Evaluation of EEG Foundation Models for Event-Based Burst-Suppression Detection in ICU

Abstract: Burst suppression (BS) is a clinically relevant electroencephalographic (EEG) pattern used to monitor sedation depth and brain activity in critically ill patients, particularly during induced coma in Intensive Care Units (ICUs). Automatic burst detection remains challenging because BS patterns vary substantially between patients and annotated datasets are scarce. Recently, EEG Foundation Models (FMs) have shown promise across several downstream EEG applications, but their usefulness for BS detection remains unexplored. We present the first study to evaluate EEG FMs for burst detection in reduced-montage ICU EEG without patient-specific calibration. We compare REVE-base, LUNA-large and LuMamba-Tiny with an adaptive thresholding baseline and a task-specific EEGNet baseline. Additionally, we complement conventional EEG window-based classification with event-based burst detection evaluation. This helps assessing clinically whether burst episodes are correctly detected, reducing the impact of expected annotation variability. The best model, REVE-base, achieved the highest event-based F1-score ($0.868 \pm 0.167$) and reduced burst-per-minute error by 52.1% and 36.2% compared to EEGNet and adaptive thresholding respectively, supporting FMs for scalable EEG monitoring in ICU. Ablation experiments showed that full fine-tuning was the most effective adaptation strategy with respect to frozen-backbone training, two-step fine-tuning, and LoRA-based adaptation, improving event-based F1-score over frozen-backbone training by up to $+0.102$ for LUNA-large. With reduced labeled datasets, pretrained REVE-base outperformed random initialization by $+0.723$ event-based F1 points at 25% of the cohort, demonstrating the benefit of pretraining FM representations when adapted to burst detection with limited labeled data.

• The paper presents a systematic evaluation showing that pre-trained EEG foundation models, especially REVE-base, achieve high event-based F1 scores and significantly lower BPM error without patient-specific calibration.
• It compares foundation models against traditional techniques, demonstrating that full encoder fine-tuning outperforms partial adaptation methods like LoRA or two-step processes.
• The study highlights that large-scale pretraining dramatically improves data efficiency, making these models promising for real-time, objective ICU EEG monitoring.

Evaluation of EEG Foundation Models for Event-Based Burst-Suppression Detection in ICU

Clinical Context and Challenges

The detection of burst-suppression (BS) in ICU EEG is a critical yet technically challenging task due to considerable signal variability, patient-specific patterns, and artifacts inherent in intensive care environments. BS—defined by alternating high-voltage bursts and low-voltage suppressions—serves as a key marker for monitoring sedation depth, guiding therapy during refractory status epilepticus and intracranial hypertension, and evaluating neuroprotection [hirschAmericanClinicalNeurophysiology2021a], [maElectroencephalographicBurstSuppressionPerioperative2022]. Conventional bedside BS detection relies on manual visual inspection or simplistic EEG indices (e.g., BIS), both prone to inter-rater variability and lack direct event quantification [nasrawayHowReliableBispectral2002], [brandonwestoverRealtimeSegmentationBurst2013a], [cottenceauUseBispectralIndex2008].

Automated approaches historically employed handcrafted features and classical ML (e.g., SVM, FLD) [lofhedeClassificationBurstSuppression2008], thresholding [brandonwestoverRealtimeSegmentationBurst2013a], spectral clustering [narulaDetectionEEGBurstsuppression2021a], and task-specific CNNs (e.g., EEGNet). These are often artifact-sensitive and/or require patient-specific calibration, limiting scalability in real-world deployment [loboDoesElectroencephalographicBurst2021a].

Foundation Model Approach and Methodology

This work undertakes the first systematic evaluation of EEG foundation models (FMs) for generalized burst detection in reduced-montage ICU EEG, eliminating patient-specific calibration. The study contrasts pre-trained FMs—REVE-base [ouahidiREVEFoundationModel], LUNA-large [donerLUNAEfficientTopologyAgnostic2025], LuMamba-Tiny [broustailLuMambaLatentUnified2026]—against adaptive thresholding (RVT) and task-specific EEGNet [lawhernEEGNetCompactConvolutional2018]. The FMs leverage montage-agnostic architectures: REVE incorporates versatile positional embeddings for arbitrary electrode arrangements; LUNA unifies variable montages in a topology-agnostic latent space using learned-query cross-attention; LuMamba integrates channel unification with efficient state-space temporal modeling.

