Objective: Patients in the intensive care unit (ICU) often require continuous EEG (cEEG) monitoring due to the high risk of seizures and rhythmic and periodic patterns (RPPs). However, interpreting cEEG in real time is resource-intensive and heavily relies on specialized expertise, which is not always available. This study introduces a lightweight convolutional neural network (CNN) to automatically detect key EEG patterns, including seizures and RPPs.

Methods: We classified time-frequency spectrograms of EEG data from the Harmful Brain Activity Classification challenge, including 1950 patients. We tested our model on a subset of this dataset and a small independent cohort of ICU patients with epileptic seizures from the Geneva University Hospital.

Results: Our model showed good performance metrics on the open-source data with an AUROC score of 93% for SZ, 91% for lateralized PD, 94% for generalized PD, 87% for lateralized RDA, 89% for generalized RDA, and 88% for others. The evaluation with the Geneva University Hospital dataset also demonstrated strong temporal detection capabilities, showing an false positive rate (FPR) of 22%, 20%, and 21% at 50 s, 30 s, and 20 s before seizure onset, and a true positive rate (TPR) of 76%, 84%, and 89% at 20 s, 30 s, and 50 s after seizure onset.

Interpretation: This study presents a lightweight CNN model capable of detecting critical EEG patterns in ICU patients with minimal preprocessing. Moreover, the model's design provides reliable detection of ICU-EEG epileptic patterns shortly after their onset. These features underscore the model's potential to enhance timely EEG monitoring in resource-limited and advanced clinical contexts.