Short-horizon neonatal seizure prediction using EEG-based deep learning.

Area of Science:

• Neonatal neurology
• Computational neuroscience
• Machine learning in medicine

Background:

• Current neonatal seizure risk prediction models focus on long-term forecasts.
• Short-horizon (minute-level) prediction for neonatal seizures remains an underexplored area.
• Quantitative electroencephalography (QEEG) offers potential for detailed neurological monitoring.

Purpose of the Study:

• To investigate the feasibility of short-horizon neonatal seizure prediction using deep learning (DL) and QEEG.
• To identify optimal deep learning models and parameters for minute-level seizure prediction in neonates.

Main Methods:

• Utilized two public EEG seizure datasets comprising 132 neonates and 281 hours of data.
• Benchmarked state-of-the-art time-series deep learning methods, identifying Convolutional LSTM (ConvLSTM) as superior.
• Assessed ConvLSTM performance in a seizure alarm system with varying short prediction horizons (SPH) and seizure occurrence periods (SOP).

Main Results:

• Convolutional LSTM (ConvLSTM) demonstrated strong performance in preictal state classification.
• Optimal performance was achieved at a 3-minute seizure prediction horizon (SPH) and 7-minute seizure occurrence period (SOP), yielding an AUROC of 0.8.
• At 80% sensitivity, the false detection rate was 0.68 events/hour with a 0.36 time-in-warning.

Conclusions:

• Short-horizon neonatal seizure prediction using QEEG-based deep learning is feasible.
• The findings support the need for further validation of these advanced seizure prediction techniques.
• This approach could lead to more timely interventions for neonatal seizures.