Abstracts

Rationale: There is a growing need for seizure detection algorithms in the epilepsy world. The pediatric epilepsy space is lacking enough pediatric epileptologist as well as other trained staff to recognize a patient’s status. Therefore, the use of an automated seizure detection algorithm interfaced within the workflow of users can increase efficiency as well as save lives. The tested algorithm requires less than 10 seconds to process an hour- long EEG tracing. The algorithm was designed and tested for pediatrics in both an ambulatory and ICU monitoring setting. Testing the effectiveness and reporting the findings is crucial for the transparency of our work to the users and the epilepsy field. The purpose of this study was to determine the age dependency of the algorithm’s performance.

Methods: The Talos algorithm is built on a deterministic framework of metrics which incorporate non-linear dynamic systems methods. The data are analyzed in 20 second segments by two non-linear metrics and the results of those calculations are categorized using a fixed set of statistical rules. The design of the study consisted of two databases. The first database is the Boston Children’s Hospital/MIT pediatric database and the second is the Temple TUSZ portion of the TUH EEG Seizure Corpus. The pediatric records were separated into 4 age groups, all 21 years or younger. An event had to last 10 seconds to qualify as a seizure for the study. The MIT database does not include any absence seizures, but the Temple database included absence seizures in the ages of 3-6 and 6-12, and 94% were under 10 seconds and therefore were removed from the study.

Results: There are clear age differences in the detection performance, in which the lowest performance was with the youngest patients and the highest performance with the oldest patients, although significant detection rates are seen at all ages. The overall results proved that the as the pediatric patients get older, detection rates improve. One observation was that the Temple data from the ages 7-12 had marginally fewer records and fewer seizures compared to the ages 3-6 and 13-19. Due to the removal of absence seizures, the middle two age groups were lacking both records and patients even though the detection percentages increased, whereas ages 13-19 had very many patients and records but many did not have any seizures. However, this did not increase our false positives in the 13-19 group.

Conclusions: Overall, the algorithm performed well in terms of detections of seizures from all age groups. The algorithm performed exceptionally well in determining the first seizure in a file for a patient. Only one true miss of a patient from the MIT databased and due to the file length and seizure time, 4 patient misses were accounted for in the temple database. This is attributed to our buffer window being larger than the time from the start of the file to the first seizure. After concatenation with another file, all but one was detected. This is not the final Talos algorithm but a preliminary one that is not subject to overfitting from the databases we have real time tested on.

Funding: Complete funding provided by AGCS.