Abstracts

Rationale:
A major diagnostic limitation in the field of epilepsy is our inability to assess the electrical activity that arises from deep brain structures, without the use of intracranial electrodes. Standard clinical interpretation of scalp EEGs fails to detect 70-95% of all hippocampal epileptiform activity (HEA). Our goal was to develop an artificial intelligence tool that would allow non-invasive diagnosis and quantification of HEA, using only information extracted from a standard scalp EEG.
Method:
Our dataset consisted of more than 8395 hours of recordings of simultaneous scalp EEG (International 10-20 System with T1/T2 electrodes) and bilateral foramen ovale (FO) electrodes, from 51 patients with temporal lobe epilepsy who previously underwent evaluation at Massachusetts General Hospital.  We applied previously published algorithms for intracranial spike detection (Abou Jaoude et al, Clin Neurophysiol. 2020 Jan;131(1):133-141) and automated sleep staging (Abou Jaoude et al, Sleep. 2020 Jun 1. doi: 10.1093/sleep/zsaa112) to this dataset, to (1) extract ground truth timing information for when HEA occurred; and (2) identify sleep epochs in each recording. Using this temporal information, we automated the creation of a large training dataset with over 2 million scalp EEG examples of HEA occurring during sleep. For the holdout testing dataset, two epileptologists manually annotated 1-hour recordings of simultaneous scalp EEG and FO electrodes from each patient, resulting in a total of 18,851 expert-annotated spikes with no scalp EEG-visible correlate.  Scalp EEG signals were filtered and divided into non-overlapping 250ms epochs, which were used to train a deep neural network to detect HEA. A 5-fold cross-validation across patients was used to select the optimal training parameters and to evaluate the model’s performance. Algorithm performance is reported on the holdout testing data set, averaged across all folds.
Results:
On the receiver-operating characteristic curve, our algorithm achieved an area under the curve (AUC) of 0.86 ± 0.02. The AUC on the precision-recall curve was 0.30 ± 0.07 (chance-level ≈ 0.01). Using a threshold of 0.95 on the algorithm output, there was moderate correlation between the number of spikes detected by the algorithm and the total number of spikes annotated by experts in each 1 hour window (Pearson’s correlation coefficient r = 0.63 ± 0.17).
Conclusion:
This work demonstrates the feasibility of using deep learning to non-invasively detect, quantify, and lateralize scalp EEG-negative HEA, using only signals from a standard scalp EEG. Further work to optimize this algorithm and validate it on external datasets will be necessary to develop a robust tool that can be used in the clinical setting.  Such an algorithm will have many important applications in epilepsy and other neurological diseases.
Funding:
:ADL was funded by NIH/NINDS K23 NS101037, and the American Academy of Neurology Institute. CSJ was funded by NIH/NINDS R25 NS065743. SSC was funded by NIH/NINDS R01 NS062092 and K24 NS088568.