Long-Term Neonatal EEG Modeling with DSP and ML for Grading Hypoxic-Ischemic Encephalopathy Injury.

Hypoxic-Ischemic Encephalopathy (HIE) occurs in patients who experience a decreased flow of blood and oxygen to the brain, with the optimal window for effective treatment being within the first six hours of life. This puts a significant demand on medical professionals to accurately and effectively grade the severity of the HIE present, which is a time-consuming and challenging task. This paper proposes a novel workflow for background EEG grading, implementing a blend of Digital Signal Processing (DSP) and Machine-Learning (ML) techniques. First, the EEG signal is transformed into an amplitude and frequency modulated audio spectrogram, which enhances its relevant signal properties. The difference between EEG Grades 1 and 2 is enhanced. A convolutional neural network is then designed as a regressor to map the input image into an EEG grade, by utilizing an optimized rounding module to leverage the monotonic relationship among the grades. Using a nested cross-validation approach, an accuracy of 89.97% was achieved, in particular improving the AUC of the most challenging grades, Grade 1 and Grade 2, to 0.98 and 0.96. The results of this study show that the proposed representation and workflow increase the potential for background grading of EEG signals, increasing the accuracy of grading background patterns that are most relevant for therapeutic intervention, across large windows of time.