2D Wavelet Scalogram Training of Deep Convolutional Neural Network for Automatic Identification of Micro-Scale Sharp...

We have recently demonstrated that micro-scale Sharp waves in the first few hours EEG of asphyxiated preterm fetal sheep models are the reliable prognostic biomarkers for Hypoxic-Ischemic Encephalopathy (HIE). Higher number of sharp waves within the first 2 hours from a hypoxic insult is shown to be significantly correlated to subcortical neuronal survival in caudate nucleus of striatum. Cerebral therapeutic hypothermia is also shown to be optimally neuroprotective only if initiated as soon as possible during a short window of opportunity within the first 2-3 hours of HI insult, called the latent phase. Therefore there is an urgent necessity for reliable automated algorithms to robustly identify such biomarkers to help early diagnosis of HIE, in real time at birth, before the optimal window of opportunity for treatment is missed.We have previously introduced successful automated signal processing strategies based on the fusion of wavelet and fuzzy techniques, for real-time identification and quantification of sharp waves along a profoundly suppressed EEG/ECoG background, post HI-insult, during the latent phase of sheep models. This work, in particular, for the first time represents a novel online fusion strategy based on the combination of a deep Convolutional Neural Network (CNN) in conjunction with Wavelet Scalogram (WS) for the real-time identification and classification of micro-scale sharp wave biomarkers within the 1024Hz high resolution ECoG recordings as well as the down-sampled 256Hz signals, from in utero preterm fetal sheep. The WS-CNN classifier highlights ability in the identification of HI sharp waves with remarkable high accuracies of 95.34% for 1024Hz and 94.62% for 256Hz data tested over one hour HI ECoG within the most important interval during the first 2 hours of the latent phase, where experiments have suggested hypothermia is optimally effective.