ReMAP: Self-supervised learning to unveil brain representations and vulnerability

Jade Perdereau; Virginie Loison; Kanssa El Ayeb; Louis Gervais; Melvin Berto Strouc; Fabrice Vallée; Thomas Moreau; Jérôme Cartailler

cs.LGAug 22, 2026

ReMAP: Self-supervised learning to unveil brain representations and vulnerability

Authors: Jade Perdereau, Virginie Loison, Kanssa El Ayeb, Louis Gervais, Melvin Berto Strouc, Fabrice Vallée, Thomas Moreau, Jérôme Cartailler

Organizations: AP-HP, Hôpital Lariboisière, Paris, France · UMR-942, Inserm Délégation Régionale Paris 7, Bagnolet, France · Université Paris Cité, Boulogne-Billancourt, France · Université Paris-Saclay, Inria, CEA, Palaiseau, France ·

Abstract

General anesthesia offers a rare opportunity to observe the human brain under a standardized, controlled perturbation. Yet intraoperative electroencephalography (EEG) is almost always reduced to a single proprietary depth index, collapsing a rich trajectory into one number and discarding how a brain moves between states. Here we ask whether the geometry of that trajectory, not merely the depth it reaches, carries clinically meaningful information. Using similarity-based self-supervised learning on raw, two-electrode frontal EEG, with no labels, we place each recording within a low-dimensional space in which anesthetic depth becomes one readable axis while the shape of a patient's path encodes additional structure. We validate the representation across two cohorts and two acquisition systems totaling more than 1,000 patients. Depth of anesthesia is predicted accurately (BIS mean absolute error = 3.2, R2 = 0.82), and in the sparse-montage setting our compact ( 68k parameter) model remains competitive with EEG foundation models orders of magnitude larger (4M-157M parameters), indicating that matching the representation to the recording dominates raw scale. The learned space organizes age along its own gradient, independent from depth, without supervision. The same space also aligns with interpretable anesthetic signatures like frontal alpha, slow-delta, and burst suppression, linking this data-driven representation to established neurophysiology. On an independent cohort with longitudinal follow-up, the geometry of the early trajectory separates 30- month cognitive and mortality outcomes complementary to age (AUROC 0.86). These results suggest that the path a brain traces through anesthesia is a label-efficient correlate of latent vulnerability, motivating prospective validation.

Figures & tables

Figure 1: Depth of anesthesia prediction on VitalDB test dataset ( $n_{\text{patients}}=173$ , $n_{\text{windows}}=84{,}444$ ). Foundation models are represented in blue, hand-engineered spectral features in orange, supervised in purple and ReMAP (Proposed) in green. (A) Mean absolute error per model. (B) Mean error (bias) per model. (C) $R^{2}$ reconstruction score per model. (D) Predicted vs observed BIS plots per model. Dashed line represents the perfect fit 1–1 line, and the solid line the obtained linear fit on models’ predictions.

Figure 3: Model generalization to a clinical dataset with outcomes ( $n_{\text{patients}}=178$ shown; $n=171$ after quality control for statistical tests, $n_{\text{windows}}=58{,}974$ ). 2D projected embeddings colored by (A) Age distribution per outcome group. (B) Trajectories per outcome group. Shaded background regions show the 2D kernel density estimate of each group’s trajectory points, rendered as nine percentile bands (10th–90th, in 10% increments) with darker shading indicating higher point density. (C) Trajectories per age group. Shaded background regions show the 2D kernel density estimate of each group’s trajectory points, rendered as nine percentile bands (10th–90th, in 10% increments) with darker shading indicating higher point density. (D) Embeddings colored with alpha band power (8–13 Hz). (E) Embeddings colored with delta band power (0.5–4 Hz).

Figure 4: (A-C) Aggregated variable definition for outcome group classification (n = 171 after quality control). (D-E) Initial trajectory modelling. (A). Aggregation strategy of the 100-D trajectory into a 1-D vector based on PCA first component median (B) Aggregated feature distribution among outcome groups; asterisks denote statistical significance assessed by the two-sided Mann-Whitney U test (*p < 0.05, **p < 0.01, ***p < 0.001). (C) Receiver operating characteristic curves per outcome group; shaded bands represent 95% bootstrap confidence intervals estimated over 1,000 resampling iterations. (D) Mean reconstructed trajectories over time per outcome group, obtained by resampling; shaded regions indicate the interquartile range (25th-75th percentile). Thick solid lines over the first 15 time points represent the group-level trajectories estimated by the linear mixed model (LMM); dashed lines show the fitted exponential decay model over the full observation window (E) Early induction dynamics (first 15% of time points) of UMAP coordinate y by outcome group, with group-specific linear mixed-model fits. Asterisks at the right edge of the LMM window denote a statistically significant difference in slope relative to the healthy group (* p < 0.05; *** p < 0.001), as estimated from the LMM interaction term (time × outcome group). — Figure 4: (A-C) Aggregated variable definition for outcome group classification (n = 171 after quality control). (D-E) Initial trajectory modelling. (A). Aggregation strategy of the 100-D trajectory into a 1-D vector based on PCA first component median (B) Aggregated feature distribution among outcome groups; asterisks denote statistical significance assessed by the two-sided Mann-Whitney U test (*p < 0.05, **p < 0.01, **p < 0.001). (C) Receiver operating characteristic curves per outcome group; shaded bands represent 95% bootstrap confidence intervals estimated over 1,000 resampling iterations. (D) Mean reconstructed trajectories over time per outcome group, obtained by resampling; shaded regions indicate the interquartile range (25th-75th percentile). Thick solid lines over the first 15 time points represent the group-level trajectories estimated by the linear mixed model (LMM); dashed lines show the fitted exponential decay model over the full observation window (E) Early induction dynamics (first 15% of time points) of UMAP coordinate y by outcome group, with group-specific linear mixed-model fits. Asterisks at the right edge of the LMM window denote a statistically significant difference in slope relative to the healthy group ( p < 0.05; *** p < 0.001), as estimated from the LMM interaction term (time × outcome group).

