Abstracts

Rationale: To diagnose patients with epilepsy, routine EEG recordings are made in which neurologists search for traces of inter-ictal activity associated with epileptiform discharges. In practice, patients are only diagnosed with epilepsy if these traces are found. Routine EEG recordings last 15 to 25 minutes, and often no epileptiform behavior is seen within this time frame. After another seizure, the patient is required to have a second recording in which epileptiform activity can again be missed. Longer recordings such as in-home monitoring can yield higher diagnostic efficiency, but due to the time consuming nature of visual analysis, the burden is too large on the reviewer to make it practical on routine basis. With assistance from automated detection methods, review time can be reduced significantly. However, these methods are not yet reliable enough to perform this task. We present a novel approach for inter-ictal spike detection to overcome this limitation. Our method is based on smart templates that learn and gain experience with each classification it makes. Methods: The method can be divided into several steps: (i) Inter-ictal spike detection is performed by cross correlating a set of templates with an EEG recording. For each template, regions are found where high correlations (>0.9) exist. From these, a template can propose regions of epileptiform activity with a measure of certainty. (ii) All proposals are pooled together and grouped based on their location in time. Sufficiently large groups are considered as possible regions for epileptiform activity. (iii) A reliability measure is calculated for each group, and if it exceeds a threshold chosen by the reviewer, the region defined by the templates within that group is considered to contain epileptiform activity. A group's reliability is calculated from the weighted average of each template's track record combined with its measure of certainty for a particular event. A template's track record is found by evaluating all previous classifications made (logged by the system) and calculating the accuracy of their outcomes. By allowing the system to continuously gather information and optimize itself, sensitivity and selectivity continuously improves. Results: A database containing 714 templates was constructed from ~50 annotated EEGs. After training and optimization, a test set of ~100 EEGs was used to evaluate the algorithm. Sensitivity and the false positive rate per minute were measured given different reliability thresholds. Results show a positive outcome with sensitivities of up to 0.89 and false positive rates per minute below 0.05, depending on the threshold chosen. More improvements can still be made. The system is capable of analyzing a 20min routine EEG unsupervised in less than 15min, making it feasible in clinical practice. Conclusions: An automated detection system with high sensitivity and selectivity can greatly assist in visual interpretation of the EEG. Not only does it lessen the burden on visual analysis, but it also creates practical opportunities for in-home EEG monitoring which can improve the efficiency of epilepsy diagnostics.