Automatically detecting burst-suppression patterns in EEG recordings from ICU patients is a clinically important but challenging task due to high inter-patient variability and limited annotated data. A recent study published on arXiv presents the first evaluation of EEG foundation models for this application, comparing multiple architectures and adaptation strategies without requiring patient-specific calibration.
Clinical Context and Challenge
Burst suppression (BS) is an electroencephalographic pattern used to monitor sedation depth and brain activity in critically ill patients, particularly during induced coma in Intensive Care Units (ICUs). According to the study, automatic burst detection remains challenging because BS patterns vary substantially between patients and annotated datasets are scarce.
Models and Evaluation Methodology
The researchers compared three foundation models—REVE-base, LUNA-large, and LuMamba-Tiny—with an adaptive thresholding baseline and a task-specific EEGNet baseline. The evaluation used reduced-montage ICU EEG data without patient-specific calibration. Importantly, the study complemented conventional EEG window-based classification with an event-based burst detection evaluation, which helps assess clinically whether burst episodes are correctly detected and reduces the impact of expected annotation variability.
Performance Results
| Model | Event-based F1-score | Burst-per-minute error reduction vs. EEGNet | Burst-per-minute error reduction vs. adaptive thresholding |
|---|---|---|---|
| REVE-base | 0.868 ± 0.167 | 52.1% | 36.2% |
| EEGNet (baseline) | - | - | - |
| Adaptive thresholding | - | - | - |
REVE-base achieved the highest event-based F1-score of 0.868 ± 0.167 and reduced burst-per-minute error by 52.1% and 36.2% compared to EEGNet and adaptive thresholding respectively, supporting the use of foundation models for scalable EEG monitoring in ICU, according to the study.
Fine-Tuning Strategy Findings
Ablation experiments showed that full fine-tuning was the most effective adaptation strategy compared to frozen-backbone training, two-step fine-tuning, and LoRA-based adaptation. For LUNA-large, full fine-tuning improved event-based F1-score over frozen-backbone training by up to +0.102.
Benefit of Pretrained Representations
With reduced labeled datasets, pretrained REVE-base outperformed random initialization by +0.723 event-based F1 points when trained on 25% of the cohort, demonstrating the benefit of pretraining foundation model representations when adapted to burst detection with limited labeled data. This finding is particularly relevant for clinical settings where large annotated datasets are difficult to obtain.
Implications for Healthcare AI
The study provides evidence that EEG foundation models can be effectively adapted for burst-suppression detection, potentially enabling more reliable and scalable automated monitoring in ICU environments. The event-based evaluation approach offers a clinically meaningful metric that accounts for annotation variability, and the success of full fine-tuning suggests that even with limited data, the rich representations learned during pretraining can be leveraged to achieve strong performance.
