Neurophysiologic Encephalopathy Severity Index (NESI) - Data and Code

Abstract

Disorders of consciousness are assessed with multiple bedside scales that measure overlapping but non-identical aspects of neurological function, producing fragmented and often discordant assessments in acute and critical care. This resource accompanies the Neurophysiologic Encephalopathy Severity Index (NESI), a continuous electroencephalography (EEG)-based measure of brain dysfunction that places patients on a single physiological continuum spanning alertness, delirium, sedation, and coma.

NESI is derived by extracting generalized EEG representations from a clinical EEG foundation model (MORGOTH), compressing them with a contrastive encoder, and learning a continuous severity score with a pairwise ranking head trained jointly across four clinical scales (RASS, GCS, CAM-S, and ICANS). On a held-out test set, NESI agreed strongly with all four scales (Spearman correlations 0.72-0.84), showed higher AUROC than GCS for in-hospital mortality, and tracked propofol effect-site concentration more closely than RASS.

This release provides the curated 10-minute EEG segments, the MORGOTH feature activations, the assembled training/validation/test data, trained model checkpoints, the propofol pharmacokinetic and medication-exposure derivatives, and per-segment metadata, together with the complete analysis code (mirrored on GitHub). It is intended to let credentialed users retrain the models and reproduce the figures and tables from the ground up.

Background

Accurate assessment of neurologic status is essential to the care of acute and critically ill patients, yet current practice approximates the continuum of acute brain dysfunction through a patchwork of behavioral scales developed in different eras for distinct populations: the Glasgow Coma Scale (GCS), the Richmond Agitation-Sedation Scale (RASS), the Confusion Assessment Method-Severity (CAM-S), and immune effector cell-associated neurotoxicity syndrome (ICANS) grading, among others. These scales measure overlapping but non-identical constructs, can be discordant, and are limited by inter-rater variability and by focal deficits that mask underlying status.

EEG provides a direct measure of cerebral activity that captures the physiology underlying acute encephalopathy, and prior machine-learning work has shown that EEG carries enough information to approximate individual behavioral scores. However, those instruments were built for specific populations and single targets and were never integrated into a common framework. We treat the scales as noisy observations of one shared latent construct - the severity of acute encephalopathy - and recover that physiologic dimension directly from EEG. The result, NESI, unifies the four scales on a single continuous axis, which lets scores from different instruments be related to one another under a shared construct.

Methods

Cohorts and data source

Data were drawn from the Harvard EEG Database (HEEDB), a large de-identified corpus of multichannel EEG from four Harvard-affiliated hospitals linked to electronic health records. Cohorts were assembled for RASS, GCS, and CAM-S by cross-matching behavioral-assessment timestamps with concurrent EEG; the ICANS cohort is a separate retrospective dataset with daily ICANS grades paired with continuous EEG and age/sex-matched outpatient controls. Inclusion required age ≥ 18, inpatient status, and concurrent assessment and EEG. Observations with RASS > 0 were excluded from downstream analysis.

Signal processing and quality assessment

EEG was recorded with 19 channels (international 10-20 system, 200 Hz). For RASS, GCS, and CAM-S, the 10-minute segment immediately preceding each assessment was extracted; for ICANS, which lacks per-grade timestamps, the 10-minute segment maximizing awake activity (via the MORGOTH sleep head) was selected. Signals were band-pass filtered (0.5-70 Hz) and notch-filtered (60 and 50 Hz). A signal-quality-assessment (SQA) block flagged segments with high-amplitude or flat artifacts and excluded low-quality segments.

Feature extraction and model

Each 10-minute segment was passed through the MORGOTH clinical-EEG foundation model, producing a 591×17 matrix of event-level feature activations. A supervised-contrastive ResNet encoder compresses this matrix to a 40-dimensional embedding; a pairwise ranking head (a parameter-shared Siamese MLP) then maps embeddings to a single scalar NESI value. Training uses within-scale ordinal pairs and cross-scale pairs (assembled into a unified dataset, SPECTRA) so that severity is learned on one scale-agnostic axis. Data were split subject-independently 70/10/20.

Sedation and medication analyses

Medication exposure was computed from the medication administration record over a 24-hour pre-EEG window. For propofol, effect-site concentration (C_e) was estimated with the Eleveld (2018) pharmacokinetic-pharmacodynamic model, and the variance in NESI and RASS attributable to C_e was decomposed with linear mixed-effects models (patient random intercept) and cluster-bootstrap confidence intervals.

Code

All preprocessing, training, evaluation, figure, and table code is provided in the accompanying GitHub repository (see Usage Notes), with a REPRODUCE.md mapping each figure and table to its script.

Data Description

All files live under the credentialed S3 prefix yama/. The dataset is organized by cohort, plus a NESI/ module with the assembled modeling data and the sedation/medication derivatives. EEG segments are MATLAB .mat arrays (19 channels at 200 Hz); MORGOTH activations are the per-segment 591×17 feature matrices. Data are de-identified (surrogate subject IDs; shifted dates).

