Abstracts

Rationale: Reliable assessments of increased seizure risk could pave the way to novel therapeutic interventions and warning systems for patients. The recent advent of powerful machine and deep learning methods may provide opportunity to develop such seizure risk assessments from central (EEG/ECoG) or peripheral (EKG) long-term data. While recent studies indicate that such assessments may in principle be feasible, it remains unknown whether peripheral or central data perform equally well and, more generally, how both cardiac and cerebral activities are related in terms of their ability for forecast seizures. Further, even long-term clinical data are still nowhere near the big data used in other fields thus necessitating informative features for optimal performance. Critical transitions theory dictates precursors of seizures identifiable in ECoG which may provide the features to augment deep learning methods. We here address two omissions in the current research by (1) identifying the link between peripheral (EKG)- and central (ECoG)-based methods for seizure forecasting, and (2) by assessing the utility of integrating theoretical frameworks of ictogenesis (critical transitions theory) with deep learning to forecast seizures. Methods: Using comprehensive long-term data of either EKG or ECoG from 10 patients, we develop deep convolutional neural networks (CNN) to classify segments as either interictal or preictal (Fig. 1 A). First, data is divided into 30 second segments and, for each modality, a roster of features is computed. For ECoG, we use features dictated by critical transition theory (ECoG-CRIT). For EKG algorithms, features drawn from the literature are used (EKG-LIT). For comparison, power spectral densities (PSD) are used as features in both EKG and ECoG. Segments are randomly shuffled into train and test sets. Algorithm performance is evaluated using test set F1 scores. Layerwise relevance propagation (LRP), a technique used to estimate the relative importance of each feature in forming predictions, is used to assess features relevance. Results: First, we find that using ECoG features from critical transition theory (ECoG-CRIT) in deep learning strongly improves performance compared to deep learning methods using simply the power spectrum as features (ECoG-PSD, Fig. 1 B). LRP identifies ECoG features drawn from critical transitions theory as the most informative. Second, we find that seizure forecasting from EKG using the power spectrum (EKG-PSD) leads to strong performance results with mean F1 scores comparable to the best ECoG-based methods (Fig. 1 C). Here, LRP identifies the most informative frequency range to be <= 40 Hz. Third, we demonstrate that predictive performance between EKG and ECoG is strongly correlated at both the individual patient level and the overall performance across patients (Fig. 1 D). Conclusions: Our work demonstrates that the integration of theory- (critical transition theory) and data-driven (deep learning) methods is capable of outstanding seizure forecasting performance which outperforms methods not incorporating such theoretical knowledge. We demonstrate that deep learning based on cardiac activity forecasts seizures with similar accuracy to ECoG, indicating the potential for less invasive seizure forecasting methods using peripheral sensor data. Finally, our work unravels a tight connection between cardiac and cerebral activity in terms of their ability to indicate a preictal state. Funding: NARSAD Young Investigator Grant to CM