Adaptive Prototype-Based Subtle Transient Pattern Transformers for Enhanced Neonatal Seizure Classification and...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Neonatal seizures require early detection and management to prevent adverse outcomes.
• Traditional analysis of electroencephalography (EEG) signals is time-consuming and struggles with signal nuances.
• Current AI methods for EEG analysis face challenges with large datasets, computational efficiency, and noise sensitivity.

Purpose of the Study:

• To introduce a novel multi-component framework, APSTPT, for enhanced neonatal seizure detection, classification, and severity quantification.
• To overcome the limitations of existing AI-based methods in analyzing neonatal EEG signals.

Main Methods:

• The APSTPT framework incorporates pre-processing for noise reduction and signal amplification.
• Feature extraction utilizes power spectral density and phase locking value, refined by cross-channel covariance attention.
• Prototype learning enables real-time adaptation for dynamic seizure classification, while multiscale entropy analysis assesses seizure severity.

Main Results:

• The APSTPT framework achieved high accuracy in experiments using the TUH EEG Corpus and Zenodo dataset.
• Classification accuracy reached 99.74%, and severity assessment accuracy was 98.87%.
• The framework demonstrated stable performance across different datasets and window lengths.

Conclusions:

• The APSTPT framework offers a robust and flexible solution for neonatal seizure detection and severity assessment.
• Its high accuracy and stable performance suggest suitability for real-time clinical implementation.
• This approach addresses key limitations of previous methods, paving the way for improved neonatal neurological care.