pith. sign in

arxiv: 2606.20074 · v1 · pith:XQKGS3O4new · submitted 2026-06-18 · 📡 eess.SP · cs.AI· cs.LG

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

Elisa Vasta , Thorir Mar Ingolfsson , Andrea Cossettini , Luca Benini , Tilman Beck , Emanuela Keller , Una Pale This is my paper

Pith reviewed 2026-06-26 15:52 UTC · model grok-4.3

classification 📡 eess.SP cs.AIcs.LG
keywords burst suppressionEEG foundation modelsICU monitoringevent-based detectionfine-tuningsedation depthreduced montage EEG
0
0 comments X

The pith

EEG foundation models detect burst-suppression patterns in ICU recordings without patient-specific calibration.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper evaluates EEG foundation models on the task of identifying burst-suppression events in reduced-montage ICU EEG data. It reports that REVE-base reaches the highest event-based F1-score of 0.868 and cuts burst-per-minute error by 52 percent relative to a dedicated EEGNet model and 36 percent relative to adaptive thresholding. Full fine-tuning proves more effective than frozen-backbone or LoRA approaches, and the pretrained models retain strong performance when only 25 percent of the labeled cohort is available. A reader would care because reliable automatic detection could support continuous sedation monitoring across many patients where annotated data remain limited.

Core claim

Pretrained EEG foundation models can be adapted for burst-suppression detection in ICU settings. REVE-base achieves an event-based F1-score of 0.868 plus or minus 0.167 and lowers burst-per-minute error by 52.1 percent compared with EEGNet and by 36.2 percent compared with adaptive thresholding. Full fine-tuning improves results over frozen-backbone training by up to 0.102 in F1 for LUNA-large. With only 25 percent of the data, pretrained REVE-base outperforms random initialization by 0.723 F1 points.

What carries the argument

Event-based burst detection scoring, which measures correct identification of entire burst episodes rather than isolated time windows, applied after fine-tuning of the foundation model backbones.

If this is right

  • Full fine-tuning of the entire foundation model yields higher event-based F1 scores than frozen-backbone, two-step, or LoRA adaptation.
  • Pretraining provides a clear advantage when labeled data are reduced to 25 percent of the cohort.
  • The event-based metric better reflects clinical utility by focusing on episode detection rather than window accuracy.
  • The results indicate foundation models can support scalable ICU EEG monitoring without per-patient retraining.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar fine-tuning strategies could be tested on other clinical EEG patterns that also show high patient-to-patient variability.
  • Performance gains might increase if the foundation models were pretrained on larger and more diverse ICU recordings.
  • Real-time deployment would require checking latency and robustness to common ICU artifacts not emphasized in the current evaluation.

Load-bearing premise

The datasets used capture enough real-world ICU variability that the reported performance will hold for new patients without any patient-specific adjustment.

What would settle it

REVE-base producing a lower event-based F1-score than EEGNet when tested on a new multi-center ICU EEG collection that includes greater inter-patient and inter-site differences.

Figures

Figures reproduced from arXiv: 2606.20074 by Andrea Cossettini, Elisa Vasta, Emanuela Keller, Luca Benini, Thorir Mar Ingolfsson, Tilman Beck, Una Pale.

Figure 1
Figure 1. Figure 1: REVE-base performance (mean event-based F1-score [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
read the original 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper evaluates three EEG foundation models (REVE-base, LUNA-large, LuMamba-Tiny) for event-based burst-suppression (BS) detection on reduced-montage ICU EEG without patient-specific calibration. It compares them against EEGNet and adaptive thresholding baselines, reports REVE-base as best with event-based F1-score 0.868 ± 0.167 and burst-per-minute error reductions of 52.1% and 36.2%, shows full fine-tuning as the strongest adaptation strategy, and demonstrates pretraining benefits at 25% labeled data (+0.723 F1). The work emphasizes event-based metrics to mitigate annotation variability and claims support for scalable ICU monitoring.

Significance. If the results hold under proper patient-wise evaluation, the paper would provide the first empirical evidence that EEG foundation models can outperform task-specific networks and simple baselines for BS detection in data-scarce ICU settings. The event-based scoring and limited-data ablation are concrete strengths that directly address clinical annotation challenges. The reported variability and ablation results on adaptation strategies add practical value for deployment.

major comments (2)
  1. [Abstract/Methods] Abstract and Methods: The central claim of generalizability across patients without calibration rests on the evaluation data capturing inter-patient BS variability. No information is given on cohort size (number of patients), number of centers, total recording hours, or the train/test split procedure (patient-wise vs. random or window-wise). This detail is required to assess whether the held-out F1 and error reductions support the no-calibration conclusion or are vulnerable to distribution shift.
  2. [Results] Results section (performance table or text reporting 0.868 ± 0.167): The standard deviation of 0.167 on event-based F1 is large relative to the mean difference versus baselines. The manuscript should clarify whether this variability is computed across patients, cross-validation folds, or seeds, and whether any statistical test (e.g., Wilcoxon or paired t-test) establishes that REVE-base is significantly superior; without this, the superiority claim for scalable monitoring is not yet load-bearing.
minor comments (2)
  1. [Methods] The definition and implementation of the event-based F1 (e.g., tolerance window for burst onset/offset, handling of merged events) should be stated explicitly, ideally with a reference to the scoring code or pseudocode.
  2. [Ablation experiments] Table or figure reporting the ablation on adaptation strategies (full fine-tuning vs. frozen, two-step, LoRA) would benefit from also showing parameter counts or training time to contextualize the +0.102 F1 gain for LUNA-large.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive comments. We address each major comment point-by-point below and have revised the manuscript to incorporate the requested clarifications.