Encoder	Macro AUROC	AUC cog_decline_dead	AUC cv_event	AUC healthy	F1 cog_decline_dead	F1 cv_event	F1 healthy
ReMAP	$0.786\pm 0.032$	$0.862$ $[0.797\text{--}0.915]$	$0.645$ $[0.501\text{--}0.775]$	$0.847$ $[0.781\text{--}0.908]$	$0.625$ $[0.489\text{--}0.742]$	$0.222$ $[0.091\text{--}0.342]$	$0.695$ $[0.618\text{--}0.765]$
PSD	$0.762\pm 0.038$	$0.863$ $[0.796\text{--}0.920]$	$0.555$ $[0.423\text{--}0.672]$	$0.864$ $[0.807\text{--}0.919]$	$0.634$ $[0.512\text{--}0.750]$	$0.163$ $[0.037\text{--}0.285]$	$0.763$ $[0.689\text{--}0.823]$
CBraMod	$0.767\pm 0.029$	$0.818$ $[0.741\text{--}0.883]$	$0.643$ $[0.513\text{--}0.775]$	$0.835$ $[0.771\text{--}0.897]$	$0.511$ $[0.374\text{--}0.639]$	$0.120$ $[0.000\text{--}0.241]$	$0.764$ $[0.696\text{--}0.825]$
REVE	$0.694\pm 0.035$	$0.766$ $[0.686\text{--}0.844]$	$0.576$ $[0.427\text{--}0.729]$	$0.756$ $[0.684\text{--}0.834]$	$0.465$ $[0.324\text{--}0.600]$	$0.167$ $[0.038\text{--}0.293]$	$0.699$ $[0.624\text{--}0.767]$
Supervised	$0.691\pm 0.047$	$0.797$ $[0.724\text{--}0.865]$	$0.512$ $[0.363\text{--}0.666]$	$0.799$ $[0.726\text{--}0.870]$	$0.527$ $[0.400\text{--}0.652]$	$0.109$ $[0.000\text{--}0.222]$	$0.709$ $[0.636\text{--}0.777]$
LUNA	$0.690\pm 0.058$	$0.720$ $[0.633\text{--}0.805]$	$0.587$ $[0.441\text{--}0.730]$	$0.745$ $[0.667\text{--}0.823]$	$0.337$ $[0.217\text{--}0.473]$	$0.136$ $[0.000\text{--}0.279]$	$0.676$ $[0.602\text{--}0.744]$

Table 1: Models classification performance across outcome categories and competing methods controlled by age. The original embeddings information is aggregated using the median on PCA first axis before controlling against age. Area under the receiver operating characteristic curve (AUC, with 95% bootstrap confidence intervals in brackets) are reported for three binary classification tasks: cognitive decline or death, healthy ageing, and cardiovascular event. Methods compared include our proposed model (ReMAP), the deep learning supervised baseline, EEG foundation models (CBraMod, REVE, BENDR and LUNA), and a classical power spectral density feature baseline (PSD). Higher values indicate better discriminative performance.

Figure 5: Depth of anesthesia prediction pipeline and benchmark. (A). General study design. (B). Proposed EEG embedder. (C). Pretraining computational framework and proposed self-supervised approach. (D). Development and test set splitting strategy.

Electroencephalography

Electroencephalography (EEG) studies brain-generated electrical activity to understand cognition and support diagnosis, monitoring, and brain-computer interfaces.