Top-level layout

yama/
  README.md, LICENSE.txt
  segment_index.csv          unified per-segment index (all cohorts) + source-EEG provenance
  source_eeg_files.csv       unique continuous source EEG recordings (all cohorts)
  RASS/ GCS/ CAMS/ ICANS/    per cohort: *_EEG10minSegments/, MorgothActivations/,
                             Cohort/ (HEEDB metadata), *Training_Final_Metadata.csv, ModelCheckpoints/
  Death/                     Death_EEG10minSegments/, MorgothActivations/, Cohort/ (cohort + labels)
  DiagnosisMetadtafiles/     admission-diagnosis tables
  NESI/
    Data/                    SPECTRA train/val/test feature pickles, embeddings, NESI result CSVs
    ModelCheckpoints/        trained universal model (ResNetGAP_BestModel.pth + NESI_best_model.pth)
    Bespoke_models/Results/  per-scale bespoke score CSVs
    NESI-Medication-Analysis/ propofol Eleveld PK + medication-exposure derivatives

Cohort sizes

CAM-S (526) and ICANS (784) segment counts equal the final analysis sets; RASS (118,330) and GCS (124,295) contain the full pre-filtering set. The manuscript valid-observation counts (RASS 97,620; GCS 110,836; CAM-S 526; ICANS 784; total 209,766) match NESI/Data/NESI_*_results.csv.

Index and source-EEG provenance

segment_index.csv has one row per paired observation with the segment filename, subject/study id, assessment time, raw score, NESI value, and source-EEG provenance. Provenance is resolved for every segment: complete for RASS/GCS (BidsFolder, SessionID, source begin/end), and recovered for CAM-S (source start + 10-minute snippet window) and ICANS (mapped to the companion ICANS dataset, s3://bdsp-opendata-credentialed/icans/, via ICANSFiles.xlsx; BidsFolder gives the source path and resolved_BDSPPatientID the patient id). source_eeg_files.csv lists the unique continuous source recordings the segments were cut from.

Notes

Death/Cohort/Death_labels.csv is a per-patient mortality reference; the exact mortality-analysis cohort (5,675 subjects, 1,333 in-hospital deaths) is defined by the sequential-logistic-regression code in the GitHub repository. Per-cohort de-identified HEEDB metadata are under each <cohort>/Cohort/.

Usage Notes

Code on GitHub

The complete preprocessing, training, evaluation, and figure/table code is at https://github.com/bdsp-core/NESI (public; browseable without credentialed access). See NESI/REPRODUCE.md for a figure/table→script map and NESI/requirements.txt for the Python 3.9 environment. Trained universal and bespoke model weights are committed in the repository under NESI/model/ModelCheckpoints/ and NESI/Bespoke_models/ModelCheckpoints/.

Downloading

Access requires credentialed approval and the Data Use Agreement. With credentials configured, sync a cohort with the AWS CLI, e.g.:

aws s3 sync s3://bdsp-opendata-credentialed/yama/CAMS/ ./CAMS/ --profile opendata

Loading

import pandas as pd, scipy.io as sio
idx  = pd.read_csv("segment_index.csv")                      # per-segment index + provenance
meta = pd.read_csv("CAMS/CAMSTraining_Final_Metadata.csv")
eeg  = sio.loadmat("CAMS/CAMS_10minEEGSegments/<file>.mat")   # 19 x samples @ 200 Hz

Assembled MORGOTH feature inputs and labels for modeling are in NESI/Data/NESITripletIP_Morgoth_{train,val,test}_data.pkl with matching NESI_{train,val,test}_results.csv.

Release Notes

Version 1.0.0 - initial release accompanying the NESI manuscript. Includes the curated 10-minute EEG segments, MORGOTH feature activations, assembled SPECTRA training/validation/test data, baseline model checkpoints, propofol pharmacokinetic and medication-exposure derivatives, and per-segment metadata, together with the full analysis code on GitHub.

Ethics

The study was conducted with ethical approval from the institutional review boards of Beth Israel Deaconess Medical Center (protocols #2022P000481 and #2022P000417) and Stanford University (protocol #69873), with a waiver of informed consent for retrospective analysis. All data are de-identified.

Acknowledgements

Arka Roy and Katelyn Surrao contributed equally as co-first authors. Eyal Y. Kimchi, Sahar F. Zafar, Christine A. Eckhardt, and M. Brandon Westover contributed equally as co-senior authors.

This work received support from the National Institutes of Health (R01AG073410, R01HL161253, R01NS126282, R01AG073598, R01NS131347, R01NS130119).

Conflicts of Interest

Dr. Westover is a co-founder of, serves as a scientific advisor and consultant to, and has a personal equity interest in Beacon Biosignals. The remaining authors report no competing interests.