Abstracts

Rationale: Around 8-20% of comatose adult patients experience electrographic seizures and their identification requires continuous electroencephalography (CEEG). However, labor involved in maintaining and reading the CEEG recording is significantly higher than routine (or 30 minute) duration EEG. Our current practice is to identify these electrographic seizures retrospectively with delay of few minute to hours. Real-time CEEG analysis could automate detection electrographic seizures enabling immediate identification of seizures, more accurate seizure load quantification while reducing burden on clinicians. The aim of this research is to develop and validate a seizure detection model for use among comatose adult patients. Methods: Three patient with frequent electrographic seizures with continuous EEG monitoring were identified from EEG report database. Data was recorded on 16 channels using a standard bipolar montage with 10-20 EEG recording system at 256 Hz and 16 bits. Seizure detection was performed using supervised learning algorithms trained on patient data with seizures identified by board certified epileptologists and EEG technicians. For classification, the first step was to extract features from the raw EEG, here we used power spectral density in 8 frequency bands, 0.5-4Hz, 4-8Hz, 8-13Hz, 13-30Hz, 30-47Hz, 53-75Hz, 75-103Hz, and 103-128 Hz for each channel from 20 second windows with half overlap, generating 128 features every 10 seconds. Patient 1 had 22 seizures, Patient 2 had 7 seizures, and Patient 3 had 32 seizures. The second step was to perform a preliminary classification to identify windows with significant activity from background activity using a computationally efficient Linear Discriminant Analysis (LDA). The third step was to train a cost-sensitive Support Vector Machine (cSVM) with a radial basis function, to distinguish between seizures and non-seizure windows. We counted a putative seizure detection once if multiple windows that were classified as seizures happened in close temporal proximity. We generated patient specific models as well as a generalized model for seizure detection. Performance was evaluated by measuring true positive and false positive rates of seizure detections on out of sample data averaged across a 5-fold cross validation. Results: LDA was used to classify low activity from high activity windows. LDA reduced the total number of windows from 39,748 across the 3 patients for an average reduction of 94.5% of the data for further classification without loss of any seizure data. cSVM was then trained on the high activity data and achieved 90% sensitivity in patient 1 and 100% sensitivity in patients 2 and 3 with a ratio of False positives of 0.33 for Patient 1, 1 for Patient 2 and 0.46 for Patient 3. Our goal was to keep a false positive rate lower than one false positive for every true positive and this was achieved in all three patients with high sensitivity. A generalized model was also trained on all three patients. This model was able to achieve 94% sensitivity (all missed seizures were in Patient 1) with a ratio of false positives to seizures of 0.48. However, in this model, data was used from all three patients in the training and a better test will be to train on some patients and measure prediction in new patients. When analyzing the prediction accuracy of each channel alone, it was found that there was little correlation between channel location and seizure focus. Conclusions: Using linear features, power spectral density, and two staged classifier with LDA to identify windows with activity and cSVM to distinguish seizures from non-seizure data, we were able to obtain high classification sensitivity and accuracy that achieved fewer than one false positive for every true positive. After training, this pipeline for classification is very efficient and can run much faster than real time, completing the analysis within the 10 second acquisition window. This preliminary project shows the feasibility of testing in a larger patient cohort in retrospective data. If acceptable patient specific models, or generalized models, can be developed, this algorithm will be implemented to acquire and analyze patient data in real-time. Funding: No funding