The Neurotech EEG Dataset

Abstract

The Neurotech EEG Dataset is a large clinical scalp EEG corpus comprising 23,607 EEG recordings from 4,914 patients acquired by a single EEG monitoring service provider between 2021 and 2025, totaling 212,186 hours of signal data (10.2 TB). A distinguishing feature is the large proportion of ambulatory recordings acquired in patients' homes, including multi-day studies — a real-world, out-of-hospital recording context largely absent from existing large clinical EEG corpora, which are predominantly hospital-based. Recordings span routine outpatient EEGs, ambulatory monitoring, and continuous inpatient/ICU EEG, all acquired with Natus/Xltek NeuroWorks hardware at 256 Hz using the standard International 10-20 montage. The dataset includes 226,486 technician-placed annotations — including 50,482 spike markers, 6,892 seizure markers, 21,330 sharp-wave annotations, and free-text clinical observations. De-identified patient-level clinical metadata (demographics, ICD-10 referral diagnoses, comorbidities, medications, EEG findings, and monitoring summaries) is included for the 4,812 patients with available clinical records. Data are released in BIDS-EEG format with HIPAA-compliant de-identification including per-patient date shifting and automated name scrubbing.

Background

Electroencephalography (EEG) remains the cornerstone of epilepsy diagnosis and management, yet the global shortage of trained EEG readers limits access to expert interpretation. Machine learning offers a path toward scalable automated EEG interpretation, but progress has been constrained by the scarcity of large, clinically representative public datasets. The Neurotech EEG Dataset addresses this gap by providing a large, unselected clinical EEG corpus spanning the full spectrum of clinical EEG practice — from routine 20-minute outpatient recordings to multi-day continuous ICU monitoring — with a uniquely large volume of multi-day ambulatory EEG recorded in patients' homes.

This dataset complements the Harvard EEG Database (HEEDB; ~109,000 patients, ~329,000 recordings, ~3.3 million hours across four hospitals, on the same BDSP platform): whereas HEEDB comprises routine, EMU, and ICU recordings acquired in clinical facilities, the present corpus is far smaller overall but uniquely contributes out-of-hospital, in-home ambulatory EEG, Natus/Xltek hardware, and preserved workflow-native technician annotations that capture the variability of real-world clinical EEG practice.

Methods

Recording

All recordings were acquired using Natus/Xltek NeuroWorks EEG systems with standard International 10-20 electrode placement (25-29 channels; median 28, including ECG), sampled at 256 Hz and stored in EDF+C format. Recording types include routine outpatient EEGs (<1 hour; 36%), ambulatory and short-term monitoring (1-24 hours; 53%), and prolonged continuous monitoring (>24 hours; 11%).

Clinical metadata extraction

For the 4,812 of 4,914 patients (98%) with available scanned clinical documentation, a three-stage on-premises pipeline (OCR text extraction, document segmentation, and structured field extraction using deterministic parsers plus a locally hosted open-weight LLM) produced structured, de-identified fields: demographics, ICD-10 referral diagnoses, comorbidities, medications, EEG findings, and monitoring summaries. All clinical text was processed on-device.

De-identification

Performed in compliance with HIPAA Safe Harbor: identifiers removed from EDF headers; recording dates shifted by a random per-patient offset (uniform in [-365, +365] days) with times of day preserved; patient names in annotation free-text replaced with [NAME]; embedded dates shifted by the same offset. Clinical metadata is released as structured, patient-level fields only (no free text, no dates; ages >89 top-coded to 90).

Data Description

Data Organization (BIDS)

Neurotech/
  dataset_description.json
  participants.tsv          # participant_id, age, sex
  participants.json
  README
  phenotype/                # de-identified patient-level clinical metadata
    demographics.tsv/.json
    diagnoses.tsv/.json      # full ICD-10 codes
    comorbidities.tsv/.json
    medications.tsv/.json
    eeg_findings.tsv/.json
    monitoring.tsv/.json
  sub-Neurotech1/
    ses-1/
      sub-Neurotech1_ses-1_scans.tsv
      eeg/
        sub-Neurotech1_ses-1_task-EEG_eeg.edf
        sub-Neurotech1_ses-1_task-EEG_eeg.json
        sub-Neurotech1_ses-1_task-EEG_channels.tsv
        sub-Neurotech1_ses-1_task-EEG_Xltek.csv
    ses-2/
      ...
  sub-Neurotech2/
    ...

Each session directory contains an EDF file (signal data), a JSON sidecar (recording metadata), a channels TSV, and an Xltek CSV (technician annotations). The top-level phenotype/ directory holds de-identified patient-level clinical metadata (demographics, ICD-10 diagnoses, comorbidities, medications, EEG findings, and monitoring summaries) for the 4,812 patients with clinical records. Note that many EDF files are header-only stubs (filter on n_records > 0 to select files with signal data).

Usage Notes

Recommended use cases: AI/ML development (spike detection, seizure detection, EEG quality assessment); clinical EEG research (epidemiology of EEG findings, annotation variability, ambulatory/home EEG); and methodological development (automated sleep staging and event detection on multi-day recordings).

Code: the BIDS conversion, de-identification, and EHR extraction pipeline is available at github.com/bdsp-core/Neurotech-EEG-Wrangling.

Limitations: single service provider with one hardware platform; annotations are clinical workflow annotations, not multi-expert research labels; clinical metadata covers the 98% of patients with scanned records and is structured-field only (no free text); header-only stub EDFs are included for session integrity — filter on n_records > 0 for signal-level analysis.

Ethics

This project was conducted under IRB protocol number 2022P000417, with the BIDMC IRBs granting a waiver of consent, and under a BAA between BIDMC and Neurotech. The IRB approved the publication of the dataset in a de-identified form with access restricted by a data usage agreement prohibiting attempts at re-identification. The study also complied with the Declaration of Helsinki.

Acknowledgements

The authors thank the clinical EEG technicians whose workflow annotations form a key component of this dataset.

Funding. 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

MBW is a co-founder of, serves as a scientific advisor and consultant to, and has a personal equity interest in Beacon Biosignals. Beacon Biosignals did not contribute funding and played no role in this work.