TL;DR
This paper presents the first evaluation of EEG Foundation Models (FMs) for detecting burst-suppression patterns in ICU patients without requiring patient-specific calibration. By introducing a clinically relevant, event-based evaluation framework, the authors demonstrate that fine-tuned foundation models—specifically REVE-base—significantly outperform traditional thresholding methods and task-specific neural networks, particularly when labeled data is scarce.
What Problem Does This Paper Address?
In Intensive Care Units (ICUs), patients in medically induced comas (e.g., for status epilepticus or severe brain injury) are monitored using electroencephalography (EEG). A key pattern of interest is burst suppression (BS), where periods of high-voltage brain activity (bursts) alternate with periods of flatline inactivity (suppressions). Tracking the frequency of these bursts helps clinicians titrate sedatives accurately.
Currently, monitoring BS relies heavily on manual counting by medical staff or commercial indices that don't directly track individual events. Automating this is notoriously difficult because:
- BS patterns vary drastically from patient to patient.
- ICU environments introduce significant signal noise and artifacts.
- High-quality, expert-annotated EEG datasets are exceptionally rare.
While EEG Foundation Models (FMs) have shown success in general EEG tasks, their ability to generalize to this critical, data-scarce ICU application without patient-specific tuning remained unexplored.
Methods
The authors evaluated three distinct, montage-agnostic EEG Foundation Models:
- REVE-base (69.2M parameters)
- LUNA-large (40.5M parameters)
- LuMamba-Tiny (4.1M parameters)
They compared these FMs against two baselines: Recursive Variance Thresholding (RVT) (the traditional clinical standard) and EEGNet (a compact, task-specific convolutional network).
Crucially, the authors evaluated four different adaptation strategies to transfer these pre-trained models to the burst-detection task:
- Full fine-tuning: Updating all encoder and classification head parameters.
- Frozen-backbone training: Keeping the pre-trained encoder static and only training the classifier.
- LoRA-based fine-tuning: Using low-rank adapters for parameter-efficient tuning.
- Two-step fine-tuning: Training the classifier first with a frozen backbone before fully fine-tuning the entire network.
A Shift to Event-Based Evaluation
Instead of relying solely on standard window-based metrics (which classify fixed 2-second chunks of EEG), the authors introduced an event-based evaluation framework. This approach groups overlapping windows into distinct "burst events" using a 0.5-second tolerance window, mimicking how a human clinician actually counts bursts per minute (BPM).
Datasets
The study utilized long-term ICU EEG recordings from 25 patients under deep sedation at the University Hospital Zurich (USZ).
- Montage: A reduced 6-channel frontal-temporal bipolar montage was extracted to simulate practical, rapid-deployment ICU setups.
- Annotations: Two experts manually annotated burst centers, which were converted into fixed 3.5-second burst intervals for ground-truth comparison.
Key Insights and Ideas
- Window-based metrics hide true clinical utility: Standard window-level F1-scores were consistently lower across all models because they heavily penalize minor timing mismatches at the boundaries of a burst. Event-based metrics provide a much fairer and clinically meaningful picture of model performance.
- Full fine-tuning is essential for EEG FMs: Unlike in natural language processing, where parameter-efficient methods like LoRA or frozen backbones often suffice, EEG FMs achieved their best performance when the entire encoder was allowed to adapt to the downstream task.
- Pre-training unlocks massive data efficiency: When training data was artificially restricted to simulate extreme data scarcity, the pre-trained models held a massive advantage. As shown in Figure 1, when using only 25% of the patient cohort data, the pre-trained REVE-base model maintained high accuracy, whereas the same architecture trained from scratch (random initialization) completely failed.
Key Results
- Top Performer: REVE-base achieved the highest event-based F1-score ($0.868 \pm 0.167$) and the lowest Bursts-Per-Minute (BPM) Mean Absolute Error ($0.448 \pm 0.284$).
- Error Reduction: REVE-base reduced the BPM estimation error by 52.1% compared to the task-specific EEGNet baseline and by 36.2% compared to the clinical standard RVT baseline.
- Data Scarcity Advantage: At 25% of the available training data (~5 hours), pre-trained REVE-base outperformed its randomly initialized counterpart by an absolute +0.723 event-based F1 points.
The comprehensive breakdown of model performance and adaptation strategies can be seen in the tables below:
Limitations and Future Work
The primary limitation of this study is its cohort size, being restricted to 25 patients from a single medical center.
To build on these promising results, future work will focus on validating these foundation models on larger, multi-center external cohorts. Additionally, the authors suggest exploring whether continued domain-specific pre-training on unlabelled ICU EEG data (rather than general clinical EEG) can further boost the models' zero-shot or few-shot generalization capabilities in intensive care environments.
Summary updated