read point-by-point responses

  1. Referee: [Abstract/Methods] Abstract and Methods: The central claim of generalizability across patients without calibration rests on the evaluation data capturing inter-patient BS variability. No information is given on cohort size (number of patients), number of centers, total recording hours, or the train/test split procedure (patient-wise vs. random or window-wise). This detail is required to assess whether the held-out F1 and error reductions support the no-calibration conclusion or are vulnerable to distribution shift.

    Authors: We agree these details are necessary to substantiate the generalizability claim. The manuscript will be revised to explicitly state the cohort size, number of centers, total recording hours, and confirm the patient-wise train/test split in both the Abstract and Methods sections. This will directly address the concern about potential distribution shift and strengthen the no-calibration conclusion. revision: yes

  2. Referee: [Results] Results section (performance table or text reporting 0.868 ± 0.167): The standard deviation of 0.167 on event-based F1 is large relative to the mean difference versus baselines. The manuscript should clarify whether this variability is computed across patients, cross-validation folds, or seeds, and whether any statistical test (e.g., Wilcoxon or paired t-test) establishes that REVE-base is significantly superior; without this, the superiority claim for scalable monitoring is not yet load-bearing.

    Authors: We concur that the source of the reported variability and statistical significance testing should be clarified to support the superiority claims. The ±0.167 is the standard deviation computed across patients. We will revise the Results section to explicitly note this computation method and add the outcomes of a paired statistical test (Wilcoxon signed-rank test) comparing REVE-base against the baselines. These changes will make the performance claims more robust. revision: yes

Circularity Check

0 steps flagged

Empirical model evaluation on held-out data exhibits no circularity

full rationale

The paper reports measured performance metrics (event-based F1-scores, burst-per-minute error reductions) from standard fine-tuning and evaluation of EEG foundation models against independent baselines on held-out ICU EEG data. No mathematical derivations, self-referential equations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the abstract or described methodology; results are obtained via conventional supervised adaptation and external benchmarking rather than any construction that reduces to its own inputs.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central performance claims rest on the representativeness of the ICU EEG datasets used for training and testing, the validity of the event-based scoring protocol, and the assumption that foundation model pretraining transfers without domain-specific architectural changes.

free parameters (1)
  • adaptation strategy hyperparameters
    Choices such as learning rate, epochs, and LoRA rank for fine-tuning are selected to optimize reported metrics.
axioms (1)
  • domain assumption The selected ICU EEG recordings capture the full range of burst-suppression variability across patients
    Evaluation without patient-specific calibration assumes the test set is representative of unseen patients.

pith-pipeline@v0.9.1-grok · 5873 in / 1309 out tokens · 28734 ms · 2026-06-26T15:52:25.994216+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

20 extracted references

  1. [1]

    Clinical Practice Guidelines for the Prevention and Management of Pain, Agitation/Sedation, Delirium, Immobility, and Sleep Disruption in Adult Patients in the ICU,

    J. W. Devlinet al., “Clinical Practice Guidelines for the Prevention and Management of Pain, Agitation/Sedation, Delirium, Immobility, and Sleep Disruption in Adult Patients in the ICU,”Critical Care Medicine, vol. 46, p. e825, 2018

    2018

  2. [2]

    Consensus Statement on Continuous EEG in Critically Ill Adults and Children, Part I: Indications,

    S. T. Hermanet al., “Consensus Statement on Continuous EEG in Critically Ill Adults and Children, Part I: Indications,” vol. 32, 2015

    2015

  3. [3]

    The Use of Bispectral Index to Monitor Barbiturate Coma in Severely Brain-Injured Patients with Refractory Intracranial Hypertension,

    V . Cottenceauet al., “The Use of Bispectral Index to Monitor Barbiturate Coma in Severely Brain-Injured Patients with Refractory Intracranial Hypertension,”Anesthesia & Analgesia, vol. 107, p. 1676, 2008

    2008

  4. [4]

    American Clinical Neurophysiology Society’s Standardized Critical Care EEG Terminology: 2021 Version,

    L. J. Hirschet al., “American Clinical Neurophysiology Society’s Standardized Critical Care EEG Terminology: 2021 Version,”Journal of Clinical Neurophysiology, vol. 38, pp. 1–29, 2021

    2021

  5. [5]

    Real-time segmentation of burst suppression patterns in critical care EEG monitoring,

