GROND: Automated Characterization of Periodic and Rhythmic EEG Patterns

Abstract

Objective. Periodic discharges (PDs) and rhythmic delta activity (RDA) are common electroencephalographic (EEG) patterns in critically ill patients that require detailed characterization — including lateralization, spatial localization, and frequency estimation — according to the American Clinical Neurophysiology Society (ACNS) 2021 standardized terminology. Manual characterization is subjective, time-consuming, and exhibits substantial inter-rater variability. This project presents GROND (Generalized Rhythmic and Oscillatory Neurophysiology Descriptor), a comprehensive automated system for characterizing all four major subtypes: lateralized periodic discharges (LPD), generalized periodic discharges (GPD), lateralized rhythmic delta activity (LRDA), and generalized rhythmic delta activity (GRDA).

Approach. Two complementary pipelines were developed. The PD-Profiler combines a per-channel convolutional neural network (ChannelPD-Net) with hemisphere-specific learned evidence traces (HemiCET-UNet), dynamic programming under an approximately-periodic prior, and discharge-locked topographic localization. The RDA-Profiler uses iterative narrowband Hilbert refinement (NB-Hilbert) for frequency and lateralization, with phase-locking value (PLV) analysis for spatial extent. Both pipelines were trained and evaluated on 12,425 EEG segments from 11,729 unique patients.

Main results. Four electroencephalographers — three not involved in algorithm development, plus the primary annotator — independently scored 200 stratified segments per pattern subtype, with analysis restricted to segments accepted by a majority of the four raters. The algorithm exceeded expert–expert (EE) inter-rater agreement on LPD frequency (mean expert–algorithm intraclass correlation coefficient (ICC) 0.931 vs. EE 0.897; paired segment-bootstrap Δ = +0.034, 95% CI [+0.016, +0.055], p < 0.001), and was statistically indistinguishable from EE on every other attribute tested: GPD frequency (p = 0.061), GRDA frequency (p = 0.056), LRDA frequency (p = 0.435), LPD laterality (p = 0.079) and LRDA laterality (mean expert–algorithm κ 0.961 vs. EE 0.993; p = 0.269). No attribute fell significantly below the expert–expert ceiling. PD discharge timing achieved an F1 of 0.817 at single-sample precision (mean absolute timing error 3.1 ms at 200 Hz). Lateralization area under the receiver operating characteristic curve (AUC) was 0.989 for PD hemisphere, 0.909 for LPD vs. GPD, and 0.837 for RDA. PD spatial localization reached 97.3% of expert inter-rater Jaccard agreement.

Significance. This is the first system to reach expert-level inter-rater reliability across lateralization, spatial localization, discharge timing and frequency for both periodic and rhythmic EEG patterns. In post-hoc review of discordant cases, experts judged the algorithm's frequency estimates more accurate than the original expert labels in 94% of cases. Automated characterization can now both substitute for and improve manual annotation in critical-care EEG.

Background

Continuous EEG (cEEG) monitoring in the intensive care unit routinely reveals periodic discharges and rhythmic delta activity — patterns lying along the ictal–interictal continuum (IIC) that are associated with seizures, secondary brain injury, and worse neurological outcome. Under the ACNS 2021 standardized critical care EEG terminology, these patterns are not simply detected but characterized: a complete description specifies the pattern's frequency, its laterality, its spatial distribution, and — for periodic patterns — the timing of individual discharges.

Automated cEEG analysis has concentrated overwhelmingly on the binary detection step: deciding whether a pattern is present and which subtype it belongs to. The characterization step that determines clinical meaning has remained manual, and manual characterization is slow and inconsistent. Frequency in particular is estimated by visual cycle-counting, which is labor-intensive and shows substantial disagreement between experienced readers.

This matters beyond workflow. If the features of epileptiform activity — not merely its presence — drive neurological injury, then large-scale study of the dose–response relationship between those features and outcome requires characterization that is both accurate and reproducible at scale. That is the gap GROND addresses: quantitative characterization at expert-level reliability, applied automatically, opening the door to characterization-stratified analyses and ultimately to clinical trials of antiseizure treatment stratified on pattern features.