EEG data were collected from 25 ICU patients under deep sedation, using six-channel bipolar montage. Burst annotations by clinical experts provided ground truth, and downstream models were trained on 2s windows with a 1s step. Evaluation utilized leave-one-subject-out cross-validation, assessing window-based F1, event-based F1, and burst-per-minute (BPM) mean absolute error (MAE), emphasizing metrics robust to annotation variability and clinically meaningful event detection [danSzCORESeizureCommunity2025].

Quantitative Performance and Adaptation Results

Strong numerical results substantiate the superiority of FMs. REVE-base achieved event-based F1 = $0.868 \pm 0.167$ and BPM MAE = $0.448 \pm 0.284$, outperforming EEGNet ($0.720 \pm 0.245$ F1, MAE $0.936 \pm 0.645$) and RVT ($0.812 \pm 0.233$ F1, MAE $0.702 \pm 0.369$). REVE-base reduced BPM error by 52.1% versus EEGNet and 36.2% versus RVT, generating more precise burst-count estimates with lower interfold variability and minimal calibration. Notably, FMs consistently exceeded the unsupervised spectral-clustering baseline previously validated on this cohort (MAE $0.93 \pm 1.38$) [narulaDetectionEEGBurstsuppression2021a].

Ablation studies on adaptation strategies revealed full fine-tuning as superior, increasing event-based F1 relative to frozen-backbone by $+0.102$ for LUNA-large, $+0.077$ for REVE-base, and $+0.029$ for LuMamba-Tiny. LoRA-based adaptation and two-step fine-tuning underperformed full fine-tuning, indicating that optimal performance in BS detection is achieved by updating all encoder parameters.

Crucially, pretraining provided substantial benefits under limited annotation. With only 25% of labeled training data, pretrained REVE-base exceeded random initialization by $0.448 \pm 0.284$0 F1 points, demonstrating increased data efficiency—a critical feature for practical clinical deployment where labeled ICU EEG is scarce. The magnitude of the pretraining advantage correlated with pretraining corpus size: REVE’s pretraining on >60k hours of EEG from 25k subjects led to more pronounced improvements compared to LUNA and LuMamba (21k hours).

Implications for Clinical Practice and Theory

These findings establish FMs, particularly REVE-base, as state-of-the-art for generalized, event-based BS detection in ICU EEG. The models offer robust generalization across patients and recording conditions, circumventing the need for patient-specific calibration—a major barrier in routine ICU workflows. The substantial reduction in BPM error supports clinically relevant, objective monitoring of sedation depth, facilitating data-driven titration and real-time assessment.

The demonstration that pretraining confers major efficiency gains under annotation scarcity has significant implications for future algorithm deployment, especially in neurocritical care. The event-based evaluation metric, aligned with clinical practice, underscores the need to move beyond window-level metrics in downstream EEG applications.

On the theoretical front, the results highlight that the benefits of foundation models extend to highly variable biomedical signals, reinforcing neural architectures with montage invariance and large-scale pretraining as a generalizable paradigm for clinical EEG. Fine-tuning of all encoder parameters is the key for optimal transfer, contrary to established findings in LLMs where LoRA or partial adaptation often suffice [huLoRALowRankAdaptation2021a].

Future Directions

Despite robust performance in a relatively small cohort, further research should validate FM performance on larger and external ICU datasets, encompassing wider etiologies and co-morbidities. Investigation of continued pretraining on ICU BS-specific recordings may further enhance downstream adaptation. Extending FM frameworks to unsupervised or semi-supervised event detection will address annotation limitations and support scalable neurocritical monitoring. Integration into real-time EEG monitoring platforms and benchmark evaluations across tasks such as seizure detection or depth-of-sedation tracking represents the next frontier.

Conclusion

The evaluation demonstrates that EEG foundation models, epitomized by REVE-base, surpass conventional and task-specific methods for event-based burst-suppression detection in ICU EEG, achieving superior clinical interpretability and data efficiency. Full encoder fine-tuning and large-scale pretraining are pivotal for optimal transfer. Moving forward, scalable, montage-agnostic FMs hold promise for objective, real-time ICU brain monitoring and broader clinical EEG analytics (2606.20074).