ICU-EEG Pattern Detection by a Convolutional Neural Network.

Area of Science:

• Medical Technology
• Artificial Intelligence in Medicine
• Neuroscience

Background:

• Continuous electroencephalogram (cEEG) monitoring is crucial for intensive care unit (ICU) patients at risk of seizures and rhythmic and periodic patterns (RPPs).
• Real-time cEEG interpretation demands specialized expertise and significant resources, often unavailable in clinical settings.
• Automated detection of critical EEG patterns is needed to improve patient care.

Purpose of the Study:

• To develop and evaluate a lightweight convolutional neural network (CNN) for automatic detection of seizures and RPPs in cEEG data.
• To assess the model's performance on both open-source and independent clinical datasets.
• To provide a tool that enhances timely EEG interpretation in resource-limited environments.

Main Methods:

• Classified time-frequency spectrograms of EEG data using a lightweight CNN.
• Trained and tested the model on the Harmful Brain Activity Classification challenge dataset (1950 patients).
• Validated the model on an independent cohort of ICU patients with epileptic seizures from Geneva University Hospital.

Main Results:

• Achieved high performance on open-source data with Area Under the Receiver Operating Characteristic (AUROC) scores up to 94% for various EEG patterns.
• Demonstrated strong temporal detection capabilities on the independent cohort, with a false positive rate (FPR) of 20-22% and true positive rate (TPR) of 76-89% around seizure onset.
• The model effectively detected epileptic patterns shortly after their occurrence.

Conclusions:

• A lightweight CNN model can reliably detect critical EEG patterns in ICU patients with minimal preprocessing.
• The model shows potential for improving timely EEG monitoring in both resource-limited and advanced clinical settings.
• This AI-driven approach can aid clinicians in managing critically ill patients requiring cEEG.