Abstracts

Rationale: Spike and wave discharges (SWDs) are pathognomonic EEG diagnostic events of absence seizures for people with genetic generalized epilepsy (GGE). The Genetic Absence Epilepsy Rats from Strasbourg (GAERS) are one of the most well-validated animal models of GGE, which has increased our pathophysiological knowledge of GGEs. We develop a universal SWDs detection method for both GAERS rats and GGE patients with absence seizures, using a neural network architecture that integrates graph neural network and recurrent neural network. The proposed detection framework can be conveniently applied in research and clinical settings, as our algorithm shows that is much more time-efficient than traditional manual EEG inspection and epileptiform discharge marking.

Methods: The EEG data of GAERS rats was recorded using four epidural electrodes from 192 24-hour EEG sessions recorded from 18 animals. In 11 patients with GGE and absence seizures, 24-hour scalp-EEG data were recorded using the 10-20 system. The continuous EEG recordings were segmented into 2-second epochs, in which the graph network and node features were inferred using correlation coefficients and frequency component analysis, respectively. In the total 192 sessions of GAERS recordings, the detection method was trained on 19 sessions of 24-hour EEG segments and tested on the remaining 175 sessions. The GGE SWDs detector was conducted with the leave-one-out (LOO) cross-validation method. The GGE SWDs detector trained within our patient cohort, was then validated on the TUH EEG Seizure Corpus (TUSZ) public dataset on recordings from 12 patients who had GGE with absence seizures.

Results: The SWDs detection performance on 175 24-hour EEG recordings of GAERS rats showed a sensitivity of 98.01% and a false positive (FP) rate of 0.96 per hour. The SWDs detection performance on the GGE patients showed 100% sensitivity in five out of 11 patients, while the remaining patients obtained a sensitivity over 98.9%. Moderate FP rates were seen in our patient cohort, with an average of 2.21 in the range [1.33, 3.39]. The SWDs detector trained with our patient cohort applied to the TUSZ dataset for 11 out of 12 patients showed 100% sensitivity, and the remaining patient achieved 85.71% sensitivity, for which one SWD less than 3 seconds was missed. The FP rates of ten patients in the TUSZ dataset were zero, while the FP rates in the rest two patients were 2.74 and 4.04.

Conclusions: We have developed a robust SWDs detection based on a novel graph network algorithm that shows high sensitivity and specificity for EEG recordings from both GAERS rats and GGE patients. This detector can assist researchers and neurologists with time-efficient and accurate EEG quantification of SWDs from prolonged EEG recordings.

Funding: Please list any funding that was received in support of this abstract.: PMCE work is supported by the National Health and Medical Research Council (NHMRC) Early Career Fellowships (APP1166170).