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

Area of Science:

• Neuroscience
• Artificial Intelligence
• Neonatal Medicine

Background:

• Accurate timing of hypoxic-ischemic encephalopathy (HIE) is critical for effective treatment, such as hypothermia.
• Current methods for determining injury onset in neonates are often imprecise.
• Developing reliable methods to estimate the time of hypoxia ischemia (HI) insult is a significant clinical need.

Purpose of the Study:

• To provide preclinical evidence for predicting the time elapsed since HI insult using deep learning on EEG data.
• To investigate the efficacy of transformer-based models in analyzing EEG segments for injury timing.
• To establish a potential 'hidden clock' within neonatal EEG signals for HI progression.

Main Methods:

• Utilized 5-minute EEG segments from fetal sheep subjected to HI.
• Trained Wav2Vec2 transformer models on raw EEG, spectral features, and combined data.
• Employed binary classification to distinguish between EEG segments within or beyond the initial 2-hour post-HI window.
• Validated models using leave-one-out cross-validation.

Main Results:

• Transformer models trained on raw EEG achieved the highest predictive accuracy (89.49%) and F1 score (80.82%).
• Models demonstrated the ability of transformers to extract temporally evolving, injury-related EEG patterns.
• All tested models captured relevant aspects of HI progression over time, though raw EEG performed best.
• Data augmentation with Gaussian noise did not enhance predictive performance.

Conclusions:

• Preclinical findings suggest neonatal EEG contains a 'hidden clock' reflecting HI progression.
• Transformer models can effectively predict the time elapsed since HI insult in the early post-insult window.
• Validated clinical application could refine HIE diagnosis, optimize neuroprotection, and improve clinical decision-making for neonates.