Deep neural networks for highly sensitive detection of seizure onset times in long-term EEG recordings

Introduction:
Review of long-term EEG recordings in epilepsy monitoring is labor- and resource intensive. Although software packages for semi-automated workflows achieving reasonable sensitivities and false positive rates are available, there is an unmet demand for reliable, high-sensitivity algorithms, which allow to limit review processes to inspection of computer-generated seizure annotations. We present a novel approach to deep neural network-based EEG seizure detection and validate the model on a large dataset, which had been used in a recent study comparing three commercial seizure detection packages (Koren et al., Epilepsia 2021).
Methods:
In the proposed approach a deep neural network is trained for detection of temporal seizure onsets. This is achieved by a network design enabling simultaneous perception of pre-ictal EEG and seizure onset phases to make detections. The network is based on ResNet-blocks and a feature pyramid structure. It includes one network for the detection of seizure onsets and a second network for estimation of the exact onset times. It was trained with EEG recordings from the Temple University EEG Seizure Corpus, including 490 patients and 2370 seizures. The validation study was performed using EEG recordings from 81 patients (6900 hours in total) who underwent long-term video-EEG monitoring at the Department of Neurology, Clinic Hietzing. Development and training of the network was strictly done without access to validation data. Seizures had been annotated by EEG experts based on the EEG including original video-EEG reports (Koren et al., Epilepsia 2021), the total number of seizure annotations was 790. Detection sensitivities, false positive rates, and positive prediction values were assessed for each patient separately and statistically evaluated. Additionally, we measured temporal detection offsets, which is the time difference between electrographic seizure onset labels created by reviewers and detection times from the neural network. Since the network is designed to create labels close to the seizure onset time while perceiving EEG from the ongoing seizure, the temporal detection offset should not be mistaken with real-time detection delay. Detection offsets take negative values if annotations are created prior to seizure onset time and positive values otherwise.
Results:
Mean per-patient sensitivity was 87.9%. In 61/81 patients (75%) per-patient sensitivity was above 95%. Only in 2/81 patients no seizure was detected, these two patients had only one (false-negative) seizure respectively. Mean per-patient false positive rate was 19.0/24 h. The mean per-patient positive predictive value was 9.5%. The mean temporal detection offset of automatic annotations with respect to electrographic seizure onsets was -0.67 s, first and third quantiles were -8.5 s and 4.5 s respectively. In the referenced study, the software with highest sensitivity (Persyst 13) achieved a mean per-patient sensitivity of 81.6% and a mean per-patient false positive rate of 21.6/24 h.
Conclusions:
A highly effective deep neural network architecture and training approach for detection of EEG seizure onset times was introduced. The model provides seizure detection with high sensitivity and good estimates for the seizure onset time. It outperforms the most sensitive seizure detection software in the referenced study in terms of sensitivity and false positive rate.