Abstracts

Rationale: The detection of Inter-ictal epileptiform discharges (IEDs) by means of automated methods may assist in reviewing EEGs faster and more efficiently. Although many algorithms have been proposed, all of them face a similar problem: to achieve high sensitivities, false positive rates remain high. This makes the manual search for IEDs faster than reviewing all detected events. Our objective is to find a solution that will make automated detection faster and more efficient than conventional methods.Methods: Using an automated IED detection algorithm described in Lodder and van Putten (Inter-Ictal Spike Detection using a Database of Smart Templates. Clin Neurophysiol; 2013), IEDs are detected using a database of trained templates that each represent an IED waveform. During the detection phase, templates nominate possible IEDs with high correlations to themselves and pair their nominations with a certainty value based on their own reliability from past classifications. During the review phase, the ten most likely nominations based on their certainty values are presented and the reviewer is asked to either confirm or reject each of them as an IED. After the ten events have been reviewed, the certainty values of the remaining nominations are adjusted according to their similarity with the reviewed nominations. The next ten nominations with highest certainty are then chosen, and another iteration is performed in the same manner as before. After each iteration, the confirmed events are stored and marked as detected IEDs in the EEG. The number of iterations can be chosen by the reviewer.Results: Using the same template database as described in Lodder and van Putten (2013) (2256 templates), an evaluation of the proposed method was performed on a test dataset of 15 EEGs (306 min, 244 IEDs marked by an experienced reviewer). A total of 8426 events were nominated as epileptiform events and 241 of the 244 IEDs were detected (25.8 fp/min across all certainty levels). IED nominations were presented to a reviewer using the described method, and 15 iterations (10 events per iteration) were performed on each EEG. The number of confirmed IEDs was counted after each iteration. Results show that 71% of all marked IEDs in the test set were found after five iterations, 88% after ten iterations, and 93% after fifteen iterations. The review time of each iteration was approximately twenty seconds, resulting in a total review time of five minutes per EEG.Conclusions: The use of automated IED detection algorithms are limited by their high number of false detections. Our proposed method shows that automated detection methods can be used to find IEDs in a fast and efficient manner, regardless of the high number of false detections. Compared to conventional methods, this can significantly speed up review times and make it more feasible to use long-term recordings for the diagnosis of epilepsy.