availability and subjective analysis, leading to variability.
- Black-box deep-learning models lack transparency, hindering trust and clinical adoption despite potential benefits.
Purpose of the Study:
- To develop an interpretable deep-learning system for classifying six key EEG patterns.
- To provide case-based explanations for AI predictions to enhance clinician trust and understanding.
- To evaluate the system's impact on diagnostic accuracy and its support for the ictal-interictal injury continuum hypothesis.
Main Methods:
- Developed an interpretable deep-learning model trained on 50,697 EEG samples from 2711 ICU patients.
- Classified six EEG patterns: seizure, LPDs, GPDs, LRDA, GRDA, and other, with expert-annotated data.
- Evaluated AI assistance by comparing medical professionals' diagnostic accuracy with and without the system; assessed interpretability using neighborhood agreement statistics and latent space visualization.
Main Results:
- AI assistance significantly improved mean user diagnostic accuracy from 47% to 71% (P<0.04).
- The model achieved high Area Under the Receiver Operating Characteristic Curves (AUROCs) for all classes, outperforming a black-box model (P<0.0001).
- Latent space visualization supported the ictal-interictal injury continuum hypothesis, revealing relationships between EEG patterns.
Conclusions:
- The interpretable deep-learning model significantly enhances clinicians' EEG pattern classification accuracy.
- The system's interpretable design fosters human-AI collaboration, potentially improving ICU diagnosis and patient care.
- The model offers insights into EEG patterns and supports the ictal-interictal injury continuum hypothesis.