See all 183 papers

Mar 17, 2026

Laya: A LeJEPA Approach to EEG via Latent Prediction over Reconstruction

Electroencephalography (EEG) is a widely used tool for studying brain function, with applications in clinical neuroscience, diagnosis, and brain-computer interfaces (BCIs). Recent EEG foundation models trained on large unlabeled corpora aim to learn transferable representations, but their effectiveness remains unclear; reported improvements over smaller task-specific models are often modest, sensitive to downstream adaptation and fine-tuning strategies, and limited under linear probing. We hypothesize that one contributing factor is the reliance on signal reconstruction as the primary self-supervised learning (SSL) objective, which biases representations toward high-variance artifacts rather than task-relevant neural structure. To address this limitation, we explore an SSL paradigm based on Joint Embedding Predictive Architectures (JEPA), which learn by predicting latent representations instead of reconstructing raw signals. We introduce Laya, the first EEG foundation model based on LeJEPA. We show that latent prediction yields representations that encode semantic structure in EEG: Laya embeddings track clinically meaningful state changes such as seizure onset, are resilient to noise, and achieve the strongest mean clinical accuracy under frozen linear probing, with particular gains on tasks where relevant neural patterns are subtle and easily obscured by artifacts. Controlled ablations against matched MAE variants confirm that the choice of pretraining objective, rather than architecture or data, is the primary driver of these gains.

Saarang Panchavati, Uddhav Panchavati, Hiroki Nariai +2

Department of Medical Informatics UCLA Los Angeles, USA · Department of Cognitive Science UCSD San Diego, USA · David Geffen School of Medicine UCLA Mattel Children’s Hospital Los Angeles, USA

EEG Foundation Models Self-Supervised Time Series Representation Learning Predictive Representation Learning Joint-Embedding Predictive Architecture

Ask PDF

Jul 7, 2026

STST-JEPA: Shallow-Target Spatio-Temporal Joint Embedding Prediction Architecture For EEG Self-Supervised Learning

Brain age - the age inferred from a physiological recording - is an emerging biomarker whose deviation from chronological age tracks neurological and psychiatric burden, and EEG is an attractive substrate for it because it is cheap, portable, and temporally rich. Yet EEG brain-age models must contend with cross-site montage heterogeneity, small labelled cohorts, and dominant subject-level non-stationarity, and few EEG foundation models have been shown to deliver competitive age regression across the full pediatric to older adult range in which such a biomarker would actually be deployed. We introduce STST-JEPA, a self-supervised transformer for resting-state and task EEG, pretrained on 47,703 sessions spanning ages 5-81 from the brain.space and Healthy Brain Network (HBN) corpora. The model combines a latent-prediction objective - predicting masked-token representations against an EMA-of-tokenizer target - with an auxiliary signal-reconstruction term, applied to 30-second multi-channel windows under spatiotemporal block masks. A lightweight attentive probe trained on frozen pretrained embeddings achieves a best held-out-validation mean absolute error of 3.06 years (r = 0.924) for age regression on 3,367 sessions, against a predict-the-mean baseline of approximately 10 years MAE. With light task-specific finetuning of the model's final layers, the same pretrained encoder achieves rank-1 placements - with the model's native 30-second windows - on the public NeuralBench x brain.space EEG leaderboard for sex classification (balanced accuracy 0.911), age prediction (r = 0.749), and psychopathology composite regression (r = 0.215). We further show that the model's age-prediction residual is negatively correlated with cognitive efficiency over several tasks we examined.

Roy Segal, Yoni Svechinsky, Tomer Fekete

Brain Age Prediction Self-Supervised Time Series Representation Learning Joint-Embedding Predictive Architecture EEG Foundation Models

Ask PDF

Oct 12, 2026

HyBRA-EEG: Hyperbolic Global-Local Brain-Region Aggregation for Self-Supervised EEG Representation Learning

Electroencephalography (EEG) representation learning is challenged by noisy signals, inter-subject variability, and heterogeneous electrode montages in different datasets and recording settings. Recent self-supervised EEG pretraining methods improve transferability by converting heterogeneous recordings into unified token or sequence representations, but channel-aware designs may still encode montage- or dataset-specific regularities. We propose HyBRA, a self-supervised framework based on hyperbolic global-local brain-region aggregation. HyBRA maps heterogeneous channel observations into a fixed regional token space by combining electrode coordinates, region priors, and feature-dependent hyperbolic distances. A channel-evidence injection branch preserves fine-grained discriminative cues during aggregation. The resulting tokens are processed by a Region-Time Transformer and pretrained with masked reconstruction of latent features and region-level signals. HyBRA outperforms the compared methods in average balanced accuracy on six benchmark datasets. Under a controlled six-fold evaluation on OpenBMI-MI, HyBRA matches CSBrain in balanced accuracy, improves AUROC by 1.84 percentage points and AUPRC by 5.23 percentage points, and uses 96.8% fewer parameters. Matched comparisons on TUAB and TUEV show higher balanced accuracy for hyperbolic matching than for Euclidean distance or cosine similarity.

Xin Liu, Feng Lu, Wei Li

Electroencephalography Self-Supervised Pre-Training Self-Supervised Time Series Representation Learning Invariant Representation Learning Hyperbolic Representation Learning

Ask PDF

Abstract

Electroencephalography

Explore similar work

Laya: A LeJEPA Approach to EEG via Latent Prediction over Reconstruction

STST-JEPA: Shallow-Target Spatio-Temporal Joint Embedding Prediction Architecture For EEG Self-Supervised Learning

HyBRA-EEG: Hyperbolic Global-Local Brain-Region Aggregation for Self-Supervised EEG Representation Learning