    M. Brandon Westoveret al., “Real-time segmentation of burst suppression patterns in critical care EEG monitoring,”Journal of Neuroscience Methods, vol. 219, pp. 131–141, 2013

    2013

  6. [6]

    Electroencephalographic Burst-Suppression, Perioperative Neuroprotection, Postoperative Cognitive Function, and Mortality: A Focused Narrative Review of the Literature,

    K. Maet al., “Electroencephalographic Burst-Suppression, Perioperative Neuroprotection, Postoperative Cognitive Function, and Mortality: A Focused Narrative Review of the Literature,”Anesthesia & Analgesia, vol. 135, p. 79, 2022

    2022

  7. [7]

    Does electroencephalographic burst suppression still play a role in the perioperative setting?

    F. A. Loboet al., “Does electroencephalographic burst suppression still play a role in the perioperative setting?”Best Practice & Research Clinical Anaesthesiology, vol. 35, pp. 159–169, 2021

    2021

  8. [8]

    Classification of burst and suppression in the neonatal electroencephalogram,

    J. L ¨ofhedeet al., “Classification of burst and suppression in the neonatal electroencephalogram,”Journal of Neural Engineering, vol. 5, 2008

    2008

  9. [9]

    Detecting the depth of sedation in the intensive care unit using a 2-channel electroencephalogram: An analysis with 2 machine learning models,

    E. A. Alarc ´on-Bragaet al., “Detecting the depth of sedation in the intensive care unit using a 2-channel electroencephalogram: An analysis with 2 machine learning models,”Neuroscience Informatics, vol. 6, p. 100238, 2026

    2026

  10. [10]

    Time-Frequency Domain Classifier for Propofol- Mediated Unconsciousness,

    S. Orgucet al., “Time-Frequency Domain Classifier for Propofol- Mediated Unconsciousness,” in2025 47th Annual International Confer- ence of the IEEE Engineering in Medicine and Biology Society (EMBC), 2025, pp. 1–7

    2025

  11. [11]

    Burst suppression probability algorithms: State- space methods for tracking EEG burst suppression,

    J. Chemaliet al., “Burst suppression probability algorithms: State- space methods for tracking EEG burst suppression,”Journal of Neural Engineering, vol. 10, p. 056017, 2013

    2013

  12. [12]

    IRIS: A Modular Platform for Continuous Monitoring and Caretaker Notification in the Intensive Care Unit,

    S. N. Baldassanoet al., “IRIS: A Modular Platform for Continuous Monitoring and Caretaker Notification in the Intensive Care Unit,”IEEE Journal of Biomedical and Health Informatics, vol. 24, 2020

    2020

  13. [13]

    Detection of EEG burst-suppression in neurocritical care patients using an unsupervised machine learning algorithm,

    G. Narulaet al., “Detection of EEG burst-suppression in neurocritical care patients using an unsupervised machine learning algorithm,”Clinical Neurophysiology, vol. 132, pp. 2485–2492, 2021

    2021

  14. [14]

    Monitoring burst suppression in critically ill patients: Multi-centric evaluation of a novel method,

    F. F¨urbasset al., “Monitoring burst suppression in critically ill patients: Multi-centric evaluation of a novel method,”Clinical Neurophysiology, vol. 127, pp. 2038–2046, 2016

    2038

  15. [15]

    LUNA: Efficient and Topology-Agnostic Foundation Model for EEG Signal Analysis,

    B. D ¨oneret al., “LUNA: Efficient and Topology-Agnostic Foundation Model for EEG Signal Analysis,” 2025

    2025

  16. [16]

    REVE: A Foundation Model for EEG Adapting to Any Setup with Large-Scale Pretraining on 25,000 Subjects

    Y . E. Ouahidiet al., “REVE: A Foundation Model for EEG Adapting to Any Setup with Large-Scale Pretraining on 25,000 Subjects.”

  17. [17]

    LuMamba: Latent Unified Mamba for Electrode Topology-Invariant and Efficient EEG Modeling,

    D. Broustailet al., “LuMamba: Latent Unified Mamba for Electrode Topology-Invariant and Efficient EEG Modeling,” 2026

    2026

  18. [18]

    EEGNet: A compact convolutional neural network for EEG-based brain–computer interfaces,

    V . J. Lawhernet al., “EEGNet: A compact convolutional neural network for EEG-based brain–computer interfaces,”Journal of Neural Engineering, vol. 15, p. 056013, 2018

    2018

  19. [19]

    LoRA: Low-Rank Adaptation of Large Language Models,

    E. J. Huet al., “LoRA: Low-Rank Adaptation of Large Language Models,” 2021

    2021

  20. [20]

    SzCORE: Seizure Community Open-Source Research Evaluation framework for the validation of electroencephalography-based automated seizure detection algorithms,

    J. Danet al., “SzCORE: Seizure Community Open-Source Research Evaluation framework for the validation of electroencephalography-based automated seizure detection algorithms,”Epilepsia, vol. 66, pp. 14–24, 2025

    2025