Transforming spontaneous premature neonatal EEG to unpaired spontaneous fetal MEG using a CycleGan learning approach

Abstract

AbstractA large body of electroencephalography (EEG) studies has characterized the spontaneous neural activity of premature neonates at different gestational ages. However, evaluation of normal and pathological fetal brain development is still a challenge due to the complexity of the extraction and analysis of fetal neural activity. Fetal magnetoencephalography (fMEG) is currently the only available technique to record fetal neural activity with a time resolution equivalent to that of EEG. However, the signatures and characteristics of fetal spontaneous neural activity are still largely unknown. Benefiting from progress in machine learning and artificial intelligence, we aimed to transfer premature EEG to fMEG, to characterize the manifestation of spontaneous activity using the knowledge obtained from premature EEG.In this study, 30 high-resolution EEG recordings from premature newborns and 44 fMEG recordings, both from 34 to 37 weeks of gestation (wGA) were used to develop a transfer function to predict the spontaneous neural activity of the fetus. After preprocessing, bursts of spontaneous activity were detected using the non-linear energy operator over both EEG and fMEG signals. Next, we proposed a CycleGAN-based model to transform the premature EEG to fMEG and vice versa and evaluated its performance with both time and frequency measurements on both forward and inverse conversions.In the time domain, the values were similar for the mean square error (< 5%) and correlation (0.91 ± 0.05 and 0.89 ± 0.08) for the EEG to fMEG and fMEG to EEG transformations between the original data and that generated by CycleGAN. However, considering the frequency content, the CycleGAN-based model modulated the frequency content of EEG to MEG transformed signals relative to the original signals by increasing the power, on average, in all frequency bands, except for the slow delta frequency band. Our developed model showed promising potential to generate a priori signatures of fMEG manifestations related to spontaneous neural activity. Collectively, this study represents the first steps toward identifying neurobiomarkers of fetal brain development.