Abstracts

Rationale: Rodent models of epilepsy are indispensable for probing disease circuits and testing novel therapies due to their genetic and structural similarities to the human brain. However, chronic epilepsy models utilize prolonged EEG recordings that generate substantial amounts of data, resulting in a time-intensive manual labelling process. Previously published automated detectors yield a prohibitively low positive predictive values (PPV), with a false discovery rate (FDR) on the order of 0.5/hour. To address this challenge, we introduce a generalized seizure detection model for automated evaluation of large EEG data sets. Methods: Young male SD rats (2-3 mo, n = 12) were implanted with surface electrodes, EMG pads and intrahippocampal depth electrodes bilaterally. Unilateral intrahippocampal injections of kainic acid were administered to induce epilepsy, while video and EEG recordings were recorded continuously for 3 months. Three-channel EEG data was analyzed by computing standardized features in time, frequency, and synchronization domains for 5-second windows, and seizure segments were also manually labelled by an expert. PCA was used for dimensionality reduction and maximally discriminating features identified by computing Fisher scores. Generalized linear classifiers were built with lasso regularization using these features to classify seizures versus interictal EEG segments. Results: The generalized and individualized classifiers all achieved an AUROC > 0.99 on test data, and at a threshold of 0.1, the classifier had a sensitivity of 0.99, specificity of 0.83 and an overall PPV of 0.37 with an FDR of 0.08/hour. The mean AUROC of our leave-one-out general classifier, in which no data from the test subject was included in training, was 0.88. Our PCA visualizations reflect the separability of the features along the axes constructed. Conclusions: We automatically detected all seizures in a rodent epilepsy models with a high degree of sensitivity and specificity, with an order of magnitude improvement in FDR relative to the state of the art. This detection algorithm will significantly reduce the need to manually review EEG data to identify seizures, allowing the field to leverage larger EEG data sets from subjects with epilepsy to analyze seizure dynamics. Funding: NIH T32MH020017-19NIH T32GM007753-37Bertarelli FellowshipPaul & Daisy Soros Fellowship