For one-third of patients with drug-resistant epilepsy alternative approaches must be
investigated in order to improve the quality of their life. A possible approach is to find
automatic methods to detect/predict seizures, in order to adopt interventional actions to
stop or abort the seizure or to limit its side effect. The main problem in this case is to
evaluate the reproducibility of such methods and to standardize them, because there is a lack
of availability of long-term electroencephalography (EEG) data. In this study we want to
create a large long-term EEG database, called NEED (Neuromed Epilepsy EEG Database), whos aim
is to give researchers a way to test their method in a large collection of data. The database
will contain long-term EEG recordings of 200 patients as well as extensive metadata and
standardized annotation of the data sets and will be made freely available for the download
to the research community.
Description
About 50 million people worldwide are affected by epilepsy, which is one of the most frequent
neurological diseases. It is characterized by unpredictable and sudden seizures which could
led to loss of consciousness and uncontrolled motor reactions, severely impacting the quality
of life of epilepsy patients. Only two-thirds of epilepsy patients can control seizures with
anti-epileptic drugs or through epilepsy surgery. For the remaining patients, other
therapeutics approaches should be considered. These approaches include the development of
interventional closed-loop device, able to detect seizures and trigger an interventional
operation, such as administer anti-epileptic drugs or electrically stimulate the
epileptogenic focus to abort the seizure. In order to provide efficient therapeutic
approaches to these patients, in the last decades many efforts have been made to develop
automatic methods to find reliable markers in electroencephalographic (EEG) signal, which is
the gold standard for epilepsy diagnosis, able to predict or early detect seizures in EEG.
Such methods are based on mathematical or computational approaches for EEG signal analysis,
whose aim is to extract complex measures (the so-called "features") from EEG signal, which
are not recognizable with classical visual inspection of EEG signals made by epileptologists
during the diagnosis of epilepsy, in order to use such features as precursors of incoming
seizures (seizure prediction) or as indicators of an ongoing seizure (seizure detection).
Many methods have been proposed in the last years, using linear or non-linear, to extract
features from EEG signal. Recently, some studies used also electrocardiographic signal (ECG),
which is normally recorded together with EEG in epilepsy monitoring, in order to extract
promising features from it. Although these studies showed promising results, nevertheless
they suffer of many limitations. Among them, the use of a limited number of patients and
seizures and the use of only recordings belonging to the phase preceding the seizures (the so
- called pre-ictal period). Such limitations do not allow to determine, for example, the
specificity of such algorithms, because they don't consider also recordings acquired in time
intervals far from the seizures (inter-ictal data) and could led to a overfitting of the
seizure prediction/detection model. The only thing on which all the researchers on epilepsy
agree is the existence of a pre-ictal phase, which is the phase preceding the seizure, the
ictal phase, where the seizure is "active" and inter-ictal phase, which is a period
temporarily far from the seizure. Normally, automatic methods for seizure
prediction/detection consist of three different phases: the pre-processing (artifact removal,
band-pass filtering, data segmentation, …) of EEG signals, feature extraction and feature
selection and classification. This last step usually consists in the use of machine-learning
and statistical methods in order to make decision about the prediction/detection. Basically,
these models should be able to classify each EEG instance in two classes, "seizure" or "no
seizure", using the features extracted from EEG signal. The efficacy of these models highly
depends on how efficiently they are trained and normally the more are the data used to
trained them, the more they are able to take the right decision. Therefore, the availability
of large database of data could allow to develop efficient models for epilepsy
detection/prediction. The possibility for researchers to have access to large database of
continuous and long-term EEG data of epilepsy patients could be a big opportunity to develop
efficient and reliable automatic seizure prediction/detection methods. For these reasons, in
the last years some research groups have proposed and shared public EEG database with
researchers who wants to test their automated seizure detection/prediction models. In
particular, such database have been proposed by Epilepsy Center of University of Bonn and
Freiburg and also by Children's Hospital of Boston and have been made available for free
downloading to researchers. These databases contain long-term EEG recordings acquired during
pre-surgical monitor of epilepsy patients. The number of patients included in these databases
is quite low (from a minimum of 5 to a maximum of 23 patients) and also the number of
seizures is limited (from a minimum of 59 to a maximum of 189). Furthermore, the duration of
the recordings ranges from 40 minutes to 142 hours and the number of metadata (other
information about patients and seizures) is very low. The last database has been proposed in
2008 in the framework of a EU-founded project (EPILEPSIAE), in which 6 different partners
(hospitals, universities, companies) of 4 different countries (Germany, Italy, France,
Portugal) have been involved. This database is not free but is made available for the
download upon payment in 2012. Nowadays, it is the largest epilepsy EEG database available
worldwide (http://epilepsy-database.eu/). It contains data from 275 patients, including EEG
and ECG recordings, metadata, clinical and technical annotations on the data and clinical
information about the patients. Actually, only 60 out of 275 datasets are available for the
download. The aim of this study is to create a long-term EEG database acquired on epilepsy
patients during the non-invasive presurgical monitor at Epilepsy Surgery Unit at IRCCS
Neuromed. The database will include, besides EEG and ECG recordings, clinical and technical
annotations on the data made by expert epileptologists and also clinical information about
the patient, including neuropsychological evaluations. All the data will be first anonymized,
crypted and then made available for the free download. The database will include data from
200 epilepsy patients underwent non-invasive presurgical epilepsy monitoring at Epilepsy
Surgery Unit at IRCCS Neuromed. At the end of non-invasive EEG monitoring, two expert
epileptologists will inspect EEG/ECG data in order to identify the seizures, in particular
the channels where the seizure starts and the time, and everything could be of interest for
the project. All the recordings will be exported in ASCII (American Standard Code for
Information Interchange) format using the DMS Data Management System (Nihon Kohden Europe
Gmbh) software, version 2.9.8 and stored locally. At the same time, clinical and demographic
(only gender and age) data will be acquired for each patient. In particular, for each patient
the following data will be available:
- Demographic data (gender and age);
- Clinical information (epilepsy type, seizure frequency,…);
- Neuropsychological data;
- EEG data acquired according to international 10-20 system;
- ECG data;
- Information about the recordings and the seizures (start and end time, start and end
time of each seizure, …)
All data from all the patients will be included in a single database, and each patient will
be stored in a single compressed archive. The database will be made available after the
completation of a request which can be forwarded by each researcher using a dedicated URL, in
which the researcher will fill in a form when it will be asked to provide the following data:
- Information about the applicant (name, address, affiliation, …)
- GDPR consent Once the request will be received, the compressed archive containing the
whole database will be protected with an ad-hoc alphanumeric password. Such password
will consist of two parts: the first part will be sent by email to the applicant, the
second part will be sent using the regular mail service (two-way authentication).