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

Researchers evaluated EEG Foundation Models for detecting burst-suppression patterns in ICU patients, finding that REVE-base achieved superior performance with an F1-score of 0.868 and reduced errors by up to 52% compared to existing methods. This study demonstrates the practical value of pretrained AI models for clinical EEG monitoring without patient-specific calibration, particularly when labeled data is limited.

This research addresses a critical gap in clinical AI by applying foundation models to burst-suppression detection, a medically essential pattern for monitoring sedation depth in intensive care settings. Burst suppression represents periods of brain inactivity alternating with bursts of electrical activity, and accurate detection directly impacts patient safety and treatment decisions. The study's significance lies in moving beyond traditional window-based classification to event-based detection metrics that better reflect clinical relevance.

The advancement comes at a crucial moment when healthcare systems face EEG interpretation bottlenecks due to insufficient specialist availability. Foundation models pretrained on large EEG datasets provide a pathway to scalable clinical deployment without requiring extensive patient-specific calibration. The comparison between REVE-base, LUNA-large, and LuMamba-Tiny demonstrates that larger models don't necessarily outperform smaller ones, with REVE-base achieving the best results despite not being the largest architecture tested.

The ablation studies reveal important technical insights: full fine-tuning consistently outperformed parameter-efficient methods like LoRA, suggesting that adaptation depth matters for specialized clinical tasks. Most compelling is the 0.723 F1-score improvement of pretrained REVE-base over random initialization when using only 25% of labeled data, establishing pretraining's substantial value for data-scarce medical applications.

This research positions foundation models as pragmatic tools for hospital ICUs seeking to augment clinical staff capacity. The methodology's focus on reduced-montage EEG—fewer electrode channels than standard setups—enhances real-world applicability since many ICUs operate with simpler equipment. Future validation across multiple institutions and diverse patient populations remains essential before clinical deployment.

• →REVE-base foundation model achieved 0.868 event-based F1-score for burst-suppression detection, outperforming EEGNet and adaptive thresholding baselines by 52% and 36% respectively.
• →Full fine-tuning proved most effective for clinical EEG tasks, improving performance by up to 0.102 F1-points compared to frozen-backbone approaches.
• →Pretrained models demonstrated substantial advantages with limited labeled data, showing 0.723 F1-point improvement over random initialization at 25% data availability.
• →Event-based detection metrics better capture clinical relevance than traditional window-based classification by accounting for annotation variability in ICU settings.
• →The use of reduced-montage EEG makes foundation models more practically deployable across diverse hospital ICU configurations with varying equipment capabilities.