Source-linked AI summary
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
TL;DR
Intraoperative EEG is usually compressed into a single depth index, leaving the brain’s trajectory through anesthesia poorly characterized. ReMAP uses self-supervised learning on two-electrode frontal EEG to represent that trajectory in a low-dimensional space. The representation captures anesthetic depth, organizes age independently, and separates 30-month cognitive and survival outcomes in an independent cohort.
Problem
Single-value anesthesia indices capture depth but not the trajectory of brain-state changes, limiting evidence about clinically relevant response patterns.
Method
ReMAP uses similarity-based self-supervised learning on raw two-electrode frontal EEG to embed recordings in a low-dimensional space and analyze trajectory geometry.
Results
The representation recovers anesthetic depth, separates age along an independent gradient, and provides prognostic separation for 30-month cognitive and mortality outcomes.
Takeaways & Limitations
Trajectory geometry may serve as an interpretable, label-efficient correlate of perioperative vulnerability beyond a single anesthesia-depth number.
Takeaways & Limitations
The outcome cohort is single-centre, observational, and modest in size, motivating prospective external validation in a larger multicentre cohort.
Abstract
from arXiv · showhide
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.
1 Introduction
ReMAP addresses the loss of information caused by reducing intraoperative frontal EEG to a single anesthesia-depth index by learning a low-dimensional, interpretable representation from raw, unlabeled two-electrode recordings. The framework treats anesthetic state as a trajectory and captures additional structure, including age-related organization and clinically relevant brain vulnerability.
- Clinical motivation: Routine frontal EEG monitoring is usually collapsed into a proprietary depth index, discarding information about how brain states change during anesthesia.BIS and PSI guide anesthetic titration but reduce a rich EEG signal to one number.
- Clinical motivation: Prolonged burst suppression and reduced anesthetic-induced alpha power are associated with postoperative delirium, cognitive decline, and mortality, motivating richer monitoring of individual brain responses.These EEG patterns have been interpreted as markers of systemic vulnerability rather than modifiable intraoperative causes.
- ReMAP framework: ReMAP uses similarity-based self-supervised learning to map raw two-electrode frontal EEG into a 2D representation without labels.The framework is designed for perioperative brain-state monitoring and postoperative prognosis in operating-room EEG.
- ReMAP framework: The learned representation makes anesthetic state occupy a space rather than a point, separating distinct courses that a depth index may treat as similar.Anesthetic depth becomes one readable direction, while trajectory shape carries additional structure.
- Representation structure: Age forms its own gradient in the representation, providing an unsupervised correlate of brain ageing independent of anesthetic depth.The contribution is presented alongside label-efficient depth capture and competitiveness with larger foundation models in the sparse-montage setting.
2 Results
Across 1,039 patients from VitalDB and LaribDB, ReMAP learned a compact EEG representation that predicted anesthetic depth and organized physiologic and age-related structure. Trajectory summaries transferred across cohorts and carried information about 30-month outcome groups beyond age-controlled benchmarks.
- Depth-of-anesthesia prediction: 3.2 mean absolute error and R2 0.82 were achieved for BIS prediction on the VitalDB test set, outperforming supervised, foundation-model, and classical PSD approaches.The SSL model also showed negligible bias of 0.1±4.0; its comparators included CBraMod, REVE, LUNA, BENDR, and PSD.
- Representation structure: The learned embeddings exhibited a clear anesthesia-depth gradient and distinct regions, while embedding analyses linked the space to physiological properties including spectral EEG features.The representation was projected with UMAP for visualization, while statistical analyses were performed in the original 100-D latent space.
- Longitudinal outcomes: Patient-level trajectory medians significantly discriminated outcome groups, motivating trajectory-based classification after PERMANOVA lacked power.The scalar was the prespecified per-patient median of the [PC1 / y] coordinate, summarized without reference to outcome labels.
3 Discussion
ReMAP treats anesthesia-related brain activity as a trajectory in a low-dimensional space rather than a single depth index. Across cohorts, the learned geometry preserves depth while capturing age-related structure and associations with later cognitive and survival outcomes.
- Routine anesthetic indices compress complex, time-varying frontal EEG into a scalar that provides limited information about how brain states change.
- Similarity-based self-supervised learning on raw two-channel frontal EEG constructs a low-dimensional representation without labels or hand-engineered features.
- 68k parameters enabled accurate BIS recovery from sparse frontal EEG while outperforming hand-engineered features, a supervised model, and 4M-157M-parameter foundation models.The approach was evaluated across two cohorts and two acquisition systems.
- Age organized along a trajectory-shape gradient separated from anesthetic depth, while the representation also captured interpretable alpha, slow-delta, and other physiological structure without supervision.The age gradient had cosine similarity close to zero with the depth axis, which does not prove independence.
- 30-month cognitive and survival outcomes were associated with latent trajectory information during induction in an independent LaribDB cohort, complementing age.Cardiovascular-event classification was weaker and is a natural target for complementary modalities.
- Prospective external validation in a larger, preferably multicentre cohort with multimodal fusion is needed before geometric trajectory biomarkers can be generalized.Topology-based methods such as persistence diagrams are proposed for quantitatively characterizing trajectory shape.
4 Materials and Methods
The study combined patient-level outcome labeling, raw EEG preprocessing, relative-positioning self-supervision, and benchmarking against supervised and foundation-model baselines. It analyzed smoothed UMAP trajectories with mixed-effects and permutation-based methods to quantify outcome-related geometry and clinical-variable alignment.
- Cohort and outcomes: The LaribDB comprised 178 patients grouped as healthy, cognitive decline, cardiovascular event, or death using outcomes defined 30 months after acquisition.Cognitive decline required a greater-than-one-standard-deviation MoCA decrease between baseline and 30 months.
- Data processing: EEG records were split at the patient level into 60/20/20 train, validation, and test sets, excluding windows with mean BIS outside [10, 80].Processing used Braindecode Baseconcat datasets, and foundation-model weights came from HuggingFace artifacts.
- Self-supervised learning: The self-supervised model learned relative temporal positioning from 30-second EEG segments using same-recording pairs and a binary logistic loss on absolute embedding differences.Pairs were labeled by temporal separation, with a shared encoder followed by a linear prediction layer.
- Benchmarking and evaluation: The approach was benchmarked against an end-to-end supervised CNN and EEG foundation models including BENDR, REVE, and CBraMod without fine-tuning.Evaluation used unseen test sets, regression error, absolute error, R2, macro AUC, and class-level F1; age served as a covariate.
- Trajectory analysis: Patient trajectories were resampled to T = 100 UMAP points, Savitzky–Golay-smoothed, and modeled with exponential decay and early-induction linear mixed-effects analyses.Group-specific intercepts and slopes were compared with healthy patients using outcome interactions.
- Embedding statistics: Clinical-variable alignment used covariance-based encoding patterns and cosine similarity, while trajectory differences used PERMANOVA and PERMDISP with 999 permutations.Distances to the healthy centroid were computed in the original 100D embedding space.
6 Data availability
The study used the publicly available VitalDB dataset and a LaribDB dataset available from the authors upon reasonable request, with both cohorts covered by ethics approvals.
- Data availability: VitalDB is publicly available, whereas LaribDB can be obtained from the authors upon reasonable requests.VitalDB was approved by Seoul National University Hospital’s Institutional Review Board and registered at ClinicalTrials.gov; LaribDB participants provided oral informed consent and the study received committee approval.