Epilepsy is a neurological disorder characterized by recurring epileptic seizures that can occur at any given time. A system predicting these seizures could give a patient sufficient time to bring himself to safety and to apply a fast-working anti-epileptic treatment to suppress the upcoming seizure. Many seizure detection techniques claim to be able to detect seizures before the marked seizure onset on the EEG. In this work we study the predictions of such a seizure detection system. Materials: For the experiments the MIT Scalp EEG dataset was used, which contains at least 20 hours of EEG and 3 seizures for 24 pediatric patients [1]. Methods: The data is preprocessed using a filter-bank of 8 Butterworth filters of 3 Hz wide between 0.5 and 24.5 Hz [1]. Next the energy is determined for windows of 2 seconds wide with 1 second overlap. This data is presented as input for the machine learning component based on Reservoir Computing (RC) [1]. RC uses a randomly created recurrent artificial neural network, the reservoir, to map the input to a higher dimensional space. The system is trained using a linear readout of the reservoir. After this readout a simple thresholding technique is applied for classification [1]. Experiments and results: For each patient, the system is trained on the data of the 23 other patients. During training, the 2 minutes of EEG prior or following a seizure is not used. Next the system is evaluated on the data of the considered patient. Detections which occurred 10 minutes before the marked seizure onset were considered as true positives. This resulted in a system that was able to detect 75% of the seizures with about 6 false positives per correctly detected seizure. For 11 out of 24 patients some seizures were detected before the marked seizure onset. Furthermore, in 4 of these patients at least half of the seizures were detected before the marked onset, and in a single patient all seizures were detected before the marked onset. Discussion: However, in retrospect, 65% of the early detections are caused by EEG artifacts. Most others can be attributed to inter-ictal spike and wave discharges in the EEG preceding the seizure. Only 3% of the early detections have currently an unknown cause and could be actual early detections. Although nearly all early detections can be considered as false positives. However such false positives have a significantly greater occurrence right before marked seizure onsets, but further research is needed to analyze the cause of this correlation. It might be that these artifacts contain predictive information or for example that the selection criteria for adding EEG sections to the dataset were less strict for EEG sections containing a seizure. These pitfalls call for common guidelines and datasets to evaluate early seizure detection methods. References: [1] Buteneers, P. (2012). Detection of epileptic seizures: the reservoir computing approach (Doctoral dissertation, Ghent University).