Enhanced EEG seizure recognition after hypoxia-ischemia in fetal sheep using transformer-based wavelet-scalogram deep learning

Abstract

This study introduces transformer-based deep learning models as superior alternatives to convolutional neural network (CNN) architectures for seizure detection after hypoxia–ischemia (HI) in electroencephalography (EEG) recordings in the developing brain. We further demonstrate that these models excel in classifying of varied and subtle morphologies of electrographic seizures observed in the EEG of fetal sheep after HI. The study explores model training combinations from four cohorts of Romney/Suffolk fetal sheep (HI-normothermia-term (n = 7), HI-hypothermia-term (n = 14), sham-normothermia-term (n = 5), and HI-normothermia-preterm (n = 14)), totalling 31,015 EEG segments from 17,300 h of recordings. We evaluated multiple Transformer architectures using both 2D wavelet-scalograms (WS) and 1D raw EEG signals through leave-one-out and k-fold cross-validation, comparing models’ performance with or without sham-normothermia-term data and contrasting them with CNN-based results from our previous work. The Data-Efficient image Transformer (DeiT) emerged as the top-performing model, achieving 99.60 % accuracy (AUC = 0.992) when fed with WS, surpassing the 2D WS CNN result (98.94 % accuracy, AUC = 0.967) from our prior study. Notably, the Wav2Vec2 transformer model, fed with raw 1D EEG segments, demonstrated comparable performance with 99.60 % accuracy (AUC = 0.992), outperforming the 1D CNN result (98.43 % accuracy, AUC = 0.961) from the previous study. More importantly, the transformers can adeptly distinguish between seizures in both the preterm and term brain, and under the influence of treatment with 99.92 % accuracy (AUC = 0.997). Our findings highlight the superior ability of transformer models to capture long-range dependencies within EEG data, enhancing seizure detection without the added complexity of generating 2D WS images. Transformer models show promise for generalized automated seizure detection, irrespective of the perinatal brain maturation stage and the influence of hypothermia, offering potential improvements for seizure detection after HI in newborns.