Identifying signal-dependent information about the preictal state: A comparison across ECoG, EEG and EKG using deep learning

Abstract

Background: The inability to reliably assess seizure risk is a major burden for epilepsy patients and prevents developing better treatments. Recent advances have paved the way for increasingly accurate seizure preictal state detection algorithms, primarily using electrocorticography (ECoG). To develop seizure forecasting for broad clinical and ambulatory use, however, less complex and invasive modalities are needed. Algorithms using scalp electroencephalography (EEG) and electrocardiography (EKG) have also achieved better than chance performance. But it remains unknown how much preictal information is in ECoG versus modalities amenable to everyday use - such as EKG and single channel EEG - and how to optimally extract that preictal information for seizure prediction.

Methods: We apply deep learning - a powerful method to extract information from complex data - on a large epilepsy data set containing multi-day, simultaneous recordings of EKG, ECoG, and EEG, using a variety of feature sets. We use the relative performance of our algorithms to compare the preictal information contained in each modality.

Results: We find that single-channel EKG contains a comparable amount of preictal information as scalp EEG with up to 21 channels and that preictal information is best extracted not with standard heart rate measures, but from the power spectral density. We report that preictal information is not preferentially contained in EEG or ECoG channels within the seizure onset zone.

Conclusion: Collectively, these insights may help to devise future prospective, minimally invasive long-term epilepsy monitoring trials with single-channel EKG as a particularly promising modality.

Keywords: Deep neural networks; Electrocardiogram; Epilepsy; Precision medicine; Seizure prediction.

Fig. 1

Comparing information from ECoG and EKG in terms of their abilities to distinguish pre- from interictal states. (A) Illustration of the different feature sets paired with deep learning to assess the information contained in sensor data. (B) Classification performance for individual ECoG channels in one patient. (C) Average performance of all channels within the seizure onset zone (SOZ) and channels not within the SOZ (Non-SOZ). (D) Average performances across multiple network runs for each patient and all feature sets. (E) Mean performance metrics across all patients. Except for ECOG-PSDSingleChannel (averages across channels), all plots reflect mean values across five network runs. Whiskers denote standard deviation. Performance was assessed using improvement over chance (IoC-F1) and area under the ROC curve (AUC). * p ≤0·05.

Fig. 2

Comparing information from scalp EEG, EKG and ECoG in terms of their abilities to distinguish pre- from interictal states. (A) Illustration of the different feature sets paired with deep learning to assess the information contained in sensor data. (B) Classification performance for individual EEG channels in one patient. (C) Average performance of all EEG channels over the seizure focus versus other EEG channels. (D) Average performances across multiple network runs for each patient and all feature sets. (E) Mean performance metrics across all patients. Except for EEG-PSDSingleChannel (averages across channels), all plots reflect mean values across five network runs. Whiskers denote standard deviation. Performance was assessed using improvement over chance (IoC-F1) and area under the ROC curve (AUC). * p ≤0·05.

Fig. 3

Layer-wise relevance propagation (LRP) identifies informative preictal biomarkers/frequency bands (high relevance features) for each modality. (A) LRP reveals high relevance frequency bands in ECoG. Bottom: heatmap of feature relevance in two patients. Top: median feature relevance (green) across all preictal segments. ECoG signal frequency ranges with high relevance differs between patients exhibiting a more bimodal (left, grey arrows) or unimodal (right, grey arrow) distribution. (B) LRP reveals high relevance frequency bands (approx <40 Hz) in EKG. (C) Relevance of features calculated from the ECoG signal (ECOG-SEL). Average relevance across all preictal segments, divided into grouped columns by feature subtype and ordered by average relevance, greatest to least from the left. Averages for each grouped subtype are shown in grey. Features derived from synchrony R were the most informative on average. (D) Feature similarity index indicates that high relevance features are less similar within the ECOG-PSD group compared to ECOG-SEL and EKG-PSD. * p < 1e-7. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)