Time-Series Generative Adversarial Network Approach of Deep Learning Improves Seizure Detection From the Human...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Seizure detection algorithms typically focus on the epileptogenic cortex.
• Non-localizable epilepsies often involve thalamic neuromodulation, requiring thalamic seizure detection.
• Limited curated thalamic SEEG data hinders deep learning model training.

Purpose of the Study:

• To investigate the efficacy of synthetic data generated by temporal Generative Adversarial Networks (TGAN) in augmenting limited thalamic SEEG samples.
• To improve the performance of deep learning classifiers for detecting ictal and interictal states from thalamic SEEG.
• To facilitate the development of reliable seizure detection algorithms for closed-loop neuromodulation in non-localizable epilepsies.

Main Methods:

• Utilized thalamic SEEG data from 13 epilepsy patients (84 seizures).
• Employed temporal Generative Adversarial Networks (TGAN) to generate synthetic SEEG data.
• Trained a bidirectional Long-Short Term Memory (BiLSTM) deep learning model to classify ictal and interictal states.

Main Results:

• TGAN-generated synthetic data significantly augmented the performance of the BiLSTM classifier.
• The inclusion of synthetic data improved the accuracy of the seizure detection model by 18.5%.
• Demonstrated the feasibility of using synthetic data to enhance seizure detection in limited datasets.

Conclusions:

• Synthetic data generation using TGAN is a viable strategy to overcome data limitations in thalamic SEEG analysis.
• This approach can enhance the accuracy of deep learning-based seizure detection algorithms.
• The methodology holds promise for developing patient-specific seizure detectors for closed-loop neuromodulation devices and classifying seizure onset patterns.