Information
As part of the 3rd International Workshop on Epileptic Seizure Prediction, we started a contest on the prediction of epileptic seizures to quantify the current state of the art in the field. It was decided that this contest will be continued, expanded also to the field of seizure detection.
Motivation
This 'contest' enables for the first time a world-wide comparison of performances of seizure prediction/detection algorithms on the very same data sets. Moreover, by definition the algorithms are evaluated when operating in a prospective manner.
Training & Optimization
Continuous long-term intracranial EEG data recorded from three patients are provided by the Epilepsy Center of the University Medical Center, Freiburg, Germany. The data will be partially available for download as a training set, containing at least 36 hours of EEG data for each patient and at least four seizures. This data can be used to train and optimize prediction/detection algorithms to the patients individually.
Evaluation Procedure
The optimized algorithms will be evaluated on the second part of the EEG data (testing set) to investigate the predictive/detective power of the algorithms. All results will be presented in detail at the next workshop. We will show the fraction of correctly predicted/detected seizures, i.e. the sensitivity of the algorithm, as well as the fraction of false predictions/detections, i.e. the specificity of the algorithm. Both sensitivity as well as specificity will thereby be discussed related to the prediction time / time into the seizure in the case of the detection and duration of the prediction horizon. Moreover, computational needs will also be presented.
The data are evaluated block by block. If an algorithm decides that there will be a seizure, an alarm is raised. The time of such an alarm is the begin of the next chunk of data that is read. If for instance the samples 1000-2000 are read and an event occured in this block that causes an alarm, the alarm would be raised at sample 2001. Moreover, a prediction horizon must be provided for each alarm. A seizure is predicted to begin in the prediction horizon. The prediction horizon is supposed to begin at least 10 seconds after the alarm if it should be judged as a prediction. If it is less than 10 seconds apart long it will be considered as an early detection, since the uncertainty in the determination of the seizure onset is in the order of a few seconds. If the seizure already started the time into the seizure is evaluated for seizure detection. More details can be found here.
A false prediction happens to be an alarm where no seizure begins within the corresponding prediction horizon. A true prediction is an alarm where the seizure starts in the corresponding prediction horizon. The same holds for (early) detections. The specificity is presented as the number of false predictions/detections with respect to time and with respect to the total number of predictions.
A comparison of the performance of the submitted algorithms to a random predictor will be performed. Most likely using the algorithm suggested here. The results will be most likely ranked using the sum of the squared specificity and sensitivity. Since it might be of particular importance to have algorithms with high sensitivity or sepcificity the 'best' algoirthm with respect to this will also be presented.
The evaluation will be done in a prospective way. To this aim, a program data reader is provided that reads one chunk of data at a time. The next chunk of data will only be provided upon the information whether or not the seizure prediction methods raised alarms in the previous time window. Additionally, The datareader is available for download after registering for the contest. This guarantees that it can already be used for the optimization procedure.
Computational Resources
The evaluation of the prediction algorithms will be performed on the testing data by members seizure prediction project in Freiburg, Germany, on well equipped standard PCs running Linux and Windows XP operation systems. Algorithms are expected to run in almost real time on these systems.
We accept programs that run on GNU-Linux or Microsoft Windows. If you like to submit source code, please ensure that non-standard libraries and the makefile are included. If you prefer binaries, please include all libraries, etc. Matlab routines are also welcome. Please avoid usage of specific compiler otions. Our matlab is equipped with the following tool boxes: Wavelet, Statistics, Optimization, Image Processing, and Signal Processing. If specific routines or libraries are needed the participants have to provide them or contact us such that we can try our very best to find a solution for this problem. Optimally, you provide a computer program which needs one single agrument, i.e. the patient number. If you are not sure or if you have specific questions please contact us.
To allow a timely evaluation of the submitted algorithms, please document your submission as exhaustively as possible, and clarify any possible problems beforehand.
Additional Information
- Each participant can hand in patient-optimized algorithms, i.e. one algorithm per patient.
- The groups that register for this competition will be listed on the web site of the next workshop.