Quantitative EEG-Based Deep Learning for Neonatal Seizure Detection using Conv-LSTM

Neonatal seizures cause significant morbidity and mortality, both acutely and in the long term, contributing to adverse neurodevelopmental outcomes. Traditional EEG seizure detection by human experts is constrained by limited efficiency, scalability, and objectivity, which can lead to delay diagnosis and hinder optimal outcomes. Deep learning methods have shown promise neonatal seizure detection, however, more recent DL architectures have not been robustly evaluated Here we evaluated a novel DL architecture Conv-LSTM in neonatal seizure detection. This study involved a publicly available dataset with 79 term neonates, in which 20 infants were randomly selected. The 20 EDF files were processed through bad signal removal, band-pass filtering, and fixed-length epoch segmentation. Feature extraction was then performed using Principal Component Analysis on both entropy-based and wavelet decomposition features, followed by standardization and temporal duplication. The model was able to universally qualify the neonates and achieve an overall accuracy of 86%. Out of all 20 patients, the model could accurately predict 90% of the non-seizure activity and 76% of the seizure activity. Thus, this study determines the potential of Conv-Bi-LSTMs in accurately detecting neonatal seizures based on EEG data, showing promise for reducing the mortality rates and negative long-term impacts on susceptible neonates.