Machine Learning for Short-term Seizure Forecast Using Neonatal EEG.

Area of Science:

• Neonatal neurology
• Machine learning in healthcare
• Biomedical signal processing

Background:

• Neonatal seizures require continuous monitoring in the NICU.
• Accurate short-term seizure forecasting is crucial for timely intervention.
• Hypoxic-ischemic encephalopathy (HIE) is a common diagnosis associated with neonatal seizures.

Purpose of the Study:

• To develop and validate a novel machine learning (ML) method for continuous short-term seizure forecasting.
• To improve the accuracy of seizure prediction in neonates using electroencephalogram (EEG) data.
• To enhance clinical decision-making for HIE infants in the NICU.

Main Methods:

• Utilized an Adaptive Boosting classifier for seizure prediction.
• Extracted 22 features from short-time segmented neonatal EEG data.
• Trained and tested the model on three datasets (148 neonates, diverse gestation periods, and diagnoses including HIE).

Main Results:

• Achieved a Matthews Correlation Coefficient (MCC) of 0.466±0.078 and Area Under the ROC Curve (AUROC) of 0.738±0.041.
• Outperformed the current state-of-the-art ML model (MCC 0.255±0.054, AUC 0.678±0.041).
• Demonstrated state-of-the-art performance for short-term seizure forecasting.

Conclusions:

• The proposed ML method offers a significant advancement in continuous short-term seizure forecasting.
• This tool can optimize NICU resource allocation and improve treatment decisions for newborns.
• The model's superior performance provides a promising approach for neonatal seizure management.