Transformers Predict Hypoxia-Ischemia Timing in Term Fetal Sheep EEG in the Key 2-Hour Window Post-Insult.

Effective clinical administration of therapeutic interventions such as hypothermia for neonatal hypoxic-ischemic encephalopathy (HIE) depends precisely on determining the timing of injury at birth. This study provides the first preclinical evidence supporting the prediction of time elapsed since the hypoxia ischemia (HI) insult within the key initial 2-hour post-HI window using transformer-based deep learning models trained on 5-minute EEG segments. EEG recordings, sampled at 256 Hz, from eight term fetal sheep following carotid occlusion-induced HI were used to create three datasets for model training: (1) raw EEG, (2) spectral features, and (3) their combination. A Wav2Vec2 transformer model was adapted for binary classification, categorizing EEG segments as either within or beyond the initial two-hour post-insult window. Leave-one-out cross-validation revealed that models trained on raw EEG data achieved the highest class-balanced predictive accuracy (89.49±4.17%) and F1 score (80.82±6.54%), demonstrating that transformers can effectively extract temporally evolving injury-related EEG patterns. While the raw EEG model outperformed the other feature sets, all models captured aspects of the signal relevant to injury progression over time. Gaussian noise data augmentation did not improve predictive performance, likely due to the intrinsic variability of the raw EEG signal.Clinical relevance-Our preclinical findings provide evidence that neonatal EEG contains a latent "hidden clock" encoding HI progression, which, once validated on clinical recordings, could enable precise injury timing estimation. This advancement has the potential to refine clinical decision-making and optimize neuroprotective interventions for neonates with HIE.