SeizFt: Interpretable Machine Learning for Seizure Detection Using Wearables.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Epilepsy monitoring relies on accurate seizure detection.
• Wearable electroencephalogram (EEG) devices offer continuous data collection.
• Existing seizure detection methods often lack interpretability and generalization capabilities.

Purpose of the Study:

• To develop and evaluate SeizFt, a novel, interpretable machine learning framework for automated seizure detection.
• To improve the accuracy and reduce false alarms in seizure detection using wearable EEG data.
• To enhance the generalizability and resilience of seizure detection models to EEG variations.

Main Methods:

• Utilized Fourier Transform (FT) Surrogates for data augmentation and class balancing.
• Employed an ensemble of decision trees (CatBoost classifier) for classifying EEG epochs.
• Extracted meaningful features including delta and theta waves, entropy, and fractal dimensions.

Main Results:

• SeizFt achieved first place in the Seizure Detection Grand Challenge at ICASSP 2023.
• Outperformed state-of-the-art models with a combined score of 40.15 (OVLP and EPOCH metrics).
• Demonstrated significant improvement (~30%) over the next best approach, with minimized false alarms.

Conclusions:

• SeizFt offers an accurate and interpretable solution for seizure detection using wearable EEG.
• The framework shows potential for real-time, continuous epilepsy monitoring and personalized medicine.
• Key predictive features identified can inform future seizure detection algorithm development.