NMT-Events: A Scalp EEG Dataset with Fine-Grained Temporal and Channel-Level Annotations of Abnormal Events

Overview

NMT-Events is an open scalp electroencephalography dataset developed to support research on the automatic localisation and classification of clinically relevant events in EEG recordings. It was prepared as part of the NeuroAssist project, which focuses on making automated EEG analysis more transparent by identifying when an abnormal event occurs and which EEG channels are involved.

The dataset is intended to support the development of systems that move beyond record-level classification. Instead of only predicting whether a complete EEG recording is normal or abnormal, researchers can use the detailed annotations to locate events across time and individual EEG channels.

Dataset Contents

The dataset contains 1,075 scalp EEG recordings. A total of 962 recordings are classified as normal and 113 recordings contain abnormalities.

The EEG signals are stored in European Data Format, EDF. Event annotations are provided in Comma Separated Values, CSV, format. The annotation information identifies the start time, stop time, and EEG channel location of events of interest.

The dataset includes normal EEG data and fine-grained annotations of abnormal waveforms. These annotations can be used for event detection, event localisation, signal classification, model benchmarking, and research on explainable automated EEG analysis.

Data Collection

The EEG recordings were collected in a hospital setting in Rawalpindi, Pakistan. The data acquisition system was the KT88-2400 device developed by Contec Medical Systems.

Recording sessions were conducted by a qualified EEG technician with more than six years of experience. Participants were advised not to take sleep medication or sedatives for at least one day before the recording session.

During the recording procedure, participants were first asked to lie comfortably with their eyes open. They were instructed to reduce movement and blinking to limit recording artefacts. They were then asked to close their eyes, remain relaxed, and stay awake. These recording conditions allowed the collection of both eyes-open activity and resting-state EEG activity.

Technical Recording Details

The EEG recordings were acquired at a sampling frequency of 200 Hz. A standard international 10/20 electrode placement system was used.

Each recording contains 21 EEG channels. A1 and A2 were used as reference channels. The recordings were preserved in EDF format to support compatibility with established EEG analysis software and signal processing tools.

Participant Information

The participant age range is from 6 to 93 years. Recordings from children younger than 6 years were excluded because EEG patterns undergo substantial developmental changes during early childhood.

Approximately 71.63 percent of the recordings were collected from male participants and 28.37 percent were collected from female participants.

The dataset represents a South Asian clinical population. It therefore provides useful data for researchers interested in evaluating automated EEG methods on a demographic that remains underrepresented in many publicly available neurological datasets.

Annotation Procedure

Each complete EEG recording was reviewed by a team of neurologists. The annotation process involved identifying events of interest and recording their precise start time, stop time, and EEG channel location.

The labels were created and stored using an open source annotation tool developed for the NeuroAssist project. After the initial annotation process, the labelled data was reviewed by two expert neurologists. They approved the annotations or corrected identified misclassifications.

This verification procedure was used to improve annotation quality and inter-scorer reliability.

Event Labels

The dataset provides normal EEG recordings and detailed labels for abnormal EEG waveforms. The abnormal event labels include:

Spike and Wave

Spikes

Polyspikes

Polyspikes and Wave

Sharp Wave

Sharp and Slow Wave

Delta Slow Wave

Sharp and Delta Slow Wave

For the experiments reported in the associated NeuroAssist paper, these detailed abnormal labels were combined into two broader classes.

The first class is Spike and Sharp Waves, abbreviated as SSW. This group represents epileptiform discharges and includes spike, polyspike, sharp wave, and related waveform patterns.

The second class is Slow Waves, abbreviated as SW. This group represents non-epileptiform abnormalities and includes delta slow waves, sharp and slow waves, and related slow waveform patterns.

The original detailed labels remain available in the dataset. Researchers can therefore use either the broader classes or the individual event categories according to their research objectives.

Research Applications

NMT-Events can support research in temporal EEG event localisation, channel-level abnormality localisation, automated EEG screening, epileptiform discharge detection, non-epileptiform abnormality detection, computational neurology, clinical neurophysiology, signal processing, and computer aided decision support.

The dataset can also be used to study explainable artificial intelligence in EEG analysis. A model trained with these annotations can provide information about the temporal and spatial location of an event instead of producing only a general normal or abnormal classification.

Other possible applications include machine learning benchmarking, deep learning benchmarking, wavelet-based feature analysis, class imbalance research, subject-based model evaluation, and the development of methods for processing long clinical EEG recordings.

Associated NeuroAssist Study

The associated NeuroAssist study used the dataset to compare classical machine learning methods with deep convolutional neural networks for EEG event localisation.

The classical machine learning methods included support vector machines, decision trees, and random forest classifiers. The deep learning models included VGG16, GoogLeNet, and EfficientNetB1.

Continuous Wavelet Transform representations were used to capture time and frequency information from the EEG signals. The associated study reports that the evaluated deep convolutional neural networks achieved stronger average performance than the classical machine learning methods for the investigated wavelet-based features.

The dataset is being released independently of any single modelling approach. Researchers may use the original EEG signals and annotations to develop, train, validate, and compare alternative methods.

Data Partitioning in the Associated Study

For the event localisation experiments, the EEG records were divided into two-second windows with a 50 percent overlap between consecutive windows.

A window was treated as an event window when at least 25 percent of the window overlapped with an annotated abnormal event. The experimental data was separated using a subject-based division. All windows belonging to one participant were assigned to only one of the training, validation, or test sets.

This subject-based approach was used to reduce the risk of data leakage between the experimental partitions.

Ethical Approval and Consent

The study received approval from the hospital Institutional Review Board. Informed consent was obtained from all participants before data collection.

The associated paper reports the Institutional Review Board reference as MH51214, dated 15 March 2019.

Users of this dataset are responsible for following the licence conditions, institutional requirements, and applicable ethical standards when conducting research with human EEG data.

Limitations

The dataset has a natural class imbalance because normal recordings are considerably more common than abnormal recordings. Researchers should account for this imbalance when selecting evaluation metrics, preparing training data, and interpreting model performance.

The demographic distribution is also unequal, with a larger proportion of recordings collected from male participants. The dataset was collected from a specific South Asian clinical setting, so results should not automatically be generalised to every population, hospital, EEG device, or recording protocol.

Differences in patient demographics, electrode placement, recording equipment, clinical conditions, signal quality, and annotation practices may affect model performance when methods are applied to other datasets.

Responsible Use

NMT-Events is provided for scientific research and educational use. It should not be treated as a replacement for assessment by a qualified neurologist.

Models developed using this dataset should not be used as standalone clinical diagnostic systems without appropriate external validation, regulatory review, clinical supervision, and evaluation on data from the intended population.

Licence

NMT-Events is released under the Creative Commons Attribution 4.0 International licence, CC BY 4.0.

Users may copy, share, and adapt the dataset for permitted purposes provided that appropriate credit is given to the dataset creators, the Zenodo dataset record is cited, and the related publication is acknowledged.

Related Publication

Mohammad Ali Alqarni et al. “NeuroAssist: Open-Source Automatic Event Detection in Scalp EEG.” IEEE Access, Volume 12, 2024.

DOI: 10.1109/ACCESS.2024.3492673