Deep-learning for seizure forecasting in canines with epilepsy

Area of Science:

• Veterinary Neurology
• Biomedical Engineering
• Machine Learning in Medicine

Background:

• Epilepsy affects canines, necessitating reliable seizure forecasting methods.
• Current seizure prediction methods often lack real-time applicability or require computationally intensive features.
• Ambulatory intracranial EEG (iEEG) offers a continuous data stream for seizure monitoring.

Purpose of the Study:

• To introduce a fully automated, subject-specific deep learning system for forecasting seizures using iEEG data.
• To evaluate the system's performance on a portable device in a pseudo-prospective manner.
• To compare the deep learning approach with existing seizure forecasting methods.

Main Methods:

• A convolutional neural network (CNN) system was developed for subject-specific seizure forecasting.
• A genetic algorithm optimized hyper-parameters for each canine's seizure forecasting model.
• Trained CNN models were deployed on a hand-held tablet and tested on iEEG data from four epileptic canines.

Main Results:

• The CNN models achieved statistically significant seizure forecasting rates above chance in all four canines (p < 0.01).
• The system demonstrated a mean sensitivity of 0.79 with 18% time in warning.
• The deep learning method outperformed previous approaches using computationally expensive features and standard machine learning classifiers.

Conclusions:

• The findings support the feasibility of deploying trained CNN models on a hand-held device for real-time seizure forecasting.
• This automated system offers a potential advancement in managing canine epilepsy.
• The study highlights the efficacy of deep learning for analyzing streaming iEEG data for seizure prediction.