Nonconvulsive Seizure and Status Epilepticus Detection with Deep Learning in High-Risk Adult Critically Ill

Abstract

Nonconvulsive seizure (NCS) is an electrographic seizure activity with subtle motor activity, and prolonged NCS is nonconvulsive status epilepticus (NCSE). Their delayed treatment leads to permanent neurological damage. Electroencephalogram (EEG) is mandatory to detect NCS/NCSE in critically ill patients, but its interpretation is challenging. Our multicenter study proposed a Gated Recurrent Unit (GRU) model to detect the NCS/NCSE. The model was trained with patients' clinical information and 25-component Mel-frequency cepstrum coefficients (MFCC). The target is having NCS/NCSE, and the ground truth is the diagnosis following the Salzburg criteria. As a result, the final model presents a promising performance, with an 86.7%recall rate during internal validation and a 13.3 % false-negative rate. This performance suggests that our model could be a screening tool. It would reduce the prone of underdiagnosis, provided that its performance resulting from external validation is satisfied.