Quantitative EEG reactivity and machine learning for prognostication in hypoxic-ischemic brain injury

Area of Science:

• Neuroscience
• Medical Technology
• Artificial Intelligence

Background:

• Electroencephalogram (EEG) reactivity is crucial for predicting neurological recovery post-cardiac arrest.
• Limited interrater agreement among electroencephalographers poses a challenge in EEG interpretation.
• Hypoxic-ischemic brain injury following cardiac arrest requires accurate prognostication methods.

Purpose of the Study:

• To evaluate machine learning (ML) methods for predicting long-term outcomes using EEG reactivity data.
• To assess the performance of ML models in patients with hypoxic-ischemic brain injury.
• To compare ML-based EEG reactivity assessment with expert review.

Main Methods:

• Retrospective review of clinical and EEG data from comatose cardiac arrest survivors.
• EEG reactivity assessed using sound and pain stimuli within 72 hours post-arrest.
• Quantitative EEG (QEEG) analysis of spectral, entropy, and frequency features during stimulation.
• Comparison of a random forest classifier against a penalized general linear model (GLM) and expert assessment.

Main Results:

• Fifty subjects were analyzed; 32% achieved good long-term outcomes (Cerebral Performance Category 1-2 at six months).
• ML models using QEEG reactivity demonstrated comparable performance to expert review (Random Forest AUC 0.8 vs. Expert AUC 0.69).
• The penalized GLM also showed similar performance (AUC 0.69).

Conclusions:

• Machine learning models incorporating quantitative EEG reactivity can reliably predict long-term outcomes after cardiac arrest.
• Quantitative EEG analysis offers a potential tool to aid prognostication in post-cardiac arrest care.
• This approach may help standardize and improve the accuracy of EEG-based prognostication.