Methods

Dataset and annotations

12,425 EEG segments were drawn from 11,729 unique patients, obtained by screening 13,415 candidate LPD/GPD/LRDA/GRDA segments and excluding 990 during quality review. Each segment is 10 seconds of 19-channel monopolar 200 Hz EEG (2,000 samples per channel). Subtype assignment followed the source convention: a majority of at least 10 expert votes for the crowd-labeled subset, single-rater review for the expert-only segments, and a majority of four raters for the prior Tăuţan et al. cohort.

A multi-layer annotation framework with task-specific interactive review tools produced frequency, laterality, discharge-timing and spatial-extent labels. Expert raters reviewed frequency for 4,680 segments; one expert annotated individual discharge times for 2,876 PD segments across three rounds of model-assisted review; three raters annotated channel involvement for 856 segments.

Montage

Both pipelines operate on the 18-channel longitudinal bipolar ("banana") derivation of the canonical 19-channel recording. Because a bipolar derivation suppresses the common mode and could in principle attenuate the broad, synchronous activity that defines the generalized subtypes, this choice was verified by a controlled ablation in which the algorithm, its hyperparameters, the evaluation cohort, the inclusion rule and the metrics were held fixed and only the channel derivation varied across five conditions: bipolar, common-average monopolar, spherical-spline surface Laplacian, and bipolar fused with each of the latter two. The bipolar derivation was the best-performing input on every RDA task, including on GRDA, and neither fusion improved performance.

PD-Profiler

ChannelPD-Net is a lightweight 1-D convolutional network with temporal attention pooling (69,635 trainable parameters) operating independently on each bipolar channel, jointly predicting the probability that a periodic discharge pattern is present and the log-frequency of that pattern. Hemisphere laterality is determined by comparing mean per-channel probability between hemispheres.

HemiCET-UNet (129,537 trainable parameters) is a 1-D U-Net encoder–decoder producing a dense discharge-evidence trace at the full 200 Hz resolution from 8-channel hemisphere input. Its learned evidence is combined with a handcrafted peak prior (pointiness plus Teager–Kaiser energy) through a product-boost rule that amplifies regions of agreement.

Dynamic programming with an approximately-periodic prior recovers individual discharge times from the combined evidence trace, followed by expectation–maximization template refinement. The final PD frequency is the reciprocal of the median inter-peak interval, which is more accurate than the neural frequency prior because it uses the actual detected timing sequence.

Discharge-locked topographic localization computes the mean voltage topography at the moment of each detected discharge. The 19-channel monopolar EEG is bandpass-filtered and transformed to the surface Laplacian; epochs are aligned to the global field power (GFP) peak and averaged with GFP-squared weighting, which suppresses phantom discharges.

RDA-Profiler

The RDA pipeline is entirely signal-processing based and has no trainable parameters. NB-Hilbert performs two passes: a coarse pass estimating lateralization from per-hemisphere variance and frequency from Hilbert instantaneous frequency on the top-3 dominant channels, then a narrowband pass re-filtered at the first-pass estimate that refines both. Spatial extent is quantified by a PLV–amplitude product against a dominant-hemisphere reference.

Development and evaluation

Components were selected through a systematic "contest of methods" in which more than 300 candidate algorithm variants were implemented and benchmarked before any single approach entered the production pipeline. All evaluations used patient-stratified cross-validation so that no patient appeared in both training and evaluation sets. Two independent multi-rater cohorts support the performance claims: a canonical four-rater cohort of 200 stratified segments per subtype scored with the interactive tools developed here, and the prior four-rater Tăuţan et al. cohort scored without them. Agreement is reported as ICC(3,1), Spearman correlation, mean absolute error, Cohen's κ and the Jaccard index, with paired segment-level bootstrap tests comparing expert–algorithm against expert–expert agreement.

Data Description

This release contains the complete EEG data bank, expert annotations, algorithm predictions, trained model weights and rater-facing review tools underlying the GROND manuscript. Everything is distributed from the credentialed BDSP bucket under the prefix s3://bdsp-opendata-credentialed/grond/ (2.0 GB, 53 objects).

Layout

HDF5 bank structure

/metadata/     channel_names_mono (19), channel_names_bipolar (18),
               bipolar_pair_definitions_json, fs_hz=200,
               n_samples_per_segment=2000, segment_duration_s=10, citation
/segments/{segment_id}/
     eeg       (19, 2000) float32 — monopolar referential EEG at 200 Hz
     attrs     patient_id, subtype, subtype_source, eeg_source, eeg_file,
               mat_file, montage, has_discharge_timing
/labels/{segment_id}/
     discharge_times          (N,) float32 — seconds, timing-cohort segments only
     attrs     freq_hz_consensus, freq_per_rater_json, laterality_consensus,
               laterality_per_rater_json, spatial_channels, review_status,
               review_source, gt_laterality, selected_freq
/predictions/pdchar/{segment_id}/   attrs: pdchar_freq_hz, pdchar_laterality,
                                           pdchar_spatial_extent
/predictions/tautan/{segment_id}/   attrs: tautan_freq_hz, tautan_spatial_extent
/predictions/rda_plv/{segment_id}/  attrs: rda_plv_spatial_extent
/cohorts/      segment_ids (order key), pd_gt, irr_canonical  (boolean flags)

Contents by the numbers

The bank holds 14,307 segments from 13,312 patients. Subtype composition: 4,617 LPD, 3,674 GPD, 2,040 LRDA, 3,648 GRDA, 322 seizure, 2 BIPD, 4 other. Cohort flags mark 1,094 segments in the discharge-timing ground-truth cohort and 995 in the canonical four-rater inter-rater-reliability cohort. Algorithm predictions are present for 4,232 segments (PD-Profiler) and 4,667 segments (Tăuţan et al. baseline).

Coverage caveats — please read before analysis

• The bank is broader than the manuscript's analysis cohort. The manuscript reports 12,425 segments from 11,729 patients after quality-review exclusions and subtype restriction; the bank additionally retains seizure, BIPD and "other" segments and segments excluded during review. Use the subtype attribute and the /cohorts/ flags to reconstruct the analysis cohorts rather than assuming every segment was analysed.
• Stored EEG is always 19-channel monopolar referential, whatever a given segment's montage attribute says — that attribute records the montage of the original source file (provenance), not the format stored here. Derive the 18-channel bipolar banana at runtime from /metadata/bipolar_pair_definitions_json. Storing monopolar is deliberate: it lets consumers choose bipolar, average-reference or Laplacian derivations themselves, which is exactly what the montage ablation in the paper does.
• Labels are sparse by design. No segment carries every label type. Frequency, laterality, discharge timing and spatial extent were each annotated on different, partly overlapping subsets; check for attribute presence rather than assuming completeness.
• The rda_plv prediction group is empty in this release; RDA spatial-extent predictions are regenerated from the code repository rather than shipped.
• A small number of segments could not be recovered in their original 10-second slice after a 2026 repository cleanup; per-segment provenance is recorded in the eeg_source attribute, and reproducibility tiers are documented in REPRODUCIBILITY.md in the code repository.

Usage Notes

Getting the data

Access requires BDSP credentialing and acceptance of the Data Use Agreement. Once your AWS credentials are provisioned:

# everything (2.0 GB)
aws s3 sync s3://bdsp-opendata-credentialed/grond/ ./grond-data/

# or just the self-contained data bank (1.6 GB)
aws s3 cp s3://bdsp-opendata-credentialed/grond/grond_data.h5 data/grond_data.h5

Getting the code

git clone https://github.com/bdsp-core/grond.git
cd grond
conda env create -f code/morgoth.yml && conda activate morgoth

The repository contains both pipelines, every algorithm contest entry, all evaluation scripts and the figure- and table-generation scripts for the manuscript. See REPRODUCIBILITY.md for a figure/table → script map and the run order; paper_materials/generate_all_figures.py regenerates the complete figure set from a single entry point.

Reading the bank

import h5py, numpy as np, json

h5 = h5py.File('data/grond_data.h5', 'r')

# 19-channel monopolar EEG for one segment
sid = list(h5['segments'])[0]
mono = h5['segments'][sid]['eeg'][:]                      # (19, 2000) at 200 Hz

# derive the canonical 18-channel bipolar banana
pairs = json.loads(h5['metadata'].attrs['bipolar_pair_definitions_json'])
names = [c.decode() for c in h5['metadata']['channel_names_mono'][:]]
idx   = np.array([[names.index(p['anode']), names.index(p['cathode'])] for p in pairs])
bip   = mono[idx[:, 0]] - mono[idx[:, 1]]                 # (18, 2000)

# expert consensus frequency and the per-rater detail behind it
attrs = h5['labels'][sid].attrs
freq  = attrs.get('freq_hz_consensus')
per_rater = json.loads(attrs.get('freq_per_rater_json', '{}'))

# restrict to a published cohort
ids  = [s.decode() for s in h5['cohorts']['segment_ids'][:]]
irr  = np.array(ids)[h5['cohorts']['irr_canonical'][:]]   # canonical 4-rater cohort

Suggested analyses

• Benchmark a new characterization algorithm against the same expert ceiling. The bank ships both the per-rater labels and the two baselines (GROND and the Tăuţan et al. signal-processing method), so a new method can be scored as an additional rater on the canonical cohort and compared against expert–expert agreement directly.
• Study inter-rater disagreement itself. freq_per_rater_json and laterality_per_rater_json preserve individual rater calls, making the label noise structure — not just the consensus — analysable.
• Re-derive the montage question. Because monopolar EEG is stored, bipolar, average-reference and surface-Laplacian derivations can all be computed from the same segments; the paper's ablation script is released in the repository.
• Discharge-timing research. The timing cohort provides expert-marked individual discharge times at 200 Hz resolution, supporting work on discharge morphology, inter-discharge interval dynamics and phase-locked analyses.
• Link characterization to outcome by joining segment patient identifiers to other BDSP cohorts.

Citation

Please cite the accompanying manuscript (Jing, Sun, Zhang, Byrd, Tăuţan, Basovic, Hadar, Fernandes, Goldenholz, Kim, Struck, Zafar, Westover, Journal of Neural Engineering, under revision) together with this dataset. The predecessor method is Tăuţan et al., Journal of Neural Engineering 22(6):066027, 2025, doi:10.1088/1741-2552/ae2716.

Release Notes

Version 1.0.0 — initial release accompanying the GROND manuscript in Journal of Neural Engineering. Contains the complete 14,307-segment EEG data bank with expert annotations and algorithm predictions, the trained PD-Profiler model weights, and the four independent-expert review tools used for the inter-rater reliability evaluation.

Ethics

This study was conducted under protocols approved by the institutional review boards of Massachusetts General Hospital (protocols #2023P000487 and #2024P002630) and Beth Israel Deaconess Medical Center (protocols #2022P000481 and #2022P000417). Both review boards waived the requirement for informed consent for this retrospective analysis of de-identified EEG recordings. The work was carried out in accordance with the principles of the Declaration of Helsinki.

All data in this release are de-identified: patients carry surrogate identifiers and all dates are shifted. The data remain governed by the BDSP Data Use Agreement.

Acknowledgements

The authors thank the expert electroencephalographers who annotated frequency, discharge timing, and spatial extent for the segments used in this study.

Jin Jing and Chenxi Sun contributed equally as co-first authors. Sahar F. Zafar and M. Brandon Westover contributed equally as co-senior authors.

Dr. Westover's laboratory is supported by grants from the National Institutes of Health (R01AG073410, R01HL161253, R01NS126282, R01AG073598, R01NS131347, R01NS130119) and by Amazon Web Services (AWS).

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 declare no competing interests.