Abstracts

Rationale: Continuous electroencephalography (cEEG) is essential for accurate diagnosis of seizures or status epilepticus in critically ill patients in the intensive care unit (ICU). However, manual interpretation by experienced EEG readers to review the extensive data is labor intensive and time consuming. In this study, we assessed the potential of deep learning models for automated seizure detection from cEEG recordings, thereby expediting cEEG interpretation and clinical management.

Methods: Five records of 21-channel cEEG samples recorded from five different ICU patients on Profusion EEG Software V6 were analyzed. In each record, two trained epileptologists and an EEG technician reviewed and labelled 30-minute samples of non-seizure and seizure activities. The records contained 16 electrographic seizures in total. The experimental protocol of data splitting, model training and testing is shown in Fig. 1. We used a convolutional neural network (CNN) architecture to train deep-learning models for each cEEG channel from raw signal. In addition, a data augmentation technique was used to boost the minority class (seizure event). For reporting outcomes, we used a ‘leave-one-record-out’ testing approach, where four out of five records were used for training the model and the remaining record was used for testing. An event was classified as correct if >= 75% of consecutive segments from an event were classified correctly by the model. We measured accuracy (Acc), sensitivity (Sen) and false positive rate (FPR) of the model after each iteration and reported the average performance after five iterations per channel (one iteration for each test record).

Results: The average performance of individual cEEG channels is shown in Fig.2. Of the 21 channels used, 19 channels gave zero or insignificant FPR, a highly desired outcome to reduce false alarms in the ICU. The range of Acc of the model across these 19 channels was 90.77% - 96.90% and that of Sen was 87.50% - 93.75%. Even considering all 21 channels, including the ones that have a significant FPR, the minimum Acc and Sen were 93.85% and 87.5% respectively and the maximum FPR was 30.56%.

Conclusions: In this study, we demonstrate proof-of-concept of a deep learning model to detect electrographic seizures in ICU patients. The preliminary results obtained (high Acc, high Sen and low FPR) are promising: (1) to automate the detection of electrographic seizures in the ICU, and (2) potential for rapid interpretation (saves the need for laborious manual interpretation) of cEEG recordings, leading to more timely clinical management. Further evaluation of the model on a larger dataset is needed.

Funding: Please list any funding that was received in support of this abstract.: No funding was received.