Nonconvulsive seizure and status epilepticus detection with deep learning in high-risk adult critically ill patients

Abstract

Nonconvulsive seizure (NCS) is an electrographic seizure activity without manifest motor activity, and prolonged NCS is Nonconvulsive status epilepticus (NCSE). Their delayed treatment could lead to patients’ permanent neurological damage and poor outcomes. Electroencephalogram (EEG) is mandatory for NCS/NCSE detection in critically ill patients, but its interpretation is challenging even though there are Salzburg criteria, the recent consensus on NCS/NCSE conclusion from EEG. This study is a multicenter data collection. The objective is to develop a deep learning model to detect the NCS/NCSE. The two model inputs are variable-length EEG and patients’ clinical information. The 25-component Mel-Frequency Cepstrum Coefficients (MFCC) are the EEG representation while the patients’ clinical data were directly input into the model. The deep learning model consists of two main components. Gated Recurrent Unit (GRU) performs as the feature extraction model, and Neural Network (NN) is the classifier. A total of 337 EEG diagnoses were labeled following the Salzburg criteria, and the positive cases are approximately 27 percent. The final model presents a promising performance. Its recall rate is 86.7 percent during internal validation, and the false-negative rate is 13.3 percent. IMPLICATION OF THE THESIS: The proposed model could be used as a screening tool in the clinical setting to reduce the prone to delayed treatment and underdiagnosis, provided that its performance resulting from external validation is satisfied.