Neonatal seizure detection from EEG using inception ResNetV2 feature extraction and XGBoost optimized with particle...

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Signal Processing

Background:

• Neonatal seizures are critical indicators of potential neurological damage, necessitating accurate and timely detection.
• Existing machine learning and deep learning methods for electroencephalogram (EEG) analysis face challenges like signal non-stationarity, class imbalance, and limited interpretability.
• Early detection of neonatal seizures is vital for preventing long-term neurological deficits.

Purpose of the Study:

• To introduce a novel hybrid deep learning framework for the early detection of neonatal seizures.
• To address the limitations of existing methods, including signal non-stationarity, class imbalance, and computational load.
• To enhance the accuracy and interpretability of seizure detection in neonatal EEG data.

Main Methods:

• A hybrid deep framework combining discrete wavelet transform (DWT) for signal decomposition and short-time Fourier transform (STFT) for time-frequency spectrogram generation.
• Feature extraction using Inception-ResNetV2 from spectrograms, followed by classification with an XGBoost model.
• Particle swarm optimization (PSO) was employed for offline fine-tuning of the XGBoost model to optimize performance.

Main Results:

• The framework achieved high performance on the Helsinki Neonatal EEG Dataset, with an average accuracy of 98.75%, precision of 98.56%, sensitivity of 98.36%, and specificity of 98.91%.
• Accuracy improved to 99.82% with medium analysis windows, and some cross-validation runs reached 100% accuracy.
• Analysis revealed the significance of theta and delta EEG bands in frontotemporal regions for early seizure detection.

Conclusions:

• The proposed hybrid deep framework effectively addresses key challenges in neonatal seizure detection, offering robust performance and improved interpretability.
• The integration of Inception-ResNetV2 and PSO-optimized XGBoost provides significant performance gains and insights into seizure manifestation.
• This framework presents a practical and effective solution for neonatal intensive care unit (NICU) monitoring and serves as a foundation for clinical decision-support